From cffc240a7327ae60671ff533469fc4ed4bf605de Mon Sep 17 00:00:00 2001
From: Nerogar <nerogar@arcor.de>
Date: Sun, 23 Oct 2022 14:05:25 +0200
Subject: fixed textual inversion training with inpainting models

---
 modules/textual_inversion/textual_inversion.py | 27 +++++++++++++++++++++++++-
 1 file changed, 26 insertions(+), 1 deletion(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index 0aeb0459..2630c7c9 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -224,6 +224,26 @@ def validate_train_inputs(model_name, learn_rate, batch_size, data_root, templat
     if save_model_every or create_image_every:
         assert log_directory, "Log directory is empty"
 
+def create_dummy_mask(x, width=None, height=None):
+    if shared.sd_model.model.conditioning_key in {'hybrid', 'concat'}:
+
+        # The "masked-image" in this case will just be all zeros since the entire image is masked.
+        image_conditioning = torch.zeros(x.shape[0], 3, height, width, device=x.device)
+        image_conditioning = shared.sd_model.get_first_stage_encoding(shared.sd_model.encode_first_stage(image_conditioning))
+
+        # Add the fake full 1s mask to the first dimension.
+        image_conditioning = torch.nn.functional.pad(image_conditioning, (0, 0, 0, 0, 1, 0), value=1.0)
+        image_conditioning = image_conditioning.to(x.dtype)
+
+    else:
+        # Dummy zero conditioning if we're not using inpainting model.
+        # Still takes up a bit of memory, but no encoder call.
+        # Pretty sure we can just make this a 1x1 image since its not going to be used besides its batch size.
+        image_conditioning = torch.zeros(x.shape[0], 5, 1, 1, dtype=x.dtype, device=x.device)
+
+    return image_conditioning
+
+
 def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
     save_embedding_every = save_embedding_every or 0
     create_image_every = create_image_every or 0
@@ -286,6 +306,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc
     forced_filename = "<none>"
     embedding_yet_to_be_embedded = False
 
+    img_c = None
     pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step)
     for i, entries in pbar:
         embedding.step = i + ititial_step
@@ -299,8 +320,12 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc
 
         with torch.autocast("cuda"):
             c = cond_model([entry.cond_text for entry in entries])
+            if img_c is None:
+                img_c = create_dummy_mask(c, training_width, training_height)
+
             x = torch.stack([entry.latent for entry in entries]).to(devices.device)
-            loss = shared.sd_model(x, c)[0]
+            cond = {"c_concat": [img_c], "c_crossattn": [c]}
+            loss = shared.sd_model(x, cond)[0]
             del x
 
             losses[embedding.step % losses.shape[0]] = loss.item()
-- 
cgit v1.2.3


From 39541d7725bc42f456a604b07c50aba503a5a09a Mon Sep 17 00:00:00 2001
From: Fampai <>
Date: Fri, 4 Nov 2022 04:50:22 -0400
Subject: Fixes race condition in training when VAE is unloaded

set_current_image can attempt to use the VAE when it is unloaded to
the CPU while training
---
 modules/hypernetworks/hypernetwork.py          | 4 ++++
 modules/textual_inversion/textual_inversion.py | 5 +++++
 2 files changed, 9 insertions(+)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py
index 6e1a10cf..fcb96059 100644
--- a/modules/hypernetworks/hypernetwork.py
+++ b/modules/hypernetworks/hypernetwork.py
@@ -390,7 +390,10 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
     with torch.autocast("cuda"):
         ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=hypernetwork_name, model=shared.sd_model, device=devices.device, template_file=template_file, include_cond=True, batch_size=batch_size)
 
+    old_parallel_processing_allowed = shared.parallel_processing_allowed
+
     if unload:
+        shared.parallel_processing_allowed = False
         shared.sd_model.cond_stage_model.to(devices.cpu)
         shared.sd_model.first_stage_model.to(devices.cpu)
 
@@ -531,6 +534,7 @@ Last saved image: {html.escape(last_saved_image)}<br/>
 
     filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt')
     save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename)
+    shared.parallel_processing_allowed = old_parallel_processing_allowed
 
     return hypernetwork, filename
 
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index 0aeb0459..55892c57 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -273,7 +273,11 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc
     shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..."
     with torch.autocast("cuda"):
         ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file, batch_size=batch_size)
+
+    old_parallel_processing_allowed = shared.parallel_processing_allowed
+
     if unload:
+        shared.parallel_processing_allowed = False
         shared.sd_model.first_stage_model.to(devices.cpu)
 
     embedding.vec.requires_grad = True
@@ -410,6 +414,7 @@ Last saved image: {html.escape(last_saved_image)}<br/>
     filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt')
     save_embedding(embedding, checkpoint, embedding_name, filename, remove_cached_checksum=True)
     shared.sd_model.first_stage_model.to(devices.device)
+    shared.parallel_processing_allowed = old_parallel_processing_allowed
 
     return embedding, filename
 
-- 
cgit v1.2.3


From cdc8020d13c5eef099c609b0a911ccf3568afc0d Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Sat, 19 Nov 2022 12:01:51 +0300
Subject: change StableDiffusionProcessing to internally use sampler name
 instead of sampler index

---
 modules/api/api.py                             | 26 ++++++++---------------
 modules/hypernetworks/hypernetwork.py          |  4 ++--
 modules/images.py                              |  2 +-
 modules/img2img.py                             |  4 ++--
 modules/processing.py                          | 29 +++++++++++---------------
 modules/sd_samplers.py                         | 13 +++++++++---
 modules/textual_inversion/textual_inversion.py |  4 ++--
 modules/txt2img.py                             |  3 ++-
 modules/ui.py                                  |  2 +-
 scripts/img2imgalt.py                          |  4 ++--
 scripts/xy_grid.py                             | 12 +++++------
 11 files changed, 49 insertions(+), 54 deletions(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/api/api.py b/modules/api/api.py
index 596a6616..0eccccbb 100644
--- a/modules/api/api.py
+++ b/modules/api/api.py
@@ -6,9 +6,9 @@ from threading import Lock
 from gradio.processing_utils import encode_pil_to_base64, decode_base64_to_file, decode_base64_to_image
 from fastapi import APIRouter, Depends, FastAPI, HTTPException
 import modules.shared as shared
+from modules import sd_samplers
 from modules.api.models import *
 from modules.processing import StableDiffusionProcessingTxt2Img, StableDiffusionProcessingImg2Img, process_images
-from modules.sd_samplers import all_samplers
 from modules.extras import run_extras, run_pnginfo
 from PIL import PngImagePlugin
 from modules.sd_models import checkpoints_list
@@ -25,8 +25,12 @@ def upscaler_to_index(name: str):
         raise HTTPException(status_code=400, detail=f"Invalid upscaler, needs to be on of these: {' , '.join([x.name for x in sd_upscalers])}")
 
 
-sampler_to_index = lambda name: next(filter(lambda row: name.lower() == row[1].name.lower(), enumerate(all_samplers)), None)
+def validate_sampler_name(name):
+    config = sd_samplers.all_samplers_map.get(name, None)
+    if config is None:
+        raise HTTPException(status_code=404, detail="Sampler not found")
 
+    return name
 
 def setUpscalers(req: dict):
     reqDict = vars(req)
@@ -82,14 +86,9 @@ class Api:
         self.app.add_api_route("/sdapi/v1/artists", self.get_artists, methods=["GET"], response_model=List[ArtistItem])
 
     def text2imgapi(self, txt2imgreq: StableDiffusionTxt2ImgProcessingAPI):
-        sampler_index = sampler_to_index(txt2imgreq.sampler_index)
-
-        if sampler_index is None:
-            raise HTTPException(status_code=404, detail="Sampler not found")
-
         populate = txt2imgreq.copy(update={ # Override __init__ params
             "sd_model": shared.sd_model,
-            "sampler_index": sampler_index[0],
+            "sampler_name": validate_sampler_name(txt2imgreq.sampler_index),
             "do_not_save_samples": True,
             "do_not_save_grid": True
             }
@@ -109,12 +108,6 @@ class Api:
         return TextToImageResponse(images=b64images, parameters=vars(txt2imgreq), info=processed.js())
 
     def img2imgapi(self, img2imgreq: StableDiffusionImg2ImgProcessingAPI):
-        sampler_index = sampler_to_index(img2imgreq.sampler_index)
-
-        if sampler_index is None:
-            raise HTTPException(status_code=404, detail="Sampler not found")
-
-
         init_images = img2imgreq.init_images
         if init_images is None:
             raise HTTPException(status_code=404, detail="Init image not found")
@@ -123,10 +116,9 @@ class Api:
         if mask:
             mask = decode_base64_to_image(mask)
 
-
         populate = img2imgreq.copy(update={ # Override __init__ params
             "sd_model": shared.sd_model,
-            "sampler_index": sampler_index[0],
+            "sampler_name": validate_sampler_name(img2imgreq.sampler_index),
             "do_not_save_samples": True,
             "do_not_save_grid": True,
             "mask": mask
@@ -272,7 +264,7 @@ class Api:
         return vars(shared.cmd_opts)
 
     def get_samplers(self):
-        return [{"name":sampler[0], "aliases":sampler[2], "options":sampler[3]} for sampler in all_samplers]
+        return [{"name":sampler[0], "aliases":sampler[2], "options":sampler[3]} for sampler in sd_samplers.all_samplers]
 
     def get_upscalers(self):
         upscalers = []
diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py
index 7f182712..fbb87dd1 100644
--- a/modules/hypernetworks/hypernetwork.py
+++ b/modules/hypernetworks/hypernetwork.py
@@ -12,7 +12,7 @@ import torch
 import tqdm
 from einops import rearrange, repeat
 from ldm.util import default
-from modules import devices, processing, sd_models, shared
+from modules import devices, processing, sd_models, shared, sd_samplers
 from modules.textual_inversion import textual_inversion
 from modules.textual_inversion.learn_schedule import LearnRateScheduler
 from torch import einsum
@@ -535,7 +535,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
                 p.prompt = preview_prompt
                 p.negative_prompt = preview_negative_prompt
                 p.steps = preview_steps
-                p.sampler_index = preview_sampler_index
+                p.sampler_name = sd_samplers.samplers[preview_sampler_index].name
                 p.cfg_scale = preview_cfg_scale
                 p.seed = preview_seed
                 p.width = preview_width
diff --git a/modules/images.py b/modules/images.py
index ae705cbd..26d5b7a9 100644
--- a/modules/images.py
+++ b/modules/images.py
@@ -303,7 +303,7 @@ class FilenameGenerator:
         'width': lambda self: self.image.width,
         'height': lambda self: self.image.height,
         'styles': lambda self: self.p and sanitize_filename_part(", ".join([style for style in self.p.styles if not style == "None"]) or "None", replace_spaces=False),
-        'sampler': lambda self: self.p and sanitize_filename_part(sd_samplers.samplers[self.p.sampler_index].name, replace_spaces=False),
+        'sampler': lambda self: self.p and sanitize_filename_part(self.p.sampler_name, replace_spaces=False),
         'model_hash': lambda self: getattr(self.p, "sd_model_hash", shared.sd_model.sd_model_hash),
         'date': lambda self: datetime.datetime.now().strftime('%Y-%m-%d'),
         'datetime': lambda self, *args: self.datetime(*args),  # accepts formats: [datetime], [datetime<Format>], [datetime<Format><Time Zone>]
diff --git a/modules/img2img.py b/modules/img2img.py
index be9f3653..9fc5b693 100644
--- a/modules/img2img.py
+++ b/modules/img2img.py
@@ -6,7 +6,7 @@ import traceback
 import numpy as np
 from PIL import Image, ImageOps, ImageChops
 
-from modules import devices
+from modules import devices, sd_samplers
 from modules.processing import Processed, StableDiffusionProcessingImg2Img, process_images
 from modules.shared import opts, state
 import modules.shared as shared
@@ -99,7 +99,7 @@ def img2img(mode: int, prompt: str, negative_prompt: str, prompt_style: str, pro
         seed_resize_from_h=seed_resize_from_h,
         seed_resize_from_w=seed_resize_from_w,
         seed_enable_extras=seed_enable_extras,
-        sampler_index=sampler_index,
+        sampler_index=sd_samplers.samplers_for_img2img[sampler_index].name,
         batch_size=batch_size,
         n_iter=n_iter,
         steps=steps,
diff --git a/modules/processing.py b/modules/processing.py
index 03c9143d..be2edf48 100644
--- a/modules/processing.py
+++ b/modules/processing.py
@@ -2,6 +2,7 @@ import json
 import math
 import os
 import sys
+import warnings
 
 import torch
 import numpy as np
@@ -66,19 +67,15 @@ def apply_overlay(image, paste_loc, index, overlays):
 
     return image
 
-def get_correct_sampler(p):
-    if isinstance(p, modules.processing.StableDiffusionProcessingTxt2Img):
-        return sd_samplers.samplers
-    elif isinstance(p, modules.processing.StableDiffusionProcessingImg2Img):
-        return sd_samplers.samplers_for_img2img
-    elif isinstance(p, modules.api.processing.StableDiffusionProcessingAPI):
-        return sd_samplers.samplers
 
 class StableDiffusionProcessing():
     """
     The first set of paramaters: sd_models -> do_not_reload_embeddings represent the minimum required to create a StableDiffusionProcessing
     """
-    def __init__(self, sd_model=None, outpath_samples=None, outpath_grids=None, prompt: str = "", styles: List[str] = None, seed: int = -1, subseed: int = -1, subseed_strength: float = 0, seed_resize_from_h: int = -1, seed_resize_from_w: int = -1, seed_enable_extras: bool = True, sampler_index: int = 0, batch_size: int = 1, n_iter: int = 1, steps: int = 50, cfg_scale: float = 7.0, width: int = 512, height: int = 512, restore_faces: bool = False, tiling: bool = False, do_not_save_samples: bool = False, do_not_save_grid: bool = False, extra_generation_params: Dict[Any, Any] = None, overlay_images: Any = None, negative_prompt: str = None, eta: float = None, do_not_reload_embeddings: bool = False, denoising_strength: float = 0, ddim_discretize: str = None, s_churn: float = 0.0, s_tmax: float = None, s_tmin: float = 0.0, s_noise: float = 1.0, override_settings: Dict[str, Any] = None):
+    def __init__(self, sd_model=None, outpath_samples=None, outpath_grids=None, prompt: str = "", styles: List[str] = None, seed: int = -1, subseed: int = -1, subseed_strength: float = 0, seed_resize_from_h: int = -1, seed_resize_from_w: int = -1, seed_enable_extras: bool = True, sampler_name: str = None, batch_size: int = 1, n_iter: int = 1, steps: int = 50, cfg_scale: float = 7.0, width: int = 512, height: int = 512, restore_faces: bool = False, tiling: bool = False, do_not_save_samples: bool = False, do_not_save_grid: bool = False, extra_generation_params: Dict[Any, Any] = None, overlay_images: Any = None, negative_prompt: str = None, eta: float = None, do_not_reload_embeddings: bool = False, denoising_strength: float = 0, ddim_discretize: str = None, s_churn: float = 0.0, s_tmax: float = None, s_tmin: float = 0.0, s_noise: float = 1.0, override_settings: Dict[str, Any] = None, sampler_index: int = None):
+        if sampler_index is not None:
+            warnings.warn("sampler_index argument for StableDiffusionProcessing does not do anything; use sampler_name")
+
         self.sd_model = sd_model
         self.outpath_samples: str = outpath_samples
         self.outpath_grids: str = outpath_grids
@@ -91,7 +88,7 @@ class StableDiffusionProcessing():
         self.subseed_strength: float = subseed_strength
         self.seed_resize_from_h: int = seed_resize_from_h
         self.seed_resize_from_w: int = seed_resize_from_w
-        self.sampler_index: int = sampler_index
+        self.sampler_name: str = sampler_name
         self.batch_size: int = batch_size
         self.n_iter: int = n_iter
         self.steps: int = steps
@@ -210,8 +207,7 @@ class Processed:
         self.info = info
         self.width = p.width
         self.height = p.height
-        self.sampler_index = p.sampler_index
-        self.sampler = sd_samplers.samplers[p.sampler_index].name
+        self.sampler_name = p.sampler_name
         self.cfg_scale = p.cfg_scale
         self.steps = p.steps
         self.batch_size = p.batch_size
@@ -256,8 +252,7 @@ class Processed:
             "subseed_strength": self.subseed_strength,
             "width": self.width,
             "height": self.height,
-            "sampler_index": self.sampler_index,
-            "sampler": self.sampler,
+            "sampler_name": self.sampler_name,
             "cfg_scale": self.cfg_scale,
             "steps": self.steps,
             "batch_size": self.batch_size,
@@ -384,7 +379,7 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration
 
     generation_params = {
         "Steps": p.steps,
-        "Sampler": get_correct_sampler(p)[p.sampler_index].name,
+        "Sampler": p.sampler_name,
         "CFG scale": p.cfg_scale,
         "Seed": all_seeds[index],
         "Face restoration": (opts.face_restoration_model if p.restore_faces else None),
@@ -645,7 +640,7 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing):
             self.truncate_y = int(self.firstphase_height - firstphase_height_truncated) // opt_f
 
     def sample(self, conditioning, unconditional_conditioning, seeds, subseeds, subseed_strength, prompts):
-        self.sampler = sd_samplers.create_sampler_with_index(sd_samplers.samplers, self.sampler_index, self.sd_model)
+        self.sampler = sd_samplers.create_sampler(self.sampler_name, self.sd_model)
 
         if not self.enable_hr:
             x = create_random_tensors([opt_C, self.height // opt_f, self.width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self)
@@ -706,7 +701,7 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing):
 
         shared.state.nextjob()
 
-        self.sampler = sd_samplers.create_sampler_with_index(sd_samplers.samplers, self.sampler_index, self.sd_model)
+        self.sampler = sd_samplers.create_sampler(self.sampler_name, self.sd_model)
 
         noise = create_random_tensors(samples.shape[1:], seeds=seeds, subseeds=subseeds, subseed_strength=subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self)
 
@@ -743,7 +738,7 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing):
         self.image_conditioning = None
 
     def init(self, all_prompts, all_seeds, all_subseeds):
-        self.sampler = sd_samplers.create_sampler_with_index(sd_samplers.samplers_for_img2img, self.sampler_index, self.sd_model)
+        self.sampler = sd_samplers.create_sampler(self.sampler_name, self.sd_model)
         crop_region = None
 
         if self.image_mask is not None:
diff --git a/modules/sd_samplers.py b/modules/sd_samplers.py
index 783992d2..4fe67854 100644
--- a/modules/sd_samplers.py
+++ b/modules/sd_samplers.py
@@ -46,16 +46,23 @@ all_samplers = [
     SamplerData('DDIM', lambda model: VanillaStableDiffusionSampler(ldm.models.diffusion.ddim.DDIMSampler, model), [], {}),
     SamplerData('PLMS', lambda model: VanillaStableDiffusionSampler(ldm.models.diffusion.plms.PLMSSampler, model), [], {}),
 ]
+all_samplers_map = {x.name: x for x in all_samplers}
 
 samplers = []
 samplers_for_img2img = []
 
 
-def create_sampler_with_index(list_of_configs, index, model):
-    config = list_of_configs[index]
+def create_sampler(name, model):
+    if name is not None:
+        config = all_samplers_map.get(name, None)
+    else:
+        config = all_samplers[0]
+
+    assert config is not None, f'bad sampler name: {name}'
+
     sampler = config.constructor(model)
     sampler.config = config
-    
+
     return sampler
 
 
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index 0aeb0459..5e4d8688 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -10,7 +10,7 @@ import csv
 
 from PIL import Image, PngImagePlugin
 
-from modules import shared, devices, sd_hijack, processing, sd_models, images
+from modules import shared, devices, sd_hijack, processing, sd_models, images, sd_samplers
 import modules.textual_inversion.dataset
 from modules.textual_inversion.learn_schedule import LearnRateScheduler
 
@@ -345,7 +345,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc
                 p.prompt = preview_prompt
                 p.negative_prompt = preview_negative_prompt
                 p.steps = preview_steps
-                p.sampler_index = preview_sampler_index
+                p.sampler_name = sd_samplers.samplers[preview_sampler_index].name
                 p.cfg_scale = preview_cfg_scale
                 p.seed = preview_seed
                 p.width = preview_width
diff --git a/modules/txt2img.py b/modules/txt2img.py
index 8e4e8677..c8f81176 100644
--- a/modules/txt2img.py
+++ b/modules/txt2img.py
@@ -1,4 +1,5 @@
 import modules.scripts
+from modules import sd_samplers
 from modules.processing import StableDiffusionProcessing, Processed, StableDiffusionProcessingTxt2Img, \
     StableDiffusionProcessingImg2Img, process_images
 from modules.shared import opts, cmd_opts
@@ -21,7 +22,7 @@ def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2:
         seed_resize_from_h=seed_resize_from_h,
         seed_resize_from_w=seed_resize_from_w,
         seed_enable_extras=seed_enable_extras,
-        sampler_index=sampler_index,
+        sampler_name=sd_samplers.samplers[sampler_index].name,
         batch_size=batch_size,
         n_iter=n_iter,
         steps=steps,
diff --git a/modules/ui.py b/modules/ui.py
index 5dce7f3b..2d488741 100644
--- a/modules/ui.py
+++ b/modules/ui.py
@@ -142,7 +142,7 @@ def save_files(js_data, images, do_make_zip, index):
                 filenames.append(os.path.basename(txt_fullfn))
                 fullfns.append(txt_fullfn)
 
-        writer.writerow([data["prompt"], data["seed"], data["width"], data["height"], data["sampler"], data["cfg_scale"], data["steps"], filenames[0], data["negative_prompt"]])
+        writer.writerow([data["prompt"], data["seed"], data["width"], data["height"], data["sampler_name"], data["cfg_scale"], data["steps"], filenames[0], data["negative_prompt"]])
 
     # Make Zip
     if do_make_zip:
diff --git a/scripts/img2imgalt.py b/scripts/img2imgalt.py
index 964b75c7..1229f61b 100644
--- a/scripts/img2imgalt.py
+++ b/scripts/img2imgalt.py
@@ -157,7 +157,7 @@ class Script(scripts.Script):
     def run(self, p, _, override_sampler, override_prompt, original_prompt, original_negative_prompt, override_steps, st, override_strength, cfg, randomness, sigma_adjustment):
         # Override
         if override_sampler:
-            p.sampler_index = [sampler.name for sampler in sd_samplers.samplers].index("Euler")
+            p.sampler_name = "Euler"
         if override_prompt:
             p.prompt = original_prompt
             p.negative_prompt = original_negative_prompt
@@ -191,7 +191,7 @@ class Script(scripts.Script):
             
             combined_noise = ((1 - randomness) * rec_noise + randomness * rand_noise) / ((randomness**2 + (1-randomness)**2) ** 0.5)
             
-            sampler = sd_samplers.create_sampler_with_index(sd_samplers.samplers, p.sampler_index, p.sd_model)
+            sampler = sd_samplers.create_sampler(p.sampler_name, p.sd_model)
 
             sigmas = sampler.model_wrap.get_sigmas(p.steps)
             
diff --git a/scripts/xy_grid.py b/scripts/xy_grid.py
index 417ed0d4..b0b9d84d 100644
--- a/scripts/xy_grid.py
+++ b/scripts/xy_grid.py
@@ -10,9 +10,9 @@ import numpy as np
 import modules.scripts as scripts
 import gradio as gr
 
-from modules import images
+from modules import images, sd_samplers
 from modules.hypernetworks import hypernetwork
-from modules.processing import process_images, Processed, get_correct_sampler, StableDiffusionProcessingTxt2Img
+from modules.processing import process_images, Processed, StableDiffusionProcessingTxt2Img
 from modules.shared import opts, cmd_opts, state
 import modules.shared as shared
 import modules.sd_samplers
@@ -60,9 +60,9 @@ def apply_order(p, x, xs):
     p.prompt = prompt_tmp + p.prompt
     
 
-def build_samplers_dict(p):
+def build_samplers_dict():
     samplers_dict = {}
-    for i, sampler in enumerate(get_correct_sampler(p)):
+    for i, sampler in enumerate(sd_samplers.all_samplers):
         samplers_dict[sampler.name.lower()] = i
         for alias in sampler.aliases:
             samplers_dict[alias.lower()] = i
@@ -70,7 +70,7 @@ def build_samplers_dict(p):
 
 
 def apply_sampler(p, x, xs):
-    sampler_index = build_samplers_dict(p).get(x.lower(), None)
+    sampler_index = build_samplers_dict().get(x.lower(), None)
     if sampler_index is None:
         raise RuntimeError(f"Unknown sampler: {x}")
 
@@ -78,7 +78,7 @@ def apply_sampler(p, x, xs):
 
 
 def confirm_samplers(p, xs):
-    samplers_dict = build_samplers_dict(p)
+    samplers_dict = build_samplers_dict()
     for x in xs:
         if x.lower() not in samplers_dict.keys():
             raise RuntimeError(f"Unknown sampler: {x}")
-- 
cgit v1.2.3


From bd68e35de3b7cf7547ed97d8bdf60147402133cc Mon Sep 17 00:00:00 2001
From: flamelaw <flamelaw.com3d2@gmail.com>
Date: Sun, 20 Nov 2022 12:35:26 +0900
Subject: Gradient accumulation, autocast fix, new latent sampling method, etc

---
 modules/hypernetworks/hypernetwork.py          | 269 +++++++++++----------
 modules/sd_hijack.py                           |   9 +-
 modules/sd_hijack_checkpoint.py                |  10 +
 modules/shared.py                              |   3 +-
 modules/textual_inversion/dataset.py           | 134 +++++++----
 modules/textual_inversion/textual_inversion.py | 320 ++++++++++++++-----------
 modules/ui.py                                  |  16 +-
 7 files changed, 448 insertions(+), 313 deletions(-)
 create mode 100644 modules/sd_hijack_checkpoint.py

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py
index fbb87dd1..3d3301b0 100644
--- a/modules/hypernetworks/hypernetwork.py
+++ b/modules/hypernetworks/hypernetwork.py
@@ -367,13 +367,13 @@ def report_statistics(loss_info:dict):
 
 
 
-def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
+def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, steps, shuffle_tags, tag_drop_out, latent_sampling_method, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
     # images allows training previews to have infotext. Importing it at the top causes a circular import problem.
     from modules import images
 
     save_hypernetwork_every = save_hypernetwork_every or 0
     create_image_every = create_image_every or 0
-    textual_inversion.validate_train_inputs(hypernetwork_name, learn_rate, batch_size, data_root, template_file, steps, save_hypernetwork_every, create_image_every, log_directory, name="hypernetwork")
+    textual_inversion.validate_train_inputs(hypernetwork_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_hypernetwork_every, create_image_every, log_directory, name="hypernetwork")
 
     path = shared.hypernetworks.get(hypernetwork_name, None)
     shared.loaded_hypernetwork = Hypernetwork()
@@ -403,28 +403,24 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
     hypernetwork = shared.loaded_hypernetwork
     checkpoint = sd_models.select_checkpoint()
 
-    ititial_step = hypernetwork.step or 0
-    if ititial_step >= steps:
+    initial_step = hypernetwork.step or 0
+    if initial_step >= steps:
         shared.state.textinfo = f"Model has already been trained beyond specified max steps"
         return hypernetwork, filename
 
-    scheduler = LearnRateScheduler(learn_rate, steps, ititial_step)
-    
+    scheduler = LearnRateScheduler(learn_rate, steps, initial_step)
+
     # dataset loading may take a while, so input validations and early returns should be done before this
     shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..."
-    with torch.autocast("cuda"):
-        ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=hypernetwork_name, model=shared.sd_model, device=devices.device, template_file=template_file, include_cond=True, batch_size=batch_size)
+
+    pin_memory = shared.opts.pin_memory
+
+    ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=hypernetwork_name, model=shared.sd_model, cond_model=shared.sd_model.cond_stage_model, device=devices.device, template_file=template_file, include_cond=True, batch_size=batch_size, gradient_step=gradient_step, shuffle_tags=shuffle_tags, tag_drop_out=tag_drop_out, latent_sampling_method=latent_sampling_method)
+    dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, batch_size=ds.batch_size, pin_memory=pin_memory)
 
     if unload:
         shared.sd_model.cond_stage_model.to(devices.cpu)
         shared.sd_model.first_stage_model.to(devices.cpu)
-
-    size = len(ds.indexes)
-    loss_dict = defaultdict(lambda : deque(maxlen = 1024))
-    losses = torch.zeros((size,))
-    previous_mean_losses = [0]
-    previous_mean_loss = 0
-    print("Mean loss of {} elements".format(size))
     
     weights = hypernetwork.weights()
     for weight in weights:
@@ -436,8 +432,8 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
         optimizer_name = hypernetwork.optimizer_name
     else:
         print(f"Optimizer type {hypernetwork.optimizer_name} is not defined!")
-        optimizer = torch.optim.AdamW(params=weights, lr=scheduler.learn_rate)
-        optimizer_name = 'AdamW'
+    optimizer = torch.optim.AdamW(params=weights, lr=scheduler.learn_rate)
+    optimizer_name = 'AdamW'
 
     if hypernetwork.optimizer_state_dict:  # This line must be changed if Optimizer type can be different from saved optimizer.
         try:
@@ -446,131 +442,155 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
             print("Cannot resume from saved optimizer!")
             print(e)
 
+    scaler = torch.cuda.amp.GradScaler()
+    
+    batch_size = ds.batch_size
+    gradient_step = ds.gradient_step
+    # n steps = batch_size * gradient_step * n image processed
+    steps_per_epoch = len(ds) // batch_size // gradient_step
+    max_steps_per_epoch = len(ds) // batch_size - (len(ds) // batch_size) % gradient_step
+    loss_step = 0
+    _loss_step = 0 #internal
+    # size = len(ds.indexes)
+    # loss_dict = defaultdict(lambda : deque(maxlen = 1024))
+    # losses = torch.zeros((size,))
+    # previous_mean_losses = [0]
+    # previous_mean_loss = 0
+    # print("Mean loss of {} elements".format(size))
+
     steps_without_grad = 0
 
     last_saved_file = "<none>"
     last_saved_image = "<none>"
     forced_filename = "<none>"
 
-    pbar = tqdm.tqdm(enumerate(ds), total=steps - ititial_step)
-    for i, entries in pbar:
-        hypernetwork.step = i + ititial_step
-        if len(loss_dict) > 0:
-            previous_mean_losses = [i[-1] for i in loss_dict.values()]
-            previous_mean_loss = mean(previous_mean_losses)
-            
-        scheduler.apply(optimizer, hypernetwork.step)
-        if scheduler.finished:
-            break
-
-        if shared.state.interrupted:
-            break
-
-        with torch.autocast("cuda"):
-            c = stack_conds([entry.cond for entry in entries]).to(devices.device)
-            # c = torch.vstack([entry.cond for entry in entries]).to(devices.device)
-            x = torch.stack([entry.latent for entry in entries]).to(devices.device)
-            loss = shared.sd_model(x, c)[0]
-            del x
-            del c
-
-            losses[hypernetwork.step % losses.shape[0]] = loss.item()
-            for entry in entries:
-                loss_dict[entry.filename].append(loss.item())
+    pbar = tqdm.tqdm(total=steps - initial_step)
+    try:
+        for i in range((steps-initial_step) * gradient_step):
+            if scheduler.finished:
+                break
+            if shared.state.interrupted:
+                break
+            for j, batch in enumerate(dl):
+                # works as a drop_last=True for gradient accumulation
+                if j == max_steps_per_epoch:
+                    break
+                scheduler.apply(optimizer, hypernetwork.step)
+                if scheduler.finished:
+                    break
+                if shared.state.interrupted:
+                    break
+
+                with torch.autocast("cuda"):
+                    x = batch.latent_sample.to(devices.device, non_blocking=pin_memory)
+                    if tag_drop_out != 0 or shuffle_tags:
+                        shared.sd_model.cond_stage_model.to(devices.device)
+                        c = shared.sd_model.cond_stage_model(batch.cond_text).to(devices.device, non_blocking=pin_memory)
+                        shared.sd_model.cond_stage_model.to(devices.cpu)
+                    else:
+                        c = stack_conds(batch.cond).to(devices.device, non_blocking=pin_memory)
+                    loss = shared.sd_model(x, c)[0] / gradient_step
+                    del x
+                    del c
+
+                    _loss_step += loss.item()
+                scaler.scale(loss).backward()
+                # go back until we reach gradient accumulation steps
+                if (j + 1) % gradient_step != 0:
+                    continue
+                # print(f"grad:{weights[0].grad.detach().cpu().abs().mean().item():.7f}")
+                # scaler.unscale_(optimizer)
+                # print(f"grad:{weights[0].grad.detach().cpu().abs().mean().item():.15f}")
+                # torch.nn.utils.clip_grad_norm_(weights, max_norm=1.0)
+                # print(f"grad:{weights[0].grad.detach().cpu().abs().mean().item():.15f}")
+                scaler.step(optimizer)
+                scaler.update()
+                hypernetwork.step += 1
+                pbar.update()
+                optimizer.zero_grad(set_to_none=True)
+                loss_step = _loss_step
+                _loss_step = 0
+
+                steps_done = hypernetwork.step + 1
                 
-            optimizer.zero_grad()
-            weights[0].grad = None
-            loss.backward()
-
-            if weights[0].grad is None:
-                steps_without_grad += 1
-            else:
-                steps_without_grad = 0
-            assert steps_without_grad < 10, 'no gradient found for the trained weight after backward() for 10 steps in a row; this is a bug; training cannot continue'
-
-            optimizer.step()
-
-        steps_done = hypernetwork.step + 1
-
-        if torch.isnan(losses[hypernetwork.step % losses.shape[0]]): 
-            raise RuntimeError("Loss diverged.")
-        
-        if len(previous_mean_losses) > 1:
-            std = stdev(previous_mean_losses)
-        else:
-            std = 0
-        dataset_loss_info = f"dataset loss:{mean(previous_mean_losses):.3f}" + u"\u00B1" + f"({std / (len(previous_mean_losses) ** 0.5):.3f})"
-        pbar.set_description(dataset_loss_info)
-
-        if hypernetwork_dir is not None and steps_done % save_hypernetwork_every == 0:
-            # Before saving, change name to match current checkpoint.
-            hypernetwork_name_every = f'{hypernetwork_name}-{steps_done}'
-            last_saved_file = os.path.join(hypernetwork_dir, f'{hypernetwork_name_every}.pt')
-            hypernetwork.optimizer_name = optimizer_name
-            if shared.opts.save_optimizer_state:
-                hypernetwork.optimizer_state_dict = optimizer.state_dict()
-            save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, last_saved_file)
-            hypernetwork.optimizer_state_dict = None  # dereference it after saving, to save memory.
-
-        textual_inversion.write_loss(log_directory, "hypernetwork_loss.csv", hypernetwork.step, len(ds), {
-            "loss": f"{previous_mean_loss:.7f}",
-            "learn_rate": scheduler.learn_rate
-        })
-
-        if images_dir is not None and steps_done % create_image_every == 0:
-            forced_filename = f'{hypernetwork_name}-{steps_done}'
-            last_saved_image = os.path.join(images_dir, forced_filename)
-
-            optimizer.zero_grad()
-            shared.sd_model.cond_stage_model.to(devices.device)
-            shared.sd_model.first_stage_model.to(devices.device)
-
-            p = processing.StableDiffusionProcessingTxt2Img(
-                sd_model=shared.sd_model,
-                do_not_save_grid=True,
-                do_not_save_samples=True,
-            )
-
-            if preview_from_txt2img:
-                p.prompt = preview_prompt
-                p.negative_prompt = preview_negative_prompt
-                p.steps = preview_steps
-                p.sampler_name = sd_samplers.samplers[preview_sampler_index].name
-                p.cfg_scale = preview_cfg_scale
-                p.seed = preview_seed
-                p.width = preview_width
-                p.height = preview_height
-            else:
-                p.prompt = entries[0].cond_text
-                p.steps = 20
+                epoch_num = hypernetwork.step // steps_per_epoch
+                epoch_step = hypernetwork.step % steps_per_epoch
+
+                pbar.set_description(f"[Epoch {epoch_num}: {epoch_step+1}/{steps_per_epoch}]loss: {loss_step:.7f}")
+                if hypernetwork_dir is not None and steps_done % save_hypernetwork_every == 0:
+                    # Before saving, change name to match current checkpoint.
+                    hypernetwork_name_every = f'{hypernetwork_name}-{steps_done}'
+                    last_saved_file = os.path.join(hypernetwork_dir, f'{hypernetwork_name_every}.pt')
+                    hypernetwork.optimizer_name = optimizer_name
+                    if shared.opts.save_optimizer_state:
+                        hypernetwork.optimizer_state_dict = optimizer.state_dict()
+                    save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, last_saved_file)
+                    hypernetwork.optimizer_state_dict = None  # dereference it after saving, to save memory.
+
+                textual_inversion.write_loss(log_directory, "hypernetwork_loss.csv", hypernetwork.step, steps_per_epoch, {
+                    "loss": f"{loss_step:.7f}",
+                    "learn_rate": scheduler.learn_rate
+                })
+
+                if images_dir is not None and steps_done % create_image_every == 0:
+                    forced_filename = f'{hypernetwork_name}-{steps_done}'
+                    last_saved_image = os.path.join(images_dir, forced_filename)
+
+                    shared.sd_model.cond_stage_model.to(devices.device)
+                    shared.sd_model.first_stage_model.to(devices.device)
+
+                    p = processing.StableDiffusionProcessingTxt2Img(
+                        sd_model=shared.sd_model,
+                        do_not_save_grid=True,
+                        do_not_save_samples=True,
+                    )
+
+                    if preview_from_txt2img:
+                        p.prompt = preview_prompt
+                        p.negative_prompt = preview_negative_prompt
+                        p.steps = preview_steps
+                        p.sampler_name = sd_samplers.samplers[preview_sampler_index].name
+                        p.cfg_scale = preview_cfg_scale
+                        p.seed = preview_seed
+                        p.width = preview_width
+                        p.height = preview_height
+                    else:
+                        p.prompt = batch.cond_text[0]
+                        p.steps = 20
+                        p.width = training_width
+                        p.height = training_height
 
-            preview_text = p.prompt
+                    preview_text = p.prompt
 
-            processed = processing.process_images(p)
-            image = processed.images[0] if len(processed.images)>0 else None
+                    processed = processing.process_images(p)
+                    image = processed.images[0] if len(processed.images) > 0 else None
 
-            if unload:
-                shared.sd_model.cond_stage_model.to(devices.cpu)
-                shared.sd_model.first_stage_model.to(devices.cpu)
+                    if unload:
+                        shared.sd_model.cond_stage_model.to(devices.cpu)
+                        shared.sd_model.first_stage_model.to(devices.cpu)
 
-            if image is not None:
-                shared.state.current_image = image
-                last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False)
-                last_saved_image += f", prompt: {preview_text}"
+                    if image is not None:
+                        shared.state.current_image = image
+                        last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False)
+                        last_saved_image += f", prompt: {preview_text}"
 
-        shared.state.job_no = hypernetwork.step
+                shared.state.job_no = hypernetwork.step
 
-        shared.state.textinfo = f"""
+                shared.state.textinfo = f"""
 <p>
-Loss: {previous_mean_loss:.7f}<br/>
+Loss: {loss_step:.7f}<br/>
 Step: {hypernetwork.step}<br/>
-Last prompt: {html.escape(entries[0].cond_text)}<br/>
+Last prompt: {html.escape(batch.cond_text[0])}<br/>
 Last saved hypernetwork: {html.escape(last_saved_file)}<br/>
 Last saved image: {html.escape(last_saved_image)}<br/>
 </p>
 """
-        
-    report_statistics(loss_dict)
+    except Exception:
+        print(traceback.format_exc(), file=sys.stderr)
+    finally:
+        pbar.leave = False
+        pbar.close()
+        #report_statistics(loss_dict)
 
     filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt')
     hypernetwork.optimizer_name = optimizer_name
@@ -579,6 +599,9 @@ Last saved image: {html.escape(last_saved_image)}<br/>
     save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename)
     del optimizer
     hypernetwork.optimizer_state_dict = None  # dereference it after saving, to save memory.
+    shared.sd_model.cond_stage_model.to(devices.device)
+    shared.sd_model.first_stage_model.to(devices.device)
+
     return hypernetwork, filename
 
 def save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename):
diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index eaedac13..29c8b561 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -8,7 +8,7 @@ from torch import einsum
 from torch.nn.functional import silu
 
 import modules.textual_inversion.textual_inversion
-from modules import prompt_parser, devices, sd_hijack_optimizations, shared
+from modules import prompt_parser, devices, sd_hijack_optimizations, shared, sd_hijack_checkpoint
 from modules.shared import opts, device, cmd_opts
 from modules.sd_hijack_optimizations import invokeAI_mps_available
 
@@ -59,6 +59,10 @@ def undo_optimizations():
 def get_target_prompt_token_count(token_count):
     return math.ceil(max(token_count, 1) / 75) * 75
 
+def fix_checkpoint():
+    ldm.modules.attention.BasicTransformerBlock.forward = sd_hijack_checkpoint.BasicTransformerBlock_forward
+    ldm.modules.diffusionmodules.openaimodel.ResBlock.forward = sd_hijack_checkpoint.ResBlock_forward
+    ldm.modules.diffusionmodules.openaimodel.AttentionBlock.forward = sd_hijack_checkpoint.AttentionBlock_forward
 
 class StableDiffusionModelHijack:
     fixes = None
@@ -78,6 +82,7 @@ class StableDiffusionModelHijack:
         self.clip = m.cond_stage_model
 
         apply_optimizations()
+        fix_checkpoint()
 
         def flatten(el):
             flattened = [flatten(children) for children in el.children()]
@@ -303,7 +308,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
             batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text_old(text)
         else:
             batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text(text)
-
+        
         self.hijack.comments += hijack_comments
 
         if len(used_custom_terms) > 0:
diff --git a/modules/sd_hijack_checkpoint.py b/modules/sd_hijack_checkpoint.py
new file mode 100644
index 00000000..5712972f
--- /dev/null
+++ b/modules/sd_hijack_checkpoint.py
@@ -0,0 +1,10 @@
+from torch.utils.checkpoint import checkpoint
+
+def BasicTransformerBlock_forward(self, x, context=None):
+    return checkpoint(self._forward, x, context)
+
+def AttentionBlock_forward(self, x):
+    return checkpoint(self._forward, x)
+
+def ResBlock_forward(self, x, emb):
+    return checkpoint(self._forward, x, emb)
\ No newline at end of file
diff --git a/modules/shared.py b/modules/shared.py
index a4457305..3704ce23 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -322,8 +322,7 @@ options_templates.update(options_section(('system', "System"), {
 
 options_templates.update(options_section(('training', "Training"), {
     "unload_models_when_training": OptionInfo(False, "Move VAE and CLIP to RAM when training if possible. Saves VRAM."),
-    "shuffle_tags": OptionInfo(False, "Shuffleing tags by ',' when create texts."),
-    "tag_drop_out": OptionInfo(0, "Dropout tags when create texts", gr.Slider, {"minimum": 0, "maximum": 1, "step": 0.1}),
+    "pin_memory": OptionInfo(False, "Turn on pin_memory for DataLoader. Makes training slightly faster but can increase memory usage."),
     "save_optimizer_state": OptionInfo(False, "Saves Optimizer state as separate *.optim file. Training can be resumed with HN itself and matching optim file."),
     "dataset_filename_word_regex": OptionInfo("", "Filename word regex"),
     "dataset_filename_join_string": OptionInfo(" ", "Filename join string"),
diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py
index eb75c376..d594b49d 100644
--- a/modules/textual_inversion/dataset.py
+++ b/modules/textual_inversion/dataset.py
@@ -3,7 +3,7 @@ import numpy as np
 import PIL
 import torch
 from PIL import Image
-from torch.utils.data import Dataset
+from torch.utils.data import Dataset, DataLoader
 from torchvision import transforms
 
 import random
@@ -11,25 +11,28 @@ import tqdm
 from modules import devices, shared
 import re
 
+from ldm.modules.distributions.distributions import DiagonalGaussianDistribution
+
 re_numbers_at_start = re.compile(r"^[-\d]+\s*")
 
 
 class DatasetEntry:
-    def __init__(self, filename=None, latent=None, filename_text=None):
+    def __init__(self, filename=None, filename_text=None, latent_dist=None, latent_sample=None, cond=None, cond_text=None, pixel_values=None):
         self.filename = filename
-        self.latent = latent
         self.filename_text = filename_text
-        self.cond = None
-        self.cond_text = None
+        self.latent_dist = latent_dist
+        self.latent_sample = latent_sample
+        self.cond = cond
+        self.cond_text = cond_text
+        self.pixel_values = pixel_values
 
 
 class PersonalizedBase(Dataset):
-    def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, device=None, template_file=None, include_cond=False, batch_size=1):
+    def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, cond_model=None, device=None, template_file=None, include_cond=False, batch_size=1, gradient_step=1, shuffle_tags=False, tag_drop_out=0, latent_sampling_method='once'):        
         re_word = re.compile(shared.opts.dataset_filename_word_regex) if len(shared.opts.dataset_filename_word_regex) > 0 else None
-
+        
         self.placeholder_token = placeholder_token
 
-        self.batch_size = batch_size
         self.width = width
         self.height = height
         self.flip = transforms.RandomHorizontalFlip(p=flip_p)
@@ -45,11 +48,16 @@ class PersonalizedBase(Dataset):
         assert os.path.isdir(data_root), "Dataset directory doesn't exist"
         assert os.listdir(data_root), "Dataset directory is empty"
 
-        cond_model = shared.sd_model.cond_stage_model
-
         self.image_paths = [os.path.join(data_root, file_path) for file_path in os.listdir(data_root)]
+
+        
+        self.shuffle_tags = shuffle_tags
+        self.tag_drop_out = tag_drop_out
+
         print("Preparing dataset...")
         for path in tqdm.tqdm(self.image_paths):
+            if shared.state.interrupted:
+                raise Exception("inturrupted")
             try:
                 image = Image.open(path).convert('RGB').resize((self.width, self.height), PIL.Image.BICUBIC)
             except Exception:
@@ -71,37 +79,58 @@ class PersonalizedBase(Dataset):
             npimage = np.array(image).astype(np.uint8)
             npimage = (npimage / 127.5 - 1.0).astype(np.float32)
 
-            torchdata = torch.from_numpy(npimage).to(device=device, dtype=torch.float32)
-            torchdata = torch.moveaxis(torchdata, 2, 0)
-
-            init_latent = model.get_first_stage_encoding(model.encode_first_stage(torchdata.unsqueeze(dim=0))).squeeze()
-            init_latent = init_latent.to(devices.cpu)
-
-            entry = DatasetEntry(filename=path, filename_text=filename_text, latent=init_latent)
-
-            if include_cond:
+            torchdata = torch.from_numpy(npimage).permute(2, 0, 1).to(device=device, dtype=torch.float32)
+            latent_sample = None
+
+            with torch.autocast("cuda"):
+                latent_dist = model.encode_first_stage(torchdata.unsqueeze(dim=0))
+
+            if latent_sampling_method == "once" or (latent_sampling_method == "deterministic" and not isinstance(latent_dist, DiagonalGaussianDistribution)):
+                latent_sample = model.get_first_stage_encoding(latent_dist).squeeze().to(devices.cpu)
+                latent_sampling_method = "once"
+                entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample)
+            elif latent_sampling_method == "deterministic":
+                # Works only for DiagonalGaussianDistribution
+                latent_dist.std = 0
+                latent_sample = model.get_first_stage_encoding(latent_dist).squeeze().to(devices.cpu)
+                entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample)
+            elif latent_sampling_method == "random":
+                entry = DatasetEntry(filename=path, filename_text=filename_text, latent_dist=latent_dist)
+
+            if not (self.tag_drop_out != 0 or self.shuffle_tags):
                 entry.cond_text = self.create_text(filename_text)
-                entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0)
 
-            self.dataset.append(entry)
-
-        assert len(self.dataset) > 0, "No images have been found in the dataset."
-        self.length = len(self.dataset) * repeats // batch_size
+            if include_cond and not (self.tag_drop_out != 0 or self.shuffle_tags):
+                with torch.autocast("cuda"):
+                    entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0)
+            # elif not include_cond:
+            #     _, _, _, _, hijack_fixes, token_count = cond_model.process_text([entry.cond_text])
+            #     max_n = token_count // 75
+            #     index_list = [ [] for _ in range(max_n + 1) ]
+            #     for n, (z, _) in hijack_fixes[0]:
+            #         index_list[n].append(z)
+            #     with torch.autocast("cuda"):
+            #         entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0)
+            #     entry.emb_index = index_list
 
-        self.dataset_length = len(self.dataset)
-        self.indexes = None
-        self.shuffle()
+            self.dataset.append(entry)
+            del torchdata
+            del latent_dist
+            del latent_sample
 
-    def shuffle(self):
-        self.indexes = np.random.permutation(self.dataset_length)
+        self.length = len(self.dataset)
+        assert self.length > 0, "No images have been found in the dataset."
+        self.batch_size = min(batch_size, self.length)
+        self.gradient_step = min(gradient_step, self.length // self.batch_size)
+        self.latent_sampling_method = latent_sampling_method
 
     def create_text(self, filename_text):
         text = random.choice(self.lines)
         text = text.replace("[name]", self.placeholder_token)
         tags = filename_text.split(',')
-        if shared.opts.tag_drop_out != 0:
-            tags = [t for t in tags if random.random() > shared.opts.tag_drop_out]
-        if shared.opts.shuffle_tags:
+        if self.tag_drop_out != 0:
+            tags = [t for t in tags if random.random() > self.tag_drop_out]
+        if self.shuffle_tags:
             random.shuffle(tags)
         text = text.replace("[filewords]", ','.join(tags))
         return text
@@ -110,19 +139,28 @@ class PersonalizedBase(Dataset):
         return self.length
 
     def __getitem__(self, i):
-        res = []
-
-        for j in range(self.batch_size):
-            position = i * self.batch_size + j
-            if position % len(self.indexes) == 0:
-                self.shuffle()
-
-            index = self.indexes[position % len(self.indexes)]
-            entry = self.dataset[index]
-
-            if entry.cond is None:
-                entry.cond_text = self.create_text(entry.filename_text)
-
-            res.append(entry)
-
-        return res
+        entry = self.dataset[i]
+        if self.tag_drop_out != 0 or self.shuffle_tags:
+            entry.cond_text = self.create_text(entry.filename_text)
+        if self.latent_sampling_method == "random":
+            entry.latent_sample = shared.sd_model.get_first_stage_encoding(entry.latent_dist)
+        return entry
+
+class PersonalizedDataLoader(DataLoader):
+    def __init__(self, *args, **kwargs):
+        super(PersonalizedDataLoader, self).__init__(shuffle=True, drop_last=True, *args, **kwargs)
+        self.collate_fn = collate_wrapper
+        
+
+class BatchLoader:
+    def __init__(self, data):
+        self.cond_text = [entry.cond_text for entry in data]
+        self.cond = [entry.cond for entry in data]
+        self.latent_sample = torch.stack([entry.latent_sample for entry in data]).squeeze(1)
+
+    def pin_memory(self):
+        self.latent_sample = self.latent_sample.pin_memory()
+        return self
+
+def collate_wrapper(batch):
+    return BatchLoader(batch)
\ No newline at end of file
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index 5e4d8688..1d5e3a32 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -184,7 +184,7 @@ def write_loss(log_directory, filename, step, epoch_len, values):
     if shared.opts.training_write_csv_every == 0:
         return
 
-    if (step + 1) % shared.opts.training_write_csv_every != 0:
+    if step % shared.opts.training_write_csv_every != 0:
         return
     write_csv_header = False if os.path.exists(os.path.join(log_directory, filename)) else True
 
@@ -194,21 +194,23 @@ def write_loss(log_directory, filename, step, epoch_len, values):
         if write_csv_header:
             csv_writer.writeheader()
 
-        epoch = step // epoch_len
-        epoch_step = step % epoch_len 
+        epoch = (step - 1) // epoch_len
+        epoch_step = (step - 1) % epoch_len 
 
         csv_writer.writerow({
-            "step": step + 1,
+            "step": step,
             "epoch": epoch,
-            "epoch_step": epoch_step + 1,
+            "epoch_step": epoch_step,
             **values,
         })
 
-def validate_train_inputs(model_name, learn_rate, batch_size, data_root, template_file, steps, save_model_every, create_image_every, log_directory, name="embedding"):
+def validate_train_inputs(model_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_model_every, create_image_every, log_directory, name="embedding"):
     assert model_name, f"{name} not selected"
     assert learn_rate, "Learning rate is empty or 0"
     assert isinstance(batch_size, int), "Batch size must be integer"
     assert batch_size > 0, "Batch size must be positive"
+    assert isinstance(gradient_step, int), "Gradient accumulation step must be integer"
+    assert gradient_step > 0, "Gradient accumulation step must be positive"
     assert data_root, "Dataset directory is empty"
     assert os.path.isdir(data_root), "Dataset directory doesn't exist"
     assert os.listdir(data_root), "Dataset directory is empty"
@@ -224,10 +226,10 @@ def validate_train_inputs(model_name, learn_rate, batch_size, data_root, templat
     if save_model_every or create_image_every:
         assert log_directory, "Log directory is empty"
 
-def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
+def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, steps, shuffle_tags, tag_drop_out, latent_sampling_method, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
     save_embedding_every = save_embedding_every or 0
     create_image_every = create_image_every or 0
-    validate_train_inputs(embedding_name, learn_rate, batch_size, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding")
+    validate_train_inputs(embedding_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding")
 
     shared.state.textinfo = "Initializing textual inversion training..."
     shared.state.job_count = steps
@@ -255,161 +257,205 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc
     else:
         images_embeds_dir = None
 
-    cond_model = shared.sd_model.cond_stage_model
-
     hijack = sd_hijack.model_hijack
 
     embedding = hijack.embedding_db.word_embeddings[embedding_name]
     checkpoint = sd_models.select_checkpoint()
 
-    ititial_step = embedding.step or 0
-    if ititial_step >= steps:
+    initial_step = embedding.step or 0
+    if initial_step >= steps:
         shared.state.textinfo = f"Model has already been trained beyond specified max steps"
         return embedding, filename
+    scheduler = LearnRateScheduler(learn_rate, steps, initial_step)
 
-    scheduler = LearnRateScheduler(learn_rate, steps, ititial_step)
-
-    # dataset loading may take a while, so input validations and early returns should be done before this
+   # dataset loading may take a while, so input validations and early returns should be done before this
     shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..."
-    with torch.autocast("cuda"):
-        ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file, batch_size=batch_size)
+    
+    pin_memory = shared.opts.pin_memory
+    
+    ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, cond_model=shared.sd_model.cond_stage_model, device=devices.device, template_file=template_file, batch_size=batch_size, gradient_step=gradient_step, shuffle_tags=shuffle_tags, tag_drop_out=tag_drop_out, latent_sampling_method=latent_sampling_method)
+
+    latent_sampling_method = ds.latent_sampling_method
+
+    dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, batch_size=ds.batch_size, pin_memory=False)
+
     if unload:
         shared.sd_model.first_stage_model.to(devices.cpu)
 
     embedding.vec.requires_grad = True
     optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate)
+    scaler = torch.cuda.amp.GradScaler()
 
-    losses = torch.zeros((32,))
+    batch_size = ds.batch_size
+    gradient_step = ds.gradient_step
+    # n steps = batch_size * gradient_step * n image processed
+    steps_per_epoch = len(ds) // batch_size // gradient_step
+    max_steps_per_epoch = len(ds) // batch_size - (len(ds) // batch_size) % gradient_step
+    loss_step = 0
+    _loss_step = 0 #internal
 
+    
     last_saved_file = "<none>"
     last_saved_image = "<none>"
     forced_filename = "<none>"
     embedding_yet_to_be_embedded = False
-
-    pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step)
-    for i, entries in pbar:
-        embedding.step = i + ititial_step
-
-        scheduler.apply(optimizer, embedding.step)
-        if scheduler.finished:
-            break
-
-        if shared.state.interrupted:
-            break
-
-        with torch.autocast("cuda"):
-            c = cond_model([entry.cond_text for entry in entries])
-            x = torch.stack([entry.latent for entry in entries]).to(devices.device)
-            loss = shared.sd_model(x, c)[0]
-            del x
-
-            losses[embedding.step % losses.shape[0]] = loss.item()
-
-            optimizer.zero_grad()
-            loss.backward()
-            optimizer.step()
-
-        steps_done = embedding.step + 1
-
-        epoch_num = embedding.step // len(ds)
-        epoch_step = embedding.step % len(ds)
-
-        pbar.set_description(f"[Epoch {epoch_num}: {epoch_step+1}/{len(ds)}]loss: {losses.mean():.7f}")
-
-        if embedding_dir is not None and steps_done % save_embedding_every == 0:
-            # Before saving, change name to match current checkpoint.
-            embedding_name_every = f'{embedding_name}-{steps_done}'
-            last_saved_file = os.path.join(embedding_dir, f'{embedding_name_every}.pt')
-            save_embedding(embedding, checkpoint, embedding_name_every, last_saved_file, remove_cached_checksum=True)
-            embedding_yet_to_be_embedded = True
-
-        write_loss(log_directory, "textual_inversion_loss.csv", embedding.step, len(ds), {
-            "loss": f"{losses.mean():.7f}",
-            "learn_rate": scheduler.learn_rate
-        })
-
-        if images_dir is not None and steps_done % create_image_every == 0:
-            forced_filename = f'{embedding_name}-{steps_done}'
-            last_saved_image = os.path.join(images_dir, forced_filename)
-
-            shared.sd_model.first_stage_model.to(devices.device)
-
-            p = processing.StableDiffusionProcessingTxt2Img(
-                sd_model=shared.sd_model,
-                do_not_save_grid=True,
-                do_not_save_samples=True,
-                do_not_reload_embeddings=True,
-            )
-
-            if preview_from_txt2img:
-                p.prompt = preview_prompt
-                p.negative_prompt = preview_negative_prompt
-                p.steps = preview_steps
-                p.sampler_name = sd_samplers.samplers[preview_sampler_index].name
-                p.cfg_scale = preview_cfg_scale
-                p.seed = preview_seed
-                p.width = preview_width
-                p.height = preview_height
-            else:
-                p.prompt = entries[0].cond_text
-                p.steps = 20
-                p.width = training_width
-                p.height = training_height
-
-            preview_text = p.prompt
-
-            processed = processing.process_images(p)
-            image = processed.images[0]
-
-            if unload:
-                shared.sd_model.first_stage_model.to(devices.cpu)
-
-            shared.state.current_image = image
-
-            if save_image_with_stored_embedding and os.path.exists(last_saved_file) and embedding_yet_to_be_embedded:
-
-                last_saved_image_chunks = os.path.join(images_embeds_dir, f'{embedding_name}-{steps_done}.png')
-
-                info = PngImagePlugin.PngInfo()
-                data = torch.load(last_saved_file)
-                info.add_text("sd-ti-embedding", embedding_to_b64(data))
-
-                title = "<{}>".format(data.get('name', '???'))
-
-                try:
-                    vectorSize = list(data['string_to_param'].values())[0].shape[0]
-                except Exception as e:
-                    vectorSize = '?'
-
-                checkpoint = sd_models.select_checkpoint()
-                footer_left = checkpoint.model_name
-                footer_mid = '[{}]'.format(checkpoint.hash)
-                footer_right = '{}v {}s'.format(vectorSize, steps_done)
-
-                captioned_image = caption_image_overlay(image, title, footer_left, footer_mid, footer_right)
-                captioned_image = insert_image_data_embed(captioned_image, data)
-
-                captioned_image.save(last_saved_image_chunks, "PNG", pnginfo=info)
-                embedding_yet_to_be_embedded = False
-
-            last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False)
-            last_saved_image += f", prompt: {preview_text}"
-
-        shared.state.job_no = embedding.step
-
-        shared.state.textinfo = f"""
+    
+    pbar = tqdm.tqdm(total=steps - initial_step)
+    try:
+        for i in range((steps-initial_step) * gradient_step):
+            if scheduler.finished:
+                break
+            if shared.state.interrupted:
+                break
+            for j, batch in enumerate(dl):
+                # works as a drop_last=True for gradient accumulation
+                if j == max_steps_per_epoch:
+                    break
+                scheduler.apply(optimizer, embedding.step)
+                if scheduler.finished:
+                    break
+                if shared.state.interrupted:
+                    break
+
+                with torch.autocast("cuda"):
+                    # c = stack_conds(batch.cond).to(devices.device)
+                    # mask = torch.tensor(batch.emb_index).to(devices.device, non_blocking=pin_memory)
+                    # print(mask)
+                    # c[:, 1:1+embedding.vec.shape[0]] = embedding.vec.to(devices.device, non_blocking=pin_memory)
+                    x = batch.latent_sample.to(devices.device, non_blocking=pin_memory)
+                    c = shared.sd_model.cond_stage_model(batch.cond_text)
+                    loss = shared.sd_model(x, c)[0] / gradient_step
+                    del x
+                    
+                    _loss_step += loss.item()
+                scaler.scale(loss).backward()
+                
+                # go back until we reach gradient accumulation steps
+                if (j + 1) % gradient_step != 0:
+                    continue
+                #print(f"grad:{embedding.vec.grad.detach().cpu().abs().mean().item():.7f}")
+                #scaler.unscale_(optimizer)
+                #print(f"grad:{embedding.vec.grad.detach().cpu().abs().mean().item():.7f}")
+                #torch.nn.utils.clip_grad_norm_(embedding.vec, max_norm=1.0)
+                #print(f"grad:{embedding.vec.grad.detach().cpu().abs().mean().item():.7f}")
+                scaler.step(optimizer)
+                scaler.update()
+                embedding.step += 1
+                pbar.update()
+                optimizer.zero_grad(set_to_none=True)
+                loss_step = _loss_step
+                _loss_step = 0
+
+                steps_done = embedding.step + 1
+
+                epoch_num = embedding.step // steps_per_epoch
+                epoch_step = embedding.step % steps_per_epoch
+
+                pbar.set_description(f"[Epoch {epoch_num}: {epoch_step+1}/{steps_per_epoch}]loss: {loss_step:.7f}")
+                if embedding_dir is not None and steps_done % save_embedding_every == 0:
+                    # Before saving, change name to match current checkpoint.
+                    embedding_name_every = f'{embedding_name}-{steps_done}'
+                    last_saved_file = os.path.join(embedding_dir, f'{embedding_name_every}.pt')
+                    #if shared.opts.save_optimizer_state:
+                        #embedding.optimizer_state_dict = optimizer.state_dict()
+                    save_embedding(embedding, checkpoint, embedding_name_every, last_saved_file, remove_cached_checksum=True)
+                    embedding_yet_to_be_embedded = True
+
+                write_loss(log_directory, "textual_inversion_loss.csv", embedding.step, steps_per_epoch, {
+                    "loss": f"{loss_step:.7f}",
+                    "learn_rate": scheduler.learn_rate
+                })
+
+                if images_dir is not None and steps_done % create_image_every == 0:
+                    forced_filename = f'{embedding_name}-{steps_done}'
+                    last_saved_image = os.path.join(images_dir, forced_filename)
+
+                    shared.sd_model.first_stage_model.to(devices.device)
+
+                    p = processing.StableDiffusionProcessingTxt2Img(
+                        sd_model=shared.sd_model,
+                        do_not_save_grid=True,
+                        do_not_save_samples=True,
+                        do_not_reload_embeddings=True,
+                    )
+
+                    if preview_from_txt2img:
+                        p.prompt = preview_prompt
+                        p.negative_prompt = preview_negative_prompt
+                        p.steps = preview_steps
+                        p.sampler_name = sd_samplers.samplers[preview_sampler_index].name
+                        p.cfg_scale = preview_cfg_scale
+                        p.seed = preview_seed
+                        p.width = preview_width
+                        p.height = preview_height
+                    else:
+                        p.prompt = batch.cond_text[0]
+                        p.steps = 20
+                        p.width = training_width
+                        p.height = training_height
+
+                    preview_text = p.prompt
+
+                    processed = processing.process_images(p)
+                    image = processed.images[0] if len(processed.images) > 0 else None
+
+                    if unload:
+                        shared.sd_model.first_stage_model.to(devices.cpu)
+
+                    if image is not None:
+                        shared.state.current_image = image
+                        last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False)
+                        last_saved_image += f", prompt: {preview_text}"
+
+                    if save_image_with_stored_embedding and os.path.exists(last_saved_file) and embedding_yet_to_be_embedded:
+
+                        last_saved_image_chunks = os.path.join(images_embeds_dir, f'{embedding_name}-{steps_done}.png')
+
+                        info = PngImagePlugin.PngInfo()
+                        data = torch.load(last_saved_file)
+                        info.add_text("sd-ti-embedding", embedding_to_b64(data))
+
+                        title = "<{}>".format(data.get('name', '???'))
+
+                        try:
+                            vectorSize = list(data['string_to_param'].values())[0].shape[0]
+                        except Exception as e:
+                            vectorSize = '?'
+
+                        checkpoint = sd_models.select_checkpoint()
+                        footer_left = checkpoint.model_name
+                        footer_mid = '[{}]'.format(checkpoint.hash)
+                        footer_right = '{}v {}s'.format(vectorSize, steps_done)
+
+                        captioned_image = caption_image_overlay(image, title, footer_left, footer_mid, footer_right)
+                        captioned_image = insert_image_data_embed(captioned_image, data)
+
+                        captioned_image.save(last_saved_image_chunks, "PNG", pnginfo=info)
+                        embedding_yet_to_be_embedded = False
+
+                    last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False)
+                    last_saved_image += f", prompt: {preview_text}"
+
+                shared.state.job_no = embedding.step
+
+                shared.state.textinfo = f"""
 <p>
-Loss: {losses.mean():.7f}<br/>
+Loss: {loss_step:.7f}<br/>
 Step: {embedding.step}<br/>
-Last prompt: {html.escape(entries[0].cond_text)}<br/>
+Last prompt: {html.escape(batch.cond_text[0])}<br/>
 Last saved embedding: {html.escape(last_saved_file)}<br/>
 Last saved image: {html.escape(last_saved_image)}<br/>
 </p>
 """
-
-    filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt')
-    save_embedding(embedding, checkpoint, embedding_name, filename, remove_cached_checksum=True)
-    shared.sd_model.first_stage_model.to(devices.device)
+        filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt')
+        save_embedding(embedding, checkpoint, embedding_name, filename, remove_cached_checksum=True)
+    except Exception:
+        print(traceback.format_exc(), file=sys.stderr)
+        pass
+    finally:
+        pbar.leave = False
+        pbar.close()
+        shared.sd_model.first_stage_model.to(devices.device)
 
     return embedding, filename
 
diff --git a/modules/ui.py b/modules/ui.py
index a5953fce..9d2a1cbf 100644
--- a/modules/ui.py
+++ b/modules/ui.py
@@ -1262,7 +1262,7 @@ def create_ui(wrap_gradio_gpu_call):
                         with gr.Column():
                             with gr.Row():
                                 interrupt_preprocessing = gr.Button("Interrupt")
-                                run_preprocess = gr.Button(value="Preprocess", variant='primary')
+                            run_preprocess = gr.Button(value="Preprocess", variant='primary')
 
                     process_split.change(
                         fn=lambda show: gr_show(show),
@@ -1289,6 +1289,7 @@ def create_ui(wrap_gradio_gpu_call):
                         hypernetwork_learn_rate = gr.Textbox(label='Hypernetwork Learning rate', placeholder="Hypernetwork Learning rate", value="0.00001")
 
                     batch_size = gr.Number(label='Batch size', value=1, precision=0)
+                    gradient_step = gr.Number(label='Gradient accumulation steps', value=1, precision=0)
                     dataset_directory = gr.Textbox(label='Dataset directory', placeholder="Path to directory with input images")
                     log_directory = gr.Textbox(label='Log directory', placeholder="Path to directory where to write outputs", value="textual_inversion")
                     template_file = gr.Textbox(label='Prompt template file', value=os.path.join(script_path, "textual_inversion_templates", "style_filewords.txt"))
@@ -1299,6 +1300,11 @@ def create_ui(wrap_gradio_gpu_call):
                     save_embedding_every = gr.Number(label='Save a copy of embedding to log directory every N steps, 0 to disable', value=500, precision=0)
                     save_image_with_stored_embedding = gr.Checkbox(label='Save images with embedding in PNG chunks', value=True)
                     preview_from_txt2img = gr.Checkbox(label='Read parameters (prompt, etc...) from txt2img tab when making previews', value=False)
+                    with gr.Row():
+                        shuffle_tags = gr.Checkbox(label="Shuffle tags by ',' when creating prompts.", value=False)
+                        tag_drop_out = gr.Slider(minimum=0, maximum=1, step=0.1, label="Drop out tags when creating prompts.", value=0)
+                    with gr.Row():
+                        latent_sampling_method = gr.Radio(label='Choose latent sampling method', value="once", choices=['once', 'deterministic', 'random'])
 
                     with gr.Row():
                         interrupt_training = gr.Button(value="Interrupt")
@@ -1387,11 +1393,15 @@ def create_ui(wrap_gradio_gpu_call):
                 train_embedding_name,
                 embedding_learn_rate,
                 batch_size,
+                gradient_step,
                 dataset_directory,
                 log_directory,
                 training_width,
                 training_height,
                 steps,
+                shuffle_tags,
+                tag_drop_out,
+                latent_sampling_method,
                 create_image_every,
                 save_embedding_every,
                 template_file,
@@ -1412,11 +1422,15 @@ def create_ui(wrap_gradio_gpu_call):
                 train_hypernetwork_name,
                 hypernetwork_learn_rate,
                 batch_size,
+                gradient_step,
                 dataset_directory,
                 log_directory,
                 training_width,
                 training_height,
                 steps,
+                shuffle_tags,
+                tag_drop_out,
+                latent_sampling_method,
                 create_image_every,
                 save_embedding_every,
                 template_file,
-- 
cgit v1.2.3


From 5b57f61ba47f8b11d19a5b46e7fb5a52458abae5 Mon Sep 17 00:00:00 2001
From: flamelaw <flamelaw.com3d2@gmail.com>
Date: Mon, 21 Nov 2022 10:15:46 +0900
Subject: fix pin_memory with different latent sampling method

---
 modules/hypernetworks/hypernetwork.py          |  5 ++++-
 modules/textual_inversion/dataset.py           | 23 +++++++++++++++++++----
 modules/textual_inversion/textual_inversion.py |  7 +------
 3 files changed, 24 insertions(+), 11 deletions(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py
index 3d3301b0..0128419b 100644
--- a/modules/hypernetworks/hypernetwork.py
+++ b/modules/hypernetworks/hypernetwork.py
@@ -416,7 +416,10 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step,
     pin_memory = shared.opts.pin_memory
 
     ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=hypernetwork_name, model=shared.sd_model, cond_model=shared.sd_model.cond_stage_model, device=devices.device, template_file=template_file, include_cond=True, batch_size=batch_size, gradient_step=gradient_step, shuffle_tags=shuffle_tags, tag_drop_out=tag_drop_out, latent_sampling_method=latent_sampling_method)
-    dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, batch_size=ds.batch_size, pin_memory=pin_memory)
+    
+    latent_sampling_method = ds.latent_sampling_method
+
+    dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, latent_sampling_method=latent_sampling_method, batch_size=ds.batch_size, pin_memory=pin_memory)
 
     if unload:
         shared.sd_model.cond_stage_model.to(devices.cpu)
diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py
index 110c0e09..f470324a 100644
--- a/modules/textual_inversion/dataset.py
+++ b/modules/textual_inversion/dataset.py
@@ -138,9 +138,12 @@ class PersonalizedBase(Dataset):
         return entry
 
 class PersonalizedDataLoader(DataLoader):
-    def __init__(self, *args, **kwargs):
-        super(PersonalizedDataLoader, self).__init__(shuffle=True, drop_last=True, *args, **kwargs)
-        self.collate_fn = collate_wrapper
+    def __init__(self, dataset, latent_sampling_method="once", batch_size=1, pin_memory=False):
+        super(PersonalizedDataLoader, self).__init__(dataset, shuffle=True, drop_last=True, batch_size=batch_size, pin_memory=pin_memory)
+        if latent_sampling_method == "random":
+            self.collate_fn = collate_wrapper_random
+        else:
+            self.collate_fn = collate_wrapper
         
 
 class BatchLoader:
@@ -148,10 +151,22 @@ class BatchLoader:
         self.cond_text = [entry.cond_text for entry in data]
         self.cond = [entry.cond for entry in data]
         self.latent_sample = torch.stack([entry.latent_sample for entry in data]).squeeze(1)
+        #self.emb_index = [entry.emb_index for entry in data]
+        #print(self.latent_sample.device)
 
     def pin_memory(self):
         self.latent_sample = self.latent_sample.pin_memory()
         return self
 
 def collate_wrapper(batch):
-    return BatchLoader(batch)
\ No newline at end of file
+    return BatchLoader(batch)
+
+class BatchLoaderRandom(BatchLoader):
+    def __init__(self, data):
+        super().__init__(data)
+
+    def pin_memory(self):
+        return self
+
+def collate_wrapper_random(batch):
+    return BatchLoaderRandom(batch)
\ No newline at end of file
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index 1d5e3a32..3036e48a 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -277,7 +277,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
 
     latent_sampling_method = ds.latent_sampling_method
 
-    dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, batch_size=ds.batch_size, pin_memory=False)
+    dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, latent_sampling_method=latent_sampling_method, batch_size=ds.batch_size, pin_memory=pin_memory)
 
     if unload:
         shared.sd_model.first_stage_model.to(devices.cpu)
@@ -333,11 +333,6 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
                 # go back until we reach gradient accumulation steps
                 if (j + 1) % gradient_step != 0:
                     continue
-                #print(f"grad:{embedding.vec.grad.detach().cpu().abs().mean().item():.7f}")
-                #scaler.unscale_(optimizer)
-                #print(f"grad:{embedding.vec.grad.detach().cpu().abs().mean().item():.7f}")
-                #torch.nn.utils.clip_grad_norm_(embedding.vec, max_norm=1.0)
-                #print(f"grad:{embedding.vec.grad.detach().cpu().abs().mean().item():.7f}")
                 scaler.step(optimizer)
                 scaler.update()
                 embedding.step += 1
-- 
cgit v1.2.3


From 89d8ecff09b426ddc89eb5b432825f8f4c218051 Mon Sep 17 00:00:00 2001
From: flamelaw <flamelaw.com3d2@gmail.com>
Date: Wed, 23 Nov 2022 02:49:01 +0900
Subject: small fixes

---
 modules/hypernetworks/hypernetwork.py          | 6 +++---
 modules/textual_inversion/textual_inversion.py | 2 +-
 2 files changed, 4 insertions(+), 4 deletions(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py
index 0128419b..4541af18 100644
--- a/modules/hypernetworks/hypernetwork.py
+++ b/modules/hypernetworks/hypernetwork.py
@@ -435,8 +435,8 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step,
         optimizer_name = hypernetwork.optimizer_name
     else:
         print(f"Optimizer type {hypernetwork.optimizer_name} is not defined!")
-    optimizer = torch.optim.AdamW(params=weights, lr=scheduler.learn_rate)
-    optimizer_name = 'AdamW'
+        optimizer = torch.optim.AdamW(params=weights, lr=scheduler.learn_rate)
+        optimizer_name = 'AdamW'
 
     if hypernetwork.optimizer_state_dict:  # This line must be changed if Optimizer type can be different from saved optimizer.
         try:
@@ -582,7 +582,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step,
                 shared.state.textinfo = f"""
 <p>
 Loss: {loss_step:.7f}<br/>
-Step: {hypernetwork.step}<br/>
+Step: {steps_done}<br/>
 Last prompt: {html.escape(batch.cond_text[0])}<br/>
 Last saved hypernetwork: {html.escape(last_saved_file)}<br/>
 Last saved image: {html.escape(last_saved_image)}<br/>
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index 3036e48a..fee08e33 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -436,7 +436,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
                 shared.state.textinfo = f"""
 <p>
 Loss: {loss_step:.7f}<br/>
-Step: {embedding.step}<br/>
+Step: {steps_done}<br/>
 Last prompt: {html.escape(batch.cond_text[0])}<br/>
 Last saved embedding: {html.escape(last_saved_file)}<br/>
 Last saved image: {html.escape(last_saved_image)}<br/>
-- 
cgit v1.2.3


From ce6911158b5b2f9cf79b405a1f368f875492044d Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Sat, 26 Nov 2022 16:10:46 +0300
Subject: Add support Stable Diffusion 2.0

---
 README.md                                      |  21 +-
 launch.py                                      |  12 +-
 modules/paths.py                               |   2 +-
 modules/sd_hijack.py                           | 297 +++---------------------
 modules/sd_hijack_clip.py                      | 301 +++++++++++++++++++++++++
 modules/sd_hijack_inpainting.py                |  20 +-
 modules/sd_hijack_open_clip.py                 |  37 +++
 modules/sd_samplers.py                         |  14 +-
 modules/shared.py                              |  34 ++-
 modules/textual_inversion/textual_inversion.py |   7 +-
 modules/ui.py                                  |  13 +-
 requirements.txt                               |   1 +
 requirements_versions.txt                      |   1 +
 v1-inference.yaml                              |  70 ++++++
 webui.py                                       |   5 +-
 15 files changed, 504 insertions(+), 331 deletions(-)
 create mode 100644 modules/sd_hijack_clip.py
 create mode 100644 modules/sd_hijack_open_clip.py
 create mode 100644 v1-inference.yaml

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/README.md b/README.md
index 5f5ab3aa..8a4ffade 100644
--- a/README.md
+++ b/README.md
@@ -84,26 +84,7 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web
 - API
 - Support for dedicated [inpainting model](https://github.com/runwayml/stable-diffusion#inpainting-with-stable-diffusion) by RunwayML. 
 - via extension: [Aesthetic Gradients](https://github.com/AUTOMATIC1111/stable-diffusion-webui-aesthetic-gradients), a way to generate images with a specific aesthetic by using clip images embds (implementation of [https://github.com/vicgalle/stable-diffusion-aesthetic-gradients](https://github.com/vicgalle/stable-diffusion-aesthetic-gradients))
-
-## Where are Aesthetic Gradients?!?!
-Aesthetic Gradients are now an extension. You can install it using git:
-
-```commandline
-git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui-aesthetic-gradients extensions/aesthetic-gradients
-```
-
-After running this command, make sure that you have `aesthetic-gradients` dir in webui's `extensions` directory and restart
-the UI. The interface for Aesthetic Gradients should appear exactly the same as it was.
-
-## Where is History/Image browser?!?!
-Image browser is now an extension. You can install it using git:
-
-```commandline
-git clone https://github.com/yfszzx/stable-diffusion-webui-images-browser extensions/images-browser
-```
-
-After running this command, make sure that you have `images-browser` dir in webui's `extensions` directory and restart
-the UI. The interface for Image browser should appear exactly the same as it was.
+- [Stable Diffusion 2.0](https://github.com/Stability-AI/stablediffusion) support - see [wiki](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#stable-diffusion-20) for instructions
 
 ## Installation and Running
 Make sure the required [dependencies](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Dependencies) are met and follow the instructions available for both [NVidia](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs) (recommended) and [AMD](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs) GPUs.
diff --git a/launch.py b/launch.py
index d2f1055c..b1626cb5 100644
--- a/launch.py
+++ b/launch.py
@@ -134,18 +134,19 @@ def prepare_enviroment():
 
     gfpgan_package = os.environ.get('GFPGAN_PACKAGE', "git+https://github.com/TencentARC/GFPGAN.git@8d2447a2d918f8eba5a4a01463fd48e45126a379")
     clip_package = os.environ.get('CLIP_PACKAGE', "git+https://github.com/openai/CLIP.git@d50d76daa670286dd6cacf3bcd80b5e4823fc8e1")
+    openclip_package = os.environ.get('OPENCLIP_PACKAGE', "git+https://github.com/mlfoundations/open_clip.git@bb6e834e9c70d9c27d0dc3ecedeebeaeb1ffad6b")
 
     xformers_windows_package = os.environ.get('XFORMERS_WINDOWS_PACKAGE', 'https://github.com/C43H66N12O12S2/stable-diffusion-webui/releases/download/f/xformers-0.0.14.dev0-cp310-cp310-win_amd64.whl')
 
-    stable_diffusion_repo = os.environ.get('STABLE_DIFFUSION_REPO', "https://github.com/CompVis/stable-diffusion.git")
+    stable_diffusion_repo = os.environ.get('STABLE_DIFFUSION_REPO', "https://github.com/Stability-AI/stablediffusion.git")
     taming_transformers_repo = os.environ.get('TAMING_TRANSFORMERS_REPO', "https://github.com/CompVis/taming-transformers.git")
     k_diffusion_repo = os.environ.get('K_DIFFUSION_REPO', 'https://github.com/crowsonkb/k-diffusion.git')
     codeformer_repo = os.environ.get('CODEFORMER_REPO', 'https://github.com/sczhou/CodeFormer.git')
     blip_repo = os.environ.get('BLIP_REPO', 'https://github.com/salesforce/BLIP.git')
 
-    stable_diffusion_commit_hash = os.environ.get('STABLE_DIFFUSION_COMMIT_HASH', "69ae4b35e0a0f6ee1af8bb9a5d0016ccb27e36dc")
+    stable_diffusion_commit_hash = os.environ.get('STABLE_DIFFUSION_COMMIT_HASH', "47b6b607fdd31875c9279cd2f4f16b92e4ea958e")
     taming_transformers_commit_hash = os.environ.get('TAMING_TRANSFORMERS_COMMIT_HASH', "24268930bf1dce879235a7fddd0b2355b84d7ea6")
-    k_diffusion_commit_hash = os.environ.get('K_DIFFUSION_COMMIT_HASH', "60e5042ca0da89c14d1dd59d73883280f8fce991")
+    k_diffusion_commit_hash = os.environ.get('K_DIFFUSION_COMMIT_HASH', "5b3af030dd83e0297272d861c19477735d0317ec")
     codeformer_commit_hash = os.environ.get('CODEFORMER_COMMIT_HASH', "c5b4593074ba6214284d6acd5f1719b6c5d739af")
     blip_commit_hash = os.environ.get('BLIP_COMMIT_HASH', "48211a1594f1321b00f14c9f7a5b4813144b2fb9")
 
@@ -179,6 +180,9 @@ def prepare_enviroment():
     if not is_installed("clip"):
         run_pip(f"install {clip_package}", "clip")
 
+    if not is_installed("open_clip"):
+        run_pip(f"install {openclip_package}", "open_clip")
+
     if (not is_installed("xformers") or reinstall_xformers) and xformers:
         if platform.system() == "Windows":
             if platform.python_version().startswith("3.10"):
@@ -196,7 +200,7 @@ def prepare_enviroment():
 
     os.makedirs(dir_repos, exist_ok=True)
 
-    git_clone(stable_diffusion_repo, repo_dir('stable-diffusion'), "Stable Diffusion", stable_diffusion_commit_hash)
+    git_clone(stable_diffusion_repo, repo_dir('stable-diffusion-stability-ai'), "Stable Diffusion", stable_diffusion_commit_hash)
     git_clone(taming_transformers_repo, repo_dir('taming-transformers'), "Taming Transformers", taming_transformers_commit_hash)
     git_clone(k_diffusion_repo, repo_dir('k-diffusion'), "K-diffusion", k_diffusion_commit_hash)
     git_clone(codeformer_repo, repo_dir('CodeFormer'), "CodeFormer", codeformer_commit_hash)
diff --git a/modules/paths.py b/modules/paths.py
index 1e7a2fbc..4dd03a35 100644
--- a/modules/paths.py
+++ b/modules/paths.py
@@ -9,7 +9,7 @@ sys.path.insert(0, script_path)
 
 # search for directory of stable diffusion in following places
 sd_path = None
-possible_sd_paths = [os.path.join(script_path, 'repositories/stable-diffusion'), '.', os.path.dirname(script_path)]
+possible_sd_paths = [os.path.join(script_path, 'repositories/stable-diffusion-stability-ai'), '.', os.path.dirname(script_path)]
 for possible_sd_path in possible_sd_paths:
     if os.path.exists(os.path.join(possible_sd_path, 'ldm/models/diffusion/ddpm.py')):
         sd_path = os.path.abspath(possible_sd_path)
diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index eaedac13..d5243fd3 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -9,18 +9,29 @@ from torch.nn.functional import silu
 
 import modules.textual_inversion.textual_inversion
 from modules import prompt_parser, devices, sd_hijack_optimizations, shared
-from modules.shared import opts, device, cmd_opts
+from modules.shared import cmd_opts
+from modules import sd_hijack_clip, sd_hijack_open_clip
+
 from modules.sd_hijack_optimizations import invokeAI_mps_available
 
 import ldm.modules.attention
 import ldm.modules.diffusionmodules.model
 import ldm.models.diffusion.ddim
 import ldm.models.diffusion.plms
+import ldm.modules.encoders.modules
 
 attention_CrossAttention_forward = ldm.modules.attention.CrossAttention.forward
 diffusionmodules_model_nonlinearity = ldm.modules.diffusionmodules.model.nonlinearity
 diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.AttnBlock.forward
 
+# new memory efficient cross attention blocks do not support hypernets and we already
+# have memory efficient cross attention anyway, so this disables SD2.0's memory efficient cross attention
+ldm.modules.attention.MemoryEfficientCrossAttention = ldm.modules.attention.CrossAttention
+ldm.modules.attention.BasicTransformerBlock.ATTENTION_MODES["softmax-xformers"] = ldm.modules.attention.CrossAttention
+
+# silence new console spam from SD2
+ldm.modules.attention.print = lambda *args: None
+ldm.modules.diffusionmodules.model.print = lambda *args: None
 
 def apply_optimizations():
     undo_optimizations()
@@ -49,16 +60,11 @@ def apply_optimizations():
 
 
 def undo_optimizations():
-    from modules.hypernetworks import hypernetwork
-
-    ldm.modules.attention.CrossAttention.forward = hypernetwork.attention_CrossAttention_forward
+    ldm.modules.attention.CrossAttention.forward = attention_CrossAttention_forward  # this stops hypernets from working
     ldm.modules.diffusionmodules.model.nonlinearity = diffusionmodules_model_nonlinearity
     ldm.modules.diffusionmodules.model.AttnBlock.forward = diffusionmodules_model_AttnBlock_forward
 
 
-def get_target_prompt_token_count(token_count):
-    return math.ceil(max(token_count, 1) / 75) * 75
-
 
 class StableDiffusionModelHijack:
     fixes = None
@@ -70,10 +76,13 @@ class StableDiffusionModelHijack:
     embedding_db = modules.textual_inversion.textual_inversion.EmbeddingDatabase(cmd_opts.embeddings_dir)
 
     def hijack(self, m):
-        model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
-
-        model_embeddings.token_embedding = EmbeddingsWithFixes(model_embeddings.token_embedding, self)
-        m.cond_stage_model = FrozenCLIPEmbedderWithCustomWords(m.cond_stage_model, self)
+        if type(m.cond_stage_model) == ldm.modules.encoders.modules.FrozenCLIPEmbedder:
+            model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
+            model_embeddings.token_embedding = EmbeddingsWithFixes(model_embeddings.token_embedding, self)
+            m.cond_stage_model = sd_hijack_clip.FrozenCLIPEmbedderWithCustomWords(m.cond_stage_model, self)
+        elif type(m.cond_stage_model) == ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder:
+            m.cond_stage_model.model.token_embedding = EmbeddingsWithFixes(m.cond_stage_model.model.token_embedding, self)
+            m.cond_stage_model = sd_hijack_open_clip.FrozenOpenCLIPEmbedderWithCustomWords(m.cond_stage_model, self)
 
         self.clip = m.cond_stage_model
 
@@ -89,12 +98,15 @@ class StableDiffusionModelHijack:
         self.layers = flatten(m)
 
     def undo_hijack(self, m):
-        if type(m.cond_stage_model) == FrozenCLIPEmbedderWithCustomWords:
+        if type(m.cond_stage_model) == sd_hijack_clip.FrozenCLIPEmbedderWithCustomWords:
             m.cond_stage_model = m.cond_stage_model.wrapped
 
-        model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
-        if type(model_embeddings.token_embedding) == EmbeddingsWithFixes:
-            model_embeddings.token_embedding = model_embeddings.token_embedding.wrapped
+            model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
+            if type(model_embeddings.token_embedding) == EmbeddingsWithFixes:
+                model_embeddings.token_embedding = model_embeddings.token_embedding.wrapped
+        elif type(m.cond_stage_model) == sd_hijack_open_clip.FrozenOpenCLIPEmbedderWithCustomWords:
+            m.cond_stage_model.wrapped.model.token_embedding = m.cond_stage_model.wrapped.model.token_embedding.wrapped
+            m.cond_stage_model = m.cond_stage_model.wrapped
 
         self.apply_circular(False)
         self.layers = None
@@ -114,261 +126,8 @@ class StableDiffusionModelHijack:
 
     def tokenize(self, text):
         _, remade_batch_tokens, _, _, _, token_count = self.clip.process_text([text])
-        return remade_batch_tokens[0], token_count, get_target_prompt_token_count(token_count)
-
-
-class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
-    def __init__(self, wrapped, hijack):
-        super().__init__()
-        self.wrapped = wrapped
-        self.hijack: StableDiffusionModelHijack = hijack
-        self.tokenizer = wrapped.tokenizer
-        self.token_mults = {}
-
-        self.comma_token = [v for k, v in self.tokenizer.get_vocab().items() if k == ',</w>'][0]
-
-        tokens_with_parens = [(k, v) for k, v in self.tokenizer.get_vocab().items() if '(' in k or ')' in k or '[' in k or ']' in k]
-        for text, ident in tokens_with_parens:
-            mult = 1.0
-            for c in text:
-                if c == '[':
-                    mult /= 1.1
-                if c == ']':
-                    mult *= 1.1
-                if c == '(':
-                    mult *= 1.1
-                if c == ')':
-                    mult /= 1.1
-
-            if mult != 1.0:
-                self.token_mults[ident] = mult
-
-    def tokenize_line(self, line, used_custom_terms, hijack_comments):
-        id_end = self.wrapped.tokenizer.eos_token_id
-
-        if opts.enable_emphasis:
-            parsed = prompt_parser.parse_prompt_attention(line)
-        else:
-            parsed = [[line, 1.0]]
-
-        tokenized = self.wrapped.tokenizer([text for text, _ in parsed], truncation=False, add_special_tokens=False)["input_ids"]
-
-        fixes = []
-        remade_tokens = []
-        multipliers = []
-        last_comma = -1
-
-        for tokens, (text, weight) in zip(tokenized, parsed):
-            i = 0
-            while i < len(tokens):
-                token = tokens[i]
-
-                embedding, embedding_length_in_tokens = self.hijack.embedding_db.find_embedding_at_position(tokens, i)
-
-                if token == self.comma_token:
-                    last_comma = len(remade_tokens)
-                elif opts.comma_padding_backtrack != 0 and max(len(remade_tokens), 1) % 75 == 0 and last_comma != -1 and len(remade_tokens) - last_comma <= opts.comma_padding_backtrack:
-                    last_comma += 1
-                    reloc_tokens = remade_tokens[last_comma:]
-                    reloc_mults = multipliers[last_comma:]
-
-                    remade_tokens = remade_tokens[:last_comma]
-                    length = len(remade_tokens)
-
-                    rem = int(math.ceil(length / 75)) * 75 - length
-                    remade_tokens += [id_end] * rem + reloc_tokens
-                    multipliers = multipliers[:last_comma] + [1.0] * rem + reloc_mults
-
-                if embedding is None:
-                    remade_tokens.append(token)
-                    multipliers.append(weight)
-                    i += 1
-                else:
-                    emb_len = int(embedding.vec.shape[0])
-                    iteration = len(remade_tokens) // 75
-                    if (len(remade_tokens) + emb_len) // 75 != iteration:
-                        rem = (75 * (iteration + 1) - len(remade_tokens))
-                        remade_tokens += [id_end] * rem
-                        multipliers += [1.0] * rem
-                        iteration += 1
-                    fixes.append((iteration, (len(remade_tokens) % 75, embedding)))
-                    remade_tokens += [0] * emb_len
-                    multipliers += [weight] * emb_len
-                    used_custom_terms.append((embedding.name, embedding.checksum()))
-                    i += embedding_length_in_tokens
-
-        token_count = len(remade_tokens)
-        prompt_target_length = get_target_prompt_token_count(token_count)
-        tokens_to_add = prompt_target_length - len(remade_tokens)
-
-        remade_tokens = remade_tokens + [id_end] * tokens_to_add
-        multipliers = multipliers + [1.0] * tokens_to_add
-
-        return remade_tokens, fixes, multipliers, token_count
-
-    def process_text(self, texts):
-        used_custom_terms = []
-        remade_batch_tokens = []
-        hijack_comments = []
-        hijack_fixes = []
-        token_count = 0
-
-        cache = {}
-        batch_multipliers = []
-        for line in texts:
-            if line in cache:
-                remade_tokens, fixes, multipliers = cache[line]
-            else:
-                remade_tokens, fixes, multipliers, current_token_count = self.tokenize_line(line, used_custom_terms, hijack_comments)
-                token_count = max(current_token_count, token_count)
-
-                cache[line] = (remade_tokens, fixes, multipliers)
-
-            remade_batch_tokens.append(remade_tokens)
-            hijack_fixes.append(fixes)
-            batch_multipliers.append(multipliers)
-
-        return batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count
-
-    def process_text_old(self, text):
-        id_start = self.wrapped.tokenizer.bos_token_id
-        id_end = self.wrapped.tokenizer.eos_token_id
-        maxlen = self.wrapped.max_length  # you get to stay at 77
-        used_custom_terms = []
-        remade_batch_tokens = []
-        overflowing_words = []
-        hijack_comments = []
-        hijack_fixes = []
-        token_count = 0
-
-        cache = {}
-        batch_tokens = self.wrapped.tokenizer(text, truncation=False, add_special_tokens=False)["input_ids"]
-        batch_multipliers = []
-        for tokens in batch_tokens:
-            tuple_tokens = tuple(tokens)
-
-            if tuple_tokens in cache:
-                remade_tokens, fixes, multipliers = cache[tuple_tokens]
-            else:
-                fixes = []
-                remade_tokens = []
-                multipliers = []
-                mult = 1.0
-
-                i = 0
-                while i < len(tokens):
-                    token = tokens[i]
-
-                    embedding, embedding_length_in_tokens = self.hijack.embedding_db.find_embedding_at_position(tokens, i)
-
-                    mult_change = self.token_mults.get(token) if opts.enable_emphasis else None
-                    if mult_change is not None:
-                        mult *= mult_change
-                        i += 1
-                    elif embedding is None:
-                        remade_tokens.append(token)
-                        multipliers.append(mult)
-                        i += 1
-                    else:
-                        emb_len = int(embedding.vec.shape[0])
-                        fixes.append((len(remade_tokens), embedding))
-                        remade_tokens += [0] * emb_len
-                        multipliers += [mult] * emb_len
-                        used_custom_terms.append((embedding.name, embedding.checksum()))
-                        i += embedding_length_in_tokens
-
-                if len(remade_tokens) > maxlen - 2:
-                    vocab = {v: k for k, v in self.wrapped.tokenizer.get_vocab().items()}
-                    ovf = remade_tokens[maxlen - 2:]
-                    overflowing_words = [vocab.get(int(x), "") for x in ovf]
-                    overflowing_text = self.wrapped.tokenizer.convert_tokens_to_string(''.join(overflowing_words))
-                    hijack_comments.append(f"Warning: too many input tokens; some ({len(overflowing_words)}) have been truncated:\n{overflowing_text}\n")
-
-                token_count = len(remade_tokens)
-                remade_tokens = remade_tokens + [id_end] * (maxlen - 2 - len(remade_tokens))
-                remade_tokens = [id_start] + remade_tokens[0:maxlen - 2] + [id_end]
-                cache[tuple_tokens] = (remade_tokens, fixes, multipliers)
-
-            multipliers = multipliers + [1.0] * (maxlen - 2 - len(multipliers))
-            multipliers = [1.0] + multipliers[0:maxlen - 2] + [1.0]
-
-            remade_batch_tokens.append(remade_tokens)
-            hijack_fixes.append(fixes)
-            batch_multipliers.append(multipliers)
-        return batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count
-
-    def forward(self, text):
-        use_old = opts.use_old_emphasis_implementation
-        if use_old:
-            batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text_old(text)
-        else:
-            batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text(text)
-
-        self.hijack.comments += hijack_comments
-
-        if len(used_custom_terms) > 0:
-            self.hijack.comments.append("Used embeddings: " + ", ".join([f'{word} [{checksum}]' for word, checksum in used_custom_terms]))
-
-        if use_old:
-            self.hijack.fixes = hijack_fixes
-            return self.process_tokens(remade_batch_tokens, batch_multipliers)
-
-        z = None
-        i = 0
-        while max(map(len, remade_batch_tokens)) != 0:
-            rem_tokens = [x[75:] for x in remade_batch_tokens]
-            rem_multipliers = [x[75:] for x in batch_multipliers]
-
-            self.hijack.fixes = []
-            for unfiltered in hijack_fixes:
-                fixes = []
-                for fix in unfiltered:
-                    if fix[0] == i:
-                        fixes.append(fix[1])
-                self.hijack.fixes.append(fixes)
-
-            tokens = []
-            multipliers = []
-            for j in range(len(remade_batch_tokens)):
-                if len(remade_batch_tokens[j]) > 0:
-                    tokens.append(remade_batch_tokens[j][:75])
-                    multipliers.append(batch_multipliers[j][:75])
-                else:
-                    tokens.append([self.wrapped.tokenizer.eos_token_id] * 75)
-                    multipliers.append([1.0] * 75)
-
-            z1 = self.process_tokens(tokens, multipliers)
-            z = z1 if z is None else torch.cat((z, z1), axis=-2)
-
-            remade_batch_tokens = rem_tokens
-            batch_multipliers = rem_multipliers
-            i += 1
-
-        return z
-
-    def process_tokens(self, remade_batch_tokens, batch_multipliers):
-        if not opts.use_old_emphasis_implementation:
-            remade_batch_tokens = [[self.wrapped.tokenizer.bos_token_id] + x[:75] + [self.wrapped.tokenizer.eos_token_id] for x in remade_batch_tokens]
-            batch_multipliers = [[1.0] + x[:75] + [1.0] for x in batch_multipliers]
-
-        tokens = torch.asarray(remade_batch_tokens).to(device)
-        outputs = self.wrapped.transformer(input_ids=tokens, output_hidden_states=-opts.CLIP_stop_at_last_layers)
-
-        if opts.CLIP_stop_at_last_layers > 1:
-            z = outputs.hidden_states[-opts.CLIP_stop_at_last_layers]
-            z = self.wrapped.transformer.text_model.final_layer_norm(z)
-        else:
-            z = outputs.last_hidden_state
-
-        # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
-        batch_multipliers_of_same_length = [x + [1.0] * (75 - len(x)) for x in batch_multipliers]
-        batch_multipliers = torch.asarray(batch_multipliers_of_same_length).to(device)
-        original_mean = z.mean()
-        z *= batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
-        new_mean = z.mean()
-        z *= original_mean / new_mean
+        return remade_batch_tokens[0], token_count, sd_hijack_clip.get_target_prompt_token_count(token_count)
 
-        return z
 
 
 class EmbeddingsWithFixes(torch.nn.Module):
diff --git a/modules/sd_hijack_clip.py b/modules/sd_hijack_clip.py
new file mode 100644
index 00000000..b451d1cf
--- /dev/null
+++ b/modules/sd_hijack_clip.py
@@ -0,0 +1,301 @@
+import math
+
+import torch
+
+from modules import prompt_parser, devices
+from modules.shared import opts
+
+
+def get_target_prompt_token_count(token_count):
+    return math.ceil(max(token_count, 1) / 75) * 75
+
+
+class FrozenCLIPEmbedderWithCustomWordsBase(torch.nn.Module):
+    def __init__(self, wrapped, hijack):
+        super().__init__()
+        self.wrapped = wrapped
+        self.hijack = hijack
+
+    def tokenize(self, texts):
+        raise NotImplementedError
+
+    def encode_with_transformers(self, tokens):
+        raise NotImplementedError
+
+    def encode_embedding_init_text(self, init_text, nvpt):
+        raise NotImplementedError
+
+    def tokenize_line(self, line, used_custom_terms, hijack_comments):
+        if opts.enable_emphasis:
+            parsed = prompt_parser.parse_prompt_attention(line)
+        else:
+            parsed = [[line, 1.0]]
+
+        tokenized = self.tokenize([text for text, _ in parsed])
+
+        fixes = []
+        remade_tokens = []
+        multipliers = []
+        last_comma = -1
+
+        for tokens, (text, weight) in zip(tokenized, parsed):
+            i = 0
+            while i < len(tokens):
+                token = tokens[i]
+
+                embedding, embedding_length_in_tokens = self.hijack.embedding_db.find_embedding_at_position(tokens, i)
+
+                if token == self.comma_token:
+                    last_comma = len(remade_tokens)
+                elif opts.comma_padding_backtrack != 0 and max(len(remade_tokens), 1) % 75 == 0 and last_comma != -1 and len(remade_tokens) - last_comma <= opts.comma_padding_backtrack:
+                    last_comma += 1
+                    reloc_tokens = remade_tokens[last_comma:]
+                    reloc_mults = multipliers[last_comma:]
+
+                    remade_tokens = remade_tokens[:last_comma]
+                    length = len(remade_tokens)
+
+                    rem = int(math.ceil(length / 75)) * 75 - length
+                    remade_tokens += [self.id_end] * rem + reloc_tokens
+                    multipliers = multipliers[:last_comma] + [1.0] * rem + reloc_mults
+
+                if embedding is None:
+                    remade_tokens.append(token)
+                    multipliers.append(weight)
+                    i += 1
+                else:
+                    emb_len = int(embedding.vec.shape[0])
+                    iteration = len(remade_tokens) // 75
+                    if (len(remade_tokens) + emb_len) // 75 != iteration:
+                        rem = (75 * (iteration + 1) - len(remade_tokens))
+                        remade_tokens += [self.id_end] * rem
+                        multipliers += [1.0] * rem
+                        iteration += 1
+                    fixes.append((iteration, (len(remade_tokens) % 75, embedding)))
+                    remade_tokens += [0] * emb_len
+                    multipliers += [weight] * emb_len
+                    used_custom_terms.append((embedding.name, embedding.checksum()))
+                    i += embedding_length_in_tokens
+
+        token_count = len(remade_tokens)
+        prompt_target_length = get_target_prompt_token_count(token_count)
+        tokens_to_add = prompt_target_length - len(remade_tokens)
+
+        remade_tokens = remade_tokens + [self.id_end] * tokens_to_add
+        multipliers = multipliers + [1.0] * tokens_to_add
+
+        return remade_tokens, fixes, multipliers, token_count
+
+    def process_text(self, texts):
+        used_custom_terms = []
+        remade_batch_tokens = []
+        hijack_comments = []
+        hijack_fixes = []
+        token_count = 0
+
+        cache = {}
+        batch_multipliers = []
+        for line in texts:
+            if line in cache:
+                remade_tokens, fixes, multipliers = cache[line]
+            else:
+                remade_tokens, fixes, multipliers, current_token_count = self.tokenize_line(line, used_custom_terms, hijack_comments)
+                token_count = max(current_token_count, token_count)
+
+                cache[line] = (remade_tokens, fixes, multipliers)
+
+            remade_batch_tokens.append(remade_tokens)
+            hijack_fixes.append(fixes)
+            batch_multipliers.append(multipliers)
+
+        return batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count
+
+    def process_text_old(self, texts):
+        id_start = self.id_start
+        id_end = self.id_end
+        maxlen = self.wrapped.max_length  # you get to stay at 77
+        used_custom_terms = []
+        remade_batch_tokens = []
+        hijack_comments = []
+        hijack_fixes = []
+        token_count = 0
+
+        cache = {}
+        batch_tokens = self.tokenize(texts)
+        batch_multipliers = []
+        for tokens in batch_tokens:
+            tuple_tokens = tuple(tokens)
+
+            if tuple_tokens in cache:
+                remade_tokens, fixes, multipliers = cache[tuple_tokens]
+            else:
+                fixes = []
+                remade_tokens = []
+                multipliers = []
+                mult = 1.0
+
+                i = 0
+                while i < len(tokens):
+                    token = tokens[i]
+
+                    embedding, embedding_length_in_tokens = self.hijack.embedding_db.find_embedding_at_position(tokens, i)
+
+                    mult_change = self.token_mults.get(token) if opts.enable_emphasis else None
+                    if mult_change is not None:
+                        mult *= mult_change
+                        i += 1
+                    elif embedding is None:
+                        remade_tokens.append(token)
+                        multipliers.append(mult)
+                        i += 1
+                    else:
+                        emb_len = int(embedding.vec.shape[0])
+                        fixes.append((len(remade_tokens), embedding))
+                        remade_tokens += [0] * emb_len
+                        multipliers += [mult] * emb_len
+                        used_custom_terms.append((embedding.name, embedding.checksum()))
+                        i += embedding_length_in_tokens
+
+                if len(remade_tokens) > maxlen - 2:
+                    vocab = {v: k for k, v in self.wrapped.tokenizer.get_vocab().items()}
+                    ovf = remade_tokens[maxlen - 2:]
+                    overflowing_words = [vocab.get(int(x), "") for x in ovf]
+                    overflowing_text = self.wrapped.tokenizer.convert_tokens_to_string(''.join(overflowing_words))
+                    hijack_comments.append(f"Warning: too many input tokens; some ({len(overflowing_words)}) have been truncated:\n{overflowing_text}\n")
+
+                token_count = len(remade_tokens)
+                remade_tokens = remade_tokens + [id_end] * (maxlen - 2 - len(remade_tokens))
+                remade_tokens = [id_start] + remade_tokens[0:maxlen - 2] + [id_end]
+                cache[tuple_tokens] = (remade_tokens, fixes, multipliers)
+
+            multipliers = multipliers + [1.0] * (maxlen - 2 - len(multipliers))
+            multipliers = [1.0] + multipliers[0:maxlen - 2] + [1.0]
+
+            remade_batch_tokens.append(remade_tokens)
+            hijack_fixes.append(fixes)
+            batch_multipliers.append(multipliers)
+        return batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count
+
+    def forward(self, text):
+        use_old = opts.use_old_emphasis_implementation
+        if use_old:
+            batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text_old(text)
+        else:
+            batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text(text)
+
+        self.hijack.comments += hijack_comments
+
+        if len(used_custom_terms) > 0:
+            self.hijack.comments.append("Used embeddings: " + ", ".join([f'{word} [{checksum}]' for word, checksum in used_custom_terms]))
+
+        if use_old:
+            self.hijack.fixes = hijack_fixes
+            return self.process_tokens(remade_batch_tokens, batch_multipliers)
+
+        z = None
+        i = 0
+        while max(map(len, remade_batch_tokens)) != 0:
+            rem_tokens = [x[75:] for x in remade_batch_tokens]
+            rem_multipliers = [x[75:] for x in batch_multipliers]
+
+            self.hijack.fixes = []
+            for unfiltered in hijack_fixes:
+                fixes = []
+                for fix in unfiltered:
+                    if fix[0] == i:
+                        fixes.append(fix[1])
+                self.hijack.fixes.append(fixes)
+
+            tokens = []
+            multipliers = []
+            for j in range(len(remade_batch_tokens)):
+                if len(remade_batch_tokens[j]) > 0:
+                    tokens.append(remade_batch_tokens[j][:75])
+                    multipliers.append(batch_multipliers[j][:75])
+                else:
+                    tokens.append([self.id_end] * 75)
+                    multipliers.append([1.0] * 75)
+
+            z1 = self.process_tokens(tokens, multipliers)
+            z = z1 if z is None else torch.cat((z, z1), axis=-2)
+
+            remade_batch_tokens = rem_tokens
+            batch_multipliers = rem_multipliers
+            i += 1
+
+        return z
+
+    def process_tokens(self, remade_batch_tokens, batch_multipliers):
+        if not opts.use_old_emphasis_implementation:
+            remade_batch_tokens = [[self.id_start] + x[:75] + [self.id_end] for x in remade_batch_tokens]
+            batch_multipliers = [[1.0] + x[:75] + [1.0] for x in batch_multipliers]
+
+        tokens = torch.asarray(remade_batch_tokens).to(devices.device)
+
+        if self.id_end != self.id_pad:
+            for batch_pos in range(len(remade_batch_tokens)):
+                index = remade_batch_tokens[batch_pos].index(self.id_end)
+                tokens[batch_pos, index+1:tokens.shape[1]] = self.id_pad
+
+        z = self.encode_with_transformers(tokens)
+
+        # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
+        batch_multipliers_of_same_length = [x + [1.0] * (75 - len(x)) for x in batch_multipliers]
+        batch_multipliers = torch.asarray(batch_multipliers_of_same_length).to(devices.device)
+        original_mean = z.mean()
+        z *= batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
+        new_mean = z.mean()
+        z *= original_mean / new_mean
+
+        return z
+
+
+class FrozenCLIPEmbedderWithCustomWords(FrozenCLIPEmbedderWithCustomWordsBase):
+    def __init__(self, wrapped, hijack):
+        super().__init__(wrapped, hijack)
+        self.tokenizer = wrapped.tokenizer
+        self.comma_token = [v for k, v in self.tokenizer.get_vocab().items() if k == ',</w>'][0]
+
+        self.token_mults = {}
+        tokens_with_parens = [(k, v) for k, v in self.tokenizer.get_vocab().items() if '(' in k or ')' in k or '[' in k or ']' in k]
+        for text, ident in tokens_with_parens:
+            mult = 1.0
+            for c in text:
+                if c == '[':
+                    mult /= 1.1
+                if c == ']':
+                    mult *= 1.1
+                if c == '(':
+                    mult *= 1.1
+                if c == ')':
+                    mult /= 1.1
+
+            if mult != 1.0:
+                self.token_mults[ident] = mult
+
+        self.id_start = self.wrapped.tokenizer.bos_token_id
+        self.id_end = self.wrapped.tokenizer.eos_token_id
+        self.id_pad = self.id_end
+
+    def tokenize(self, texts):
+        tokenized = self.wrapped.tokenizer(texts, truncation=False, add_special_tokens=False)["input_ids"]
+
+        return tokenized
+
+    def encode_with_transformers(self, tokens):
+        outputs = self.wrapped.transformer(input_ids=tokens, output_hidden_states=-opts.CLIP_stop_at_last_layers)
+
+        if opts.CLIP_stop_at_last_layers > 1:
+            z = outputs.hidden_states[-opts.CLIP_stop_at_last_layers]
+            z = self.wrapped.transformer.text_model.final_layer_norm(z)
+        else:
+            z = outputs.last_hidden_state
+
+        return z
+
+    def encode_embedding_init_text(self, init_text, nvpt):
+        embedding_layer = self.wrapped.transformer.text_model.embeddings
+        ids = self.wrapped.tokenizer(init_text, max_length=nvpt, return_tensors="pt", add_special_tokens=False)["input_ids"]
+        embedded = embedding_layer.token_embedding.wrapped(ids.to(devices.device)).squeeze(0)
+
+        return embedded
diff --git a/modules/sd_hijack_inpainting.py b/modules/sd_hijack_inpainting.py
index 46714a4f..938f9a58 100644
--- a/modules/sd_hijack_inpainting.py
+++ b/modules/sd_hijack_inpainting.py
@@ -199,8 +199,8 @@ def sample_plms(self,
 
 @torch.no_grad()
 def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
-                    temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
-                    unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None):
+                  temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                  unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None, dynamic_threshold=None):
     b, *_, device = *x.shape, x.device
 
     def get_model_output(x, t):
@@ -249,6 +249,8 @@ def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=F
         pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
         if quantize_denoised:
             pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+        if dynamic_threshold is not None:
+            pred_x0 = norm_thresholding(pred_x0, dynamic_threshold)
         # direction pointing to x_t
         dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
         noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
@@ -321,12 +323,16 @@ def should_hijack_inpainting(checkpoint_info):
 
 
 def do_inpainting_hijack():
-    ldm.models.diffusion.ddpm.get_unconditional_conditioning = get_unconditional_conditioning
+    # most of this stuff seems to no longer be needed because it is already included into SD2.0
+    # LatentInpaintDiffusion remains because SD2.0's LatentInpaintDiffusion can't be loaded without specifying a checkpoint
+    # p_sample_plms is needed because PLMS can't work with dicts as conditionings
+    # this file should be cleaned up later if weverything tuens out to work fine
+
+    # ldm.models.diffusion.ddpm.get_unconditional_conditioning = get_unconditional_conditioning
     ldm.models.diffusion.ddpm.LatentInpaintDiffusion = LatentInpaintDiffusion
 
-    ldm.models.diffusion.ddim.DDIMSampler.p_sample_ddim = p_sample_ddim
-    ldm.models.diffusion.ddim.DDIMSampler.sample = sample_ddim
+    # ldm.models.diffusion.ddim.DDIMSampler.p_sample_ddim = p_sample_ddim
+    # ldm.models.diffusion.ddim.DDIMSampler.sample = sample_ddim
 
     ldm.models.diffusion.plms.PLMSSampler.p_sample_plms = p_sample_plms
-    ldm.models.diffusion.plms.PLMSSampler.sample = sample_plms
-
+    # ldm.models.diffusion.plms.PLMSSampler.sample = sample_plms
diff --git a/modules/sd_hijack_open_clip.py b/modules/sd_hijack_open_clip.py
new file mode 100644
index 00000000..f733e852
--- /dev/null
+++ b/modules/sd_hijack_open_clip.py
@@ -0,0 +1,37 @@
+import open_clip.tokenizer
+import torch
+
+from modules import sd_hijack_clip, devices
+from modules.shared import opts
+
+tokenizer = open_clip.tokenizer._tokenizer
+
+
+class FrozenOpenCLIPEmbedderWithCustomWords(sd_hijack_clip.FrozenCLIPEmbedderWithCustomWordsBase):
+    def __init__(self, wrapped, hijack):
+        super().__init__(wrapped, hijack)
+
+        self.comma_token = [v for k, v in tokenizer.encoder.items() if k == ',</w>'][0]
+        self.id_start = tokenizer.encoder["<start_of_text>"]
+        self.id_end = tokenizer.encoder["<end_of_text>"]
+        self.id_pad = 0
+
+    def tokenize(self, texts):
+        assert not opts.use_old_emphasis_implementation, 'Old emphasis implementation not supported for Open Clip'
+
+        tokenized = [tokenizer.encode(text) for text in texts]
+
+        return tokenized
+
+    def encode_with_transformers(self, tokens):
+        # set self.wrapped.layer_idx here according to opts.CLIP_stop_at_last_layers
+        z = self.wrapped.encode_with_transformer(tokens)
+
+        return z
+
+    def encode_embedding_init_text(self, init_text, nvpt):
+        ids = tokenizer.encode(init_text)
+        ids = torch.asarray([ids], device=devices.device, dtype=torch.int)
+        embedded = self.wrapped.model.token_embedding.wrapped(ids).squeeze(0)
+
+        return embedded
diff --git a/modules/sd_samplers.py b/modules/sd_samplers.py
index 4fe67854..4edd8c60 100644
--- a/modules/sd_samplers.py
+++ b/modules/sd_samplers.py
@@ -127,7 +127,8 @@ class InterruptedException(BaseException):
 class VanillaStableDiffusionSampler:
     def __init__(self, constructor, sd_model):
         self.sampler = constructor(sd_model)
-        self.orig_p_sample_ddim = self.sampler.p_sample_ddim if hasattr(self.sampler, 'p_sample_ddim') else self.sampler.p_sample_plms
+        self.is_plms = hasattr(self.sampler, 'p_sample_plms')
+        self.orig_p_sample_ddim = self.sampler.p_sample_plms if self.is_plms else self.sampler.p_sample_ddim
         self.mask = None
         self.nmask = None
         self.init_latent = None
@@ -218,7 +219,6 @@ class VanillaStableDiffusionSampler:
         self.mask = p.mask if hasattr(p, 'mask') else None
         self.nmask = p.nmask if hasattr(p, 'nmask') else None
 
-
     def adjust_steps_if_invalid(self, p, num_steps):
         if  (self.config.name == 'DDIM' and p.ddim_discretize == 'uniform') or (self.config.name == 'PLMS'):
             valid_step = 999 / (1000 // num_steps)
@@ -227,7 +227,6 @@ class VanillaStableDiffusionSampler:
         
         return num_steps
 
-
     def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None, image_conditioning=None):
         steps, t_enc = setup_img2img_steps(p, steps)
         steps = self.adjust_steps_if_invalid(p, steps)
@@ -260,9 +259,10 @@ class VanillaStableDiffusionSampler:
         steps = self.adjust_steps_if_invalid(p, steps or p.steps)
 
         # Wrap the conditioning models with additional image conditioning for inpainting model
+        # dummy_for_plms is needed because PLMS code checks the first item in the dict to have the right shape
         if image_conditioning is not None:
-            conditioning = {"c_concat": [image_conditioning], "c_crossattn": [conditioning]}
-            unconditional_conditioning = {"c_concat": [image_conditioning], "c_crossattn": [unconditional_conditioning]}
+            conditioning = {"dummy_for_plms": np.zeros((conditioning.shape[0],)), "c_crossattn": [conditioning], "c_concat": [image_conditioning]}
+            unconditional_conditioning = {"c_crossattn": [unconditional_conditioning], "c_concat": [image_conditioning]}
 
         samples_ddim = self.launch_sampling(steps, lambda: self.sampler.sample(S=steps, conditioning=conditioning, batch_size=int(x.shape[0]), shape=x[0].shape, verbose=False, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning, x_T=x, eta=self.eta)[0])
 
@@ -350,7 +350,9 @@ class TorchHijack:
 
 class KDiffusionSampler:
     def __init__(self, funcname, sd_model):
-        self.model_wrap = k_diffusion.external.CompVisDenoiser(sd_model, quantize=shared.opts.enable_quantization)
+        denoiser = k_diffusion.external.CompVisVDenoiser if sd_model.parameterization == "v" else k_diffusion.external.CompVisDenoiser
+
+        self.model_wrap = denoiser(sd_model, quantize=shared.opts.enable_quantization)
         self.funcname = funcname
         self.func = getattr(k_diffusion.sampling, self.funcname)
         self.extra_params = sampler_extra_params.get(funcname, [])
diff --git a/modules/shared.py b/modules/shared.py
index c93ae2a3..8fb1387a 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -11,17 +11,15 @@ import tqdm
 import modules.artists
 import modules.interrogate
 import modules.memmon
-import modules.sd_models
 import modules.styles
 import modules.devices as devices
-from modules import sd_samplers, sd_models, localization, sd_vae, extensions, script_loading
-from modules.hypernetworks import hypernetwork
+from modules import localization, sd_vae, extensions, script_loading
 from modules.paths import models_path, script_path, sd_path
 
 sd_model_file = os.path.join(script_path, 'model.ckpt')
 default_sd_model_file = sd_model_file
 parser = argparse.ArgumentParser()
-parser.add_argument("--config", type=str, default=os.path.join(sd_path, "configs/stable-diffusion/v1-inference.yaml"), help="path to config which constructs model",)
+parser.add_argument("--config", type=str, default=os.path.join(script_path, "v1-inference.yaml"), help="path to config which constructs model",)
 parser.add_argument("--ckpt", type=str, default=sd_model_file, help="path to checkpoint of stable diffusion model; if specified, this checkpoint will be added to the list of checkpoints and loaded",)
 parser.add_argument("--ckpt-dir", type=str, default=None, help="Path to directory with stable diffusion checkpoints")
 parser.add_argument("--gfpgan-dir", type=str, help="GFPGAN directory", default=('./src/gfpgan' if os.path.exists('./src/gfpgan') else './GFPGAN'))
@@ -121,10 +119,12 @@ xformers_available = False
 config_filename = cmd_opts.ui_settings_file
 
 os.makedirs(cmd_opts.hypernetwork_dir, exist_ok=True)
-hypernetworks = hypernetwork.list_hypernetworks(cmd_opts.hypernetwork_dir)
+hypernetworks = {}
 loaded_hypernetwork = None
 
+
 def reload_hypernetworks():
+    from modules.hypernetworks import hypernetwork
     global hypernetworks
 
     hypernetworks = hypernetwork.list_hypernetworks(cmd_opts.hypernetwork_dir)
@@ -206,10 +206,11 @@ class State:
         if self.current_latent is None:
             return
 
+        import modules.sd_samplers
         if opts.show_progress_grid:
-            self.current_image = sd_samplers.samples_to_image_grid(self.current_latent)
+            self.current_image = modules.sd_samplers.samples_to_image_grid(self.current_latent)
         else:
-            self.current_image = sd_samplers.sample_to_image(self.current_latent)
+            self.current_image = modules.sd_samplers.sample_to_image(self.current_latent)
 
         self.current_image_sampling_step = self.sampling_step
 
@@ -248,6 +249,21 @@ def options_section(section_identifier, options_dict):
     return options_dict
 
 
+def list_checkpoint_tiles():
+    import modules.sd_models
+    return modules.sd_models.checkpoint_tiles()
+
+
+def refresh_checkpoints():
+    import modules.sd_models
+    return modules.sd_models.list_models()
+
+
+def list_samplers():
+    import modules.sd_samplers
+    return modules.sd_samplers.all_samplers
+
+
 hide_dirs = {"visible": not cmd_opts.hide_ui_dir_config}
 
 options_templates = {}
@@ -333,7 +349,7 @@ options_templates.update(options_section(('training', "Training"), {
 }))
 
 options_templates.update(options_section(('sd', "Stable Diffusion"), {
-    "sd_model_checkpoint": OptionInfo(None, "Stable Diffusion checkpoint", gr.Dropdown, lambda: {"choices": modules.sd_models.checkpoint_tiles()}, refresh=sd_models.list_models),
+    "sd_model_checkpoint": OptionInfo(None, "Stable Diffusion checkpoint", gr.Dropdown, lambda: {"choices": list_checkpoint_tiles()}, refresh=refresh_checkpoints),
     "sd_checkpoint_cache": OptionInfo(0, "Checkpoints to cache in RAM", gr.Slider, {"minimum": 0, "maximum": 10, "step": 1}),
     "sd_vae": OptionInfo("auto", "SD VAE", gr.Dropdown, lambda: {"choices": sd_vae.vae_list}, refresh=sd_vae.refresh_vae_list),
     "sd_vae_as_default": OptionInfo(False, "Ignore selected VAE for stable diffusion checkpoints that have their own .vae.pt next to them"),
@@ -385,7 +401,7 @@ options_templates.update(options_section(('ui', "User interface"), {
 }))
 
 options_templates.update(options_section(('sampler-params', "Sampler parameters"), {
-    "hide_samplers": OptionInfo([], "Hide samplers in user interface (requires restart)", gr.CheckboxGroup, lambda: {"choices": [x.name for x in sd_samplers.all_samplers]}),
+    "hide_samplers": OptionInfo([], "Hide samplers in user interface (requires restart)", gr.CheckboxGroup, lambda: {"choices": [x.name for x in list_samplers()]}),
     "eta_ddim": OptionInfo(0.0, "eta (noise multiplier) for DDIM", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}),
     "eta_ancestral": OptionInfo(1.0, "eta (noise multiplier) for ancestral samplers", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}),
     "ddim_discretize": OptionInfo('uniform', "img2img DDIM discretize", gr.Radio, {"choices": ['uniform', 'quad']}),
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index 5e4d8688..a273e663 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -64,7 +64,8 @@ class EmbeddingDatabase:
 
         self.word_embeddings[embedding.name] = embedding
 
-        ids = model.cond_stage_model.tokenizer([embedding.name], add_special_tokens=False)['input_ids'][0]
+        # TODO changing between clip and open clip changes tokenization, which will cause embeddings to stop working
+        ids = model.cond_stage_model.tokenize([embedding.name])[0]
 
         first_id = ids[0]
         if first_id not in self.ids_lookup:
@@ -155,13 +156,11 @@ class EmbeddingDatabase:
 
 def create_embedding(name, num_vectors_per_token, overwrite_old, init_text='*'):
     cond_model = shared.sd_model.cond_stage_model
-    embedding_layer = cond_model.wrapped.transformer.text_model.embeddings
 
     with devices.autocast():
         cond_model([""])  # will send cond model to GPU if lowvram/medvram is active
 
-    ids = cond_model.tokenizer(init_text, max_length=num_vectors_per_token, return_tensors="pt", add_special_tokens=False)["input_ids"]
-    embedded = embedding_layer.token_embedding.wrapped(ids.to(devices.device)).squeeze(0)
+    embedded = cond_model.encode_embedding_init_text(init_text, num_vectors_per_token)
     vec = torch.zeros((num_vectors_per_token, embedded.shape[1]), device=devices.device)
 
     for i in range(num_vectors_per_token):
diff --git a/modules/ui.py b/modules/ui.py
index e6da1b2a..e5cb69d0 100644
--- a/modules/ui.py
+++ b/modules/ui.py
@@ -478,9 +478,7 @@ def create_toprow(is_img2img):
         if is_img2img:
             with gr.Column(scale=1, elem_id="interrogate_col"):
                 button_interrogate = gr.Button('Interrogate\nCLIP', elem_id="interrogate")
-
-                if cmd_opts.deepdanbooru:
-                    button_deepbooru = gr.Button('Interrogate\nDeepBooru', elem_id="deepbooru")
+                button_deepbooru = gr.Button('Interrogate\nDeepBooru', elem_id="deepbooru")
 
         with gr.Column(scale=1):
             with gr.Row():
@@ -1004,11 +1002,10 @@ def create_ui(wrap_gradio_gpu_call):
                 outputs=[img2img_prompt],
             )
 
-            if cmd_opts.deepdanbooru:
-                img2img_deepbooru.click(
-                    fn=interrogate_deepbooru,
-                    inputs=[init_img],
-                    outputs=[img2img_prompt],
+            img2img_deepbooru.click(
+                fn=interrogate_deepbooru,
+                inputs=[init_img],
+                outputs=[img2img_prompt],
             )
 
 
diff --git a/requirements.txt b/requirements.txt
index 762db4f3..e4e5ec64 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -28,3 +28,4 @@ kornia
 lark
 inflection
 GitPython
+torchsde
diff --git a/requirements_versions.txt b/requirements_versions.txt
index 662ca684..8d557fe3 100644
--- a/requirements_versions.txt
+++ b/requirements_versions.txt
@@ -25,3 +25,4 @@ kornia==0.6.7
 lark==1.1.2
 inflection==0.5.1
 GitPython==3.1.27
+torchsde==0.2.5
diff --git a/v1-inference.yaml b/v1-inference.yaml
new file mode 100644
index 00000000..d4effe56
--- /dev/null
+++ b/v1-inference.yaml
@@ -0,0 +1,70 @@
+model:
+  base_learning_rate: 1.0e-04
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: "jpg"
+    cond_stage_key: "txt"
+    image_size: 64
+    channels: 4
+    cond_stage_trainable: false   # Note: different from the one we trained before
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    scale_factor: 0.18215
+    use_ema: False
+
+    scheduler_config: # 10000 warmup steps
+      target: ldm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [ 10000 ]
+        cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
+        f_start: [ 1.e-6 ]
+        f_max: [ 1. ]
+        f_min: [ 1. ]
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 32 # unused
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [ 4, 2, 1 ]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 4, 4 ]
+        num_heads: 8
+        use_spatial_transformer: True
+        transformer_depth: 1
+        context_dim: 768
+        use_checkpoint: True
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
diff --git a/webui.py b/webui.py
index c5e5fe75..23215d1e 100644
--- a/webui.py
+++ b/webui.py
@@ -10,7 +10,7 @@ from fastapi.middleware.gzip import GZipMiddleware
 
 from modules.paths import script_path
 
-from modules import devices, sd_samplers, upscaler, extensions, localization
+from modules import shared, devices, sd_samplers, upscaler, extensions, localization
 import modules.codeformer_model as codeformer
 import modules.extras
 import modules.face_restoration
@@ -23,7 +23,6 @@ import modules.scripts
 import modules.sd_hijack
 import modules.sd_models
 import modules.sd_vae
-import modules.shared as shared
 import modules.txt2img
 import modules.script_callbacks
 
@@ -86,7 +85,7 @@ def initialize():
     shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights()))
     shared.opts.onchange("sd_vae", wrap_queued_call(lambda: modules.sd_vae.reload_vae_weights()), call=False)
     shared.opts.onchange("sd_vae_as_default", wrap_queued_call(lambda: modules.sd_vae.reload_vae_weights()), call=False)
-    shared.opts.onchange("sd_hypernetwork", wrap_queued_call(lambda: modules.hypernetworks.hypernetwork.load_hypernetwork(shared.opts.sd_hypernetwork)))
+    shared.opts.onchange("sd_hypernetwork", wrap_queued_call(lambda: shared.reload_hypernetworks()))
     shared.opts.onchange("sd_hypernetwork_strength", modules.hypernetworks.hypernetwork.apply_strength)
 
     if cmd_opts.tls_keyfile is not None and cmd_opts.tls_keyfile is not None:
-- 
cgit v1.2.3


From 755df94b2aa62eabd96f900e0dd7ddc83c2f692c Mon Sep 17 00:00:00 2001
From: flamelaw <flamelaw.com3d2@gmail.com>
Date: Sun, 27 Nov 2022 00:35:44 +0900
Subject: set TI AdamW default weight decay to 0

---
 modules/textual_inversion/textual_inversion.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index fee08e33..b9b1394f 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -283,7 +283,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
         shared.sd_model.first_stage_model.to(devices.cpu)
 
     embedding.vec.requires_grad = True
-    optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate)
+    optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate, weight_decay=0.0)
     scaler = torch.cuda.amp.GradScaler()
 
     batch_size = ds.batch_size
-- 
cgit v1.2.3


From 4d5f1691dda971ec7b461dd880426300fd54ccee Mon Sep 17 00:00:00 2001
From: brkirch <brkirch@users.noreply.github.com>
Date: Mon, 28 Nov 2022 21:36:35 -0500
Subject: Use devices.autocast instead of torch.autocast

---
 modules/hypernetworks/hypernetwork.py          | 2 +-
 modules/interrogate.py                         | 3 +--
 modules/swinir_model.py                        | 6 +-----
 modules/textual_inversion/dataset.py           | 4 ++--
 modules/textual_inversion/textual_inversion.py | 2 +-
 5 files changed, 6 insertions(+), 11 deletions(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py
index 8466887f..eb5ae372 100644
--- a/modules/hypernetworks/hypernetwork.py
+++ b/modules/hypernetworks/hypernetwork.py
@@ -495,7 +495,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step,
                 if shared.state.interrupted:
                     break
 
-                with torch.autocast("cuda"):
+                with devices.autocast():
                     x = batch.latent_sample.to(devices.device, non_blocking=pin_memory)
                     if tag_drop_out != 0 or shuffle_tags:
                         shared.sd_model.cond_stage_model.to(devices.device)
diff --git a/modules/interrogate.py b/modules/interrogate.py
index 9769aa34..40c6b082 100644
--- a/modules/interrogate.py
+++ b/modules/interrogate.py
@@ -148,8 +148,7 @@ class InterrogateModels:
 
             clip_image = self.clip_preprocess(pil_image).unsqueeze(0).type(self.dtype).to(devices.device_interrogate)
 
-            precision_scope = torch.autocast if shared.cmd_opts.precision == "autocast" else contextlib.nullcontext
-            with torch.no_grad(), precision_scope("cuda"):
+            with torch.no_grad(), devices.autocast():
                 image_features = self.clip_model.encode_image(clip_image).type(self.dtype)
 
                 image_features /= image_features.norm(dim=-1, keepdim=True)
diff --git a/modules/swinir_model.py b/modules/swinir_model.py
index facd262d..483eabd4 100644
--- a/modules/swinir_model.py
+++ b/modules/swinir_model.py
@@ -13,10 +13,6 @@ from modules.swinir_model_arch import SwinIR as net
 from modules.swinir_model_arch_v2 import Swin2SR as net2
 from modules.upscaler import Upscaler, UpscalerData
 
-precision_scope = (
-    torch.autocast if cmd_opts.precision == "autocast" else contextlib.nullcontext
-)
-
 
 class UpscalerSwinIR(Upscaler):
     def __init__(self, dirname):
@@ -112,7 +108,7 @@ def upscale(
     img = np.moveaxis(img, 2, 0) / 255
     img = torch.from_numpy(img).float()
     img = img.unsqueeze(0).to(devices.device_swinir)
-    with torch.no_grad(), precision_scope("cuda"):
+    with torch.no_grad(), devices.autocast():
         _, _, h_old, w_old = img.size()
         h_pad = (h_old // window_size + 1) * window_size - h_old
         w_pad = (w_old // window_size + 1) * window_size - w_old
diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py
index e5725f33..2dc64c3c 100644
--- a/modules/textual_inversion/dataset.py
+++ b/modules/textual_inversion/dataset.py
@@ -82,7 +82,7 @@ class PersonalizedBase(Dataset):
             torchdata = torch.from_numpy(npimage).permute(2, 0, 1).to(device=device, dtype=torch.float32)
             latent_sample = None
 
-            with torch.autocast("cuda"):
+            with devices.autocast():
                 latent_dist = model.encode_first_stage(torchdata.unsqueeze(dim=0))
 
             if latent_sampling_method == "once" or (latent_sampling_method == "deterministic" and not isinstance(latent_dist, DiagonalGaussianDistribution)):
@@ -101,7 +101,7 @@ class PersonalizedBase(Dataset):
                 entry.cond_text = self.create_text(filename_text)
 
             if include_cond and not (self.tag_drop_out != 0 or self.shuffle_tags):
-                with torch.autocast("cuda"):
+                with devices.autocast():
                     entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0)
 
             self.dataset.append(entry)
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index 4eb75cb5..daf8d1b8 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -316,7 +316,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
                 if shared.state.interrupted:
                     break
 
-                with torch.autocast("cuda"):
+                with devices.autocast():
                     # c = stack_conds(batch.cond).to(devices.device)
                     # mask = torch.tensor(batch.emb_index).to(devices.device, non_blocking=pin_memory)
                     # print(mask)
-- 
cgit v1.2.3


From c0355caefe3d82e304e6d832699d581fc8f9fbf9 Mon Sep 17 00:00:00 2001
From: Jim Hays <jim@emporatitle.com>
Date: Wed, 14 Dec 2022 21:01:32 -0500
Subject: Fix various typos

---
 README.md                                      |  4 ++--
 javascript/contextMenus.js                     | 24 ++++++++++++------------
 javascript/progressbar.js                      | 12 ++++++------
 javascript/ui.js                               |  2 +-
 modules/api/api.py                             | 18 +++++++++---------
 modules/api/models.py                          |  2 +-
 modules/images.py                              |  4 ++--
 modules/processing.py                          | 14 +++++++-------
 modules/safe.py                                |  4 ++--
 modules/scripts.py                             |  4 ++--
 modules/sd_hijack_inpainting.py                |  6 +++---
 modules/sd_hijack_unet.py                      |  2 +-
 modules/textual_inversion/dataset.py           | 10 +++++-----
 modules/textual_inversion/textual_inversion.py | 16 ++++++++--------
 scripts/prompt_matrix.py                       | 10 +++++-----
 webui.py                                       |  4 ++--
 16 files changed, 68 insertions(+), 68 deletions(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/README.md b/README.md
index 55990581..556000fb 100644
--- a/README.md
+++ b/README.md
@@ -82,8 +82,8 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web
 - Use VAEs
 - Estimated completion time in progress bar
 - API
-- Support for dedicated [inpainting model](https://github.com/runwayml/stable-diffusion#inpainting-with-stable-diffusion) by RunwayML. 
-- via extension: [Aesthetic Gradients](https://github.com/AUTOMATIC1111/stable-diffusion-webui-aesthetic-gradients), a way to generate images with a specific aesthetic by using clip images embds (implementation of [https://github.com/vicgalle/stable-diffusion-aesthetic-gradients](https://github.com/vicgalle/stable-diffusion-aesthetic-gradients))
+- Support for dedicated [inpainting model](https://github.com/runwayml/stable-diffusion#inpainting-with-stable-diffusion) by RunwayML.
+- via extension: [Aesthetic Gradients](https://github.com/AUTOMATIC1111/stable-diffusion-webui-aesthetic-gradients), a way to generate images with a specific aesthetic by using clip images embeds (implementation of [https://github.com/vicgalle/stable-diffusion-aesthetic-gradients](https://github.com/vicgalle/stable-diffusion-aesthetic-gradients))
 - [Stable Diffusion 2.0](https://github.com/Stability-AI/stablediffusion) support - see [wiki](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#stable-diffusion-20) for instructions
 
 ## Installation and Running
diff --git a/javascript/contextMenus.js b/javascript/contextMenus.js
index fe67c42e..11bcce1b 100644
--- a/javascript/contextMenus.js
+++ b/javascript/contextMenus.js
@@ -9,7 +9,7 @@ contextMenuInit = function(){
 
   function showContextMenu(event,element,menuEntries){
     let posx = event.clientX + document.body.scrollLeft + document.documentElement.scrollLeft;
-    let posy = event.clientY + document.body.scrollTop + document.documentElement.scrollTop; 
+    let posy = event.clientY + document.body.scrollTop + document.documentElement.scrollTop;
 
     let oldMenu = gradioApp().querySelector('#context-menu')
     if(oldMenu){
@@ -61,15 +61,15 @@ contextMenuInit = function(){
 
   }
 
-  function appendContextMenuOption(targetEmementSelector,entryName,entryFunction){
-    
-    currentItems = menuSpecs.get(targetEmementSelector)
-    
+  function appendContextMenuOption(targetElementSelector,entryName,entryFunction){
+
+    currentItems = menuSpecs.get(targetElementSelector)
+
     if(!currentItems){
       currentItems = []
-      menuSpecs.set(targetEmementSelector,currentItems);
+      menuSpecs.set(targetElementSelector,currentItems);
     }
-    let newItem = {'id':targetEmementSelector+'_'+uid(), 
+    let newItem = {'id':targetElementSelector+'_'+uid(),
                    'name':entryName,
                    'func':entryFunction,
                    'isNew':true}
@@ -97,7 +97,7 @@ contextMenuInit = function(){
       if(source.id && source.id.indexOf('check_progress')>-1){
         return
       }
-      
+
       let oldMenu = gradioApp().querySelector('#context-menu')
       if(oldMenu){
         oldMenu.remove()
@@ -117,7 +117,7 @@ contextMenuInit = function(){
       })
     });
     eventListenerApplied=true
-  
+
   }
 
   return [appendContextMenuOption, removeContextMenuOption, addContextMenuEventListener]
@@ -152,8 +152,8 @@ addContextMenuEventListener = initResponse[2];
     generateOnRepeat('#img2img_generate','#img2img_interrupt');
   })
 
-  let cancelGenerateForever = function(){ 
-    clearInterval(window.generateOnRepeatInterval) 
+  let cancelGenerateForever = function(){
+    clearInterval(window.generateOnRepeatInterval)
   }
 
   appendContextMenuOption('#txt2img_interrupt','Cancel generate forever',cancelGenerateForever)
@@ -162,7 +162,7 @@ addContextMenuEventListener = initResponse[2];
   appendContextMenuOption('#img2img_generate', 'Cancel generate forever',cancelGenerateForever)
 
   appendContextMenuOption('#roll','Roll three',
-    function(){ 
+    function(){
       let rollbutton = get_uiCurrentTabContent().querySelector('#roll');
       setTimeout(function(){rollbutton.click()},100)
       setTimeout(function(){rollbutton.click()},200)
diff --git a/javascript/progressbar.js b/javascript/progressbar.js
index d58737c4..d6323ed9 100644
--- a/javascript/progressbar.js
+++ b/javascript/progressbar.js
@@ -3,7 +3,7 @@ global_progressbars = {}
 galleries = {}
 galleryObservers = {}
 
-// this tracks laumnches of window.setTimeout for progressbar to prevent starting a new timeout when the previous is still running
+// this tracks launches of window.setTimeout for progressbar to prevent starting a new timeout when the previous is still running
 timeoutIds = {}
 
 function check_progressbar(id_part, id_progressbar, id_progressbar_span, id_skip, id_interrupt, id_preview, id_gallery){
@@ -20,21 +20,21 @@ function check_progressbar(id_part, id_progressbar, id_progressbar_span, id_skip
 
     var skip = id_skip ? gradioApp().getElementById(id_skip) : null
     var interrupt = gradioApp().getElementById(id_interrupt)
-    
+
     if(opts.show_progress_in_title && progressbar && progressbar.offsetParent){
         if(progressbar.innerText){
             let newtitle = '[' + progressbar.innerText.trim() + '] Stable Diffusion';
             if(document.title != newtitle){
-                document.title =  newtitle;          
+                document.title =  newtitle;
             }
         }else{
             let newtitle = 'Stable Diffusion'
             if(document.title != newtitle){
-                document.title =  newtitle;          
+                document.title =  newtitle;
             }
         }
     }
-    
+
 	if(progressbar!= null && progressbar != global_progressbars[id_progressbar]){
 	    global_progressbars[id_progressbar] = progressbar
 
@@ -63,7 +63,7 @@ function check_progressbar(id_part, id_progressbar, id_progressbar_span, id_skip
                         skip.style.display = "none"
                     }
                     interrupt.style.display = "none"
-			
+
                     //disconnect observer once generation finished, so user can close selected image if they want
                     if (galleryObservers[id_gallery]) {
                         galleryObservers[id_gallery].disconnect();
diff --git a/javascript/ui.js b/javascript/ui.js
index 2cb280e5..587dd782 100644
--- a/javascript/ui.js
+++ b/javascript/ui.js
@@ -100,7 +100,7 @@ function create_submit_args(args){
 
     // As it is currently, txt2img and img2img send back the previous output args (txt2img_gallery, generation_info, html_info) whenever you generate a new image.
     // This can lead to uploading a huge gallery of previously generated images, which leads to an unnecessary delay between submitting and beginning to generate.
-    // I don't know why gradio is seding outputs along with inputs, but we can prevent sending the image gallery here, which seems to be an issue for some.
+    // I don't know why gradio is sending outputs along with inputs, but we can prevent sending the image gallery here, which seems to be an issue for some.
     // If gradio at some point stops sending outputs, this may break something
     if(Array.isArray(res[res.length - 3])){
         res[res.length - 3] = null
diff --git a/modules/api/api.py b/modules/api/api.py
index 89935a70..33845045 100644
--- a/modules/api/api.py
+++ b/modules/api/api.py
@@ -67,10 +67,10 @@ def encode_pil_to_base64(image):
 class Api:
     def __init__(self, app: FastAPI, queue_lock: Lock):
         if shared.cmd_opts.api_auth:
-            self.credenticals = dict()
+            self.credentials = dict()
             for auth in shared.cmd_opts.api_auth.split(","):
                 user, password = auth.split(":")
-                self.credenticals[user] = password
+                self.credentials[user] = password
 
         self.router = APIRouter()
         self.app = app
@@ -93,7 +93,7 @@ class Api:
         self.add_api_route("/sdapi/v1/hypernetworks", self.get_hypernetworks, methods=["GET"], response_model=List[HypernetworkItem])
         self.add_api_route("/sdapi/v1/face-restorers", self.get_face_restorers, methods=["GET"], response_model=List[FaceRestorerItem])
         self.add_api_route("/sdapi/v1/realesrgan-models", self.get_realesrgan_models, methods=["GET"], response_model=List[RealesrganItem])
-        self.add_api_route("/sdapi/v1/prompt-styles", self.get_promp_styles, methods=["GET"], response_model=List[PromptStyleItem])
+        self.add_api_route("/sdapi/v1/prompt-styles", self.get_prompt_styles, methods=["GET"], response_model=List[PromptStyleItem])
         self.add_api_route("/sdapi/v1/artist-categories", self.get_artists_categories, methods=["GET"], response_model=List[str])
         self.add_api_route("/sdapi/v1/artists", self.get_artists, methods=["GET"], response_model=List[ArtistItem])
 
@@ -102,9 +102,9 @@ class Api:
             return self.app.add_api_route(path, endpoint, dependencies=[Depends(self.auth)], **kwargs)
         return self.app.add_api_route(path, endpoint, **kwargs)
 
-    def auth(self, credenticals: HTTPBasicCredentials = Depends(HTTPBasic())):
-        if credenticals.username in self.credenticals:
-            if compare_digest(credenticals.password, self.credenticals[credenticals.username]):
+    def auth(self, credentials: HTTPBasicCredentials = Depends(HTTPBasic())):
+        if credentials.username in self.credentials:
+            if compare_digest(credentials.password, self.credentials[credentials.username]):
                 return True
 
         raise HTTPException(status_code=401, detail="Incorrect username or password", headers={"WWW-Authenticate": "Basic"})
@@ -239,7 +239,7 @@ class Api:
     def interrogateapi(self, interrogatereq: InterrogateRequest):
         image_b64 = interrogatereq.image
         if image_b64 is None:
-            raise HTTPException(status_code=404, detail="Image not found") 
+            raise HTTPException(status_code=404, detail="Image not found")
 
         img = decode_base64_to_image(image_b64)
         img = img.convert('RGB')
@@ -252,7 +252,7 @@ class Api:
                 processed = deepbooru.model.tag(img)
             else:
                 raise HTTPException(status_code=404, detail="Model not found")
-        
+
         return InterrogateResponse(caption=processed)
 
     def interruptapi(self):
@@ -308,7 +308,7 @@ class Api:
     def get_realesrgan_models(self):
         return [{"name":x.name,"path":x.data_path, "scale":x.scale} for x in get_realesrgan_models(None)]
 
-    def get_promp_styles(self):
+    def get_prompt_styles(self):
         styleList = []
         for k in shared.prompt_styles.styles:
             style = shared.prompt_styles.styles[k]
diff --git a/modules/api/models.py b/modules/api/models.py
index f77951fc..a22bc6b3 100644
--- a/modules/api/models.py
+++ b/modules/api/models.py
@@ -128,7 +128,7 @@ class ExtrasBaseRequest(BaseModel):
     upscaling_resize: float = Field(default=2, title="Upscaling Factor", ge=1, le=4, description="By how much to upscale the image, only used when resize_mode=0.")
     upscaling_resize_w: int = Field(default=512, title="Target Width", ge=1, description="Target width for the upscaler to hit. Only used when resize_mode=1.")
     upscaling_resize_h: int = Field(default=512, title="Target Height", ge=1, description="Target height for the upscaler to hit. Only used when resize_mode=1.")
-    upscaling_crop: bool = Field(default=True, title="Crop to fit", description="Should the upscaler crop the image to fit in the choosen size?")
+    upscaling_crop: bool = Field(default=True, title="Crop to fit", description="Should the upscaler crop the image to fit in the chosen size?")
     upscaler_1: str = Field(default="None", title="Main upscaler", description=f"The name of the main upscaler to use, it has to be one of this list: {' , '.join([x.name for x in sd_upscalers])}")
     upscaler_2: str = Field(default="None", title="Secondary upscaler", description=f"The name of the secondary upscaler to use, it has to be one of this list: {' , '.join([x.name for x in sd_upscalers])}")
     extras_upscaler_2_visibility: float = Field(default=0, title="Secondary upscaler visibility", ge=0, le=1, allow_inf_nan=False, description="Sets the visibility of secondary upscaler, values should be between 0 and 1.")
diff --git a/modules/images.py b/modules/images.py
index 8146f580..93a14289 100644
--- a/modules/images.py
+++ b/modules/images.py
@@ -429,7 +429,7 @@ def save_image(image, path, basename, seed=None, prompt=None, extension='png', i
             The directory to save the image. Note, the option `save_to_dirs` will make the image to be saved into a sub directory.
         basename (`str`):
             The base filename which will be applied to `filename pattern`.
-        seed, prompt, short_filename, 
+        seed, prompt, short_filename,
         extension (`str`):
             Image file extension, default is `png`.
         pngsectionname (`str`):
@@ -590,7 +590,7 @@ def read_info_from_image(image):
 Negative prompt: {json_info["uc"]}
 Steps: {json_info["steps"]}, Sampler: {sampler}, CFG scale: {json_info["scale"]}, Seed: {json_info["seed"]}, Size: {image.width}x{image.height}, Clip skip: 2, ENSD: 31337"""
         except Exception:
-            print(f"Error parsing NovelAI iamge generation parameters:", file=sys.stderr)
+            print(f"Error parsing NovelAI image generation parameters:", file=sys.stderr)
             print(traceback.format_exc(), file=sys.stderr)
 
     return geninfo, items
diff --git a/modules/processing.py b/modules/processing.py
index 24c537d1..fe7f4faf 100644
--- a/modules/processing.py
+++ b/modules/processing.py
@@ -147,11 +147,11 @@ class StableDiffusionProcessing():
 
         # The "masked-image" in this case will just be all zeros since the entire image is masked.
         image_conditioning = torch.zeros(x.shape[0], 3, height, width, device=x.device)
-        image_conditioning = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(image_conditioning)) 
+        image_conditioning = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(image_conditioning))
 
         # Add the fake full 1s mask to the first dimension.
         image_conditioning = torch.nn.functional.pad(image_conditioning, (0, 0, 0, 0, 1, 0), value=1.0)
-        image_conditioning = image_conditioning.to(x.dtype)            
+        image_conditioning = image_conditioning.to(x.dtype)
 
         return image_conditioning
 
@@ -199,7 +199,7 @@ class StableDiffusionProcessing():
             source_image * (1.0 - conditioning_mask),
             getattr(self, "inpainting_mask_weight", shared.opts.inpainting_mask_weight)
         )
-        
+
         # Encode the new masked image using first stage of network.
         conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(conditioning_image))
 
@@ -537,7 +537,7 @@ def process_images_inner(p: StableDiffusionProcessing) -> Processed:
         for n in range(p.n_iter):
             if state.skipped:
                 state.skipped = False
-            
+
             if state.interrupted:
                 break
 
@@ -612,7 +612,7 @@ def process_images_inner(p: StableDiffusionProcessing) -> Processed:
                     image.info["parameters"] = text
                 output_images.append(image)
 
-            del x_samples_ddim 
+            del x_samples_ddim
 
             devices.torch_gc()
 
@@ -704,7 +704,7 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing):
 
         samples = samples[:, :, self.truncate_y//2:samples.shape[2]-self.truncate_y//2, self.truncate_x//2:samples.shape[3]-self.truncate_x//2]
 
-        """saves image before applying hires fix, if enabled in options; takes as an arguyment either an image or batch with latent space images"""
+        """saves image before applying hires fix, if enabled in options; takes as an argument either an image or batch with latent space images"""
         def save_intermediate(image, index):
             if not opts.save or self.do_not_save_samples or not opts.save_images_before_highres_fix:
                 return
@@ -720,7 +720,7 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing):
 
             samples = torch.nn.functional.interpolate(samples, size=(self.height // opt_f, self.width // opt_f), mode="bilinear")
 
-            # Avoid making the inpainting conditioning unless necessary as 
+            # Avoid making the inpainting conditioning unless necessary as
             # this does need some extra compute to decode / encode the image again.
             if getattr(self, "inpainting_mask_weight", shared.opts.inpainting_mask_weight) < 1.0:
                 image_conditioning = self.img2img_image_conditioning(decode_first_stage(self.sd_model, samples), samples)
diff --git a/modules/safe.py b/modules/safe.py
index 10460ad0..20e9d2fa 100644
--- a/modules/safe.py
+++ b/modules/safe.py
@@ -80,7 +80,7 @@ def check_pt(filename, extra_handler):
         # new pytorch format is a zip file
         with zipfile.ZipFile(filename) as z:
             check_zip_filenames(filename, z.namelist())
-            
+
             # find filename of data.pkl in zip file: '<directory name>/data.pkl'
             data_pkl_filenames = [f for f in z.namelist() if data_pkl_re.match(f)]
             if len(data_pkl_filenames) == 0:
@@ -108,7 +108,7 @@ def load(filename, *args, **kwargs):
 
 def load_with_extra(filename, extra_handler=None, *args, **kwargs):
     """
-    this functon is intended to be used by extensions that want to load models with
+    this function is intended to be used by extensions that want to load models with
     some extra classes in them that the usual unpickler would find suspicious.
 
     Use the extra_handler argument to specify a function that takes module and field name as text,
diff --git a/modules/scripts.py b/modules/scripts.py
index 23ca195d..722f8685 100644
--- a/modules/scripts.py
+++ b/modules/scripts.py
@@ -36,7 +36,7 @@ class Script:
     def ui(self, is_img2img):
         """this function should create gradio UI elements. See https://gradio.app/docs/#components
         The return value should be an array of all components that are used in processing.
-        Values of those returned componenbts will be passed to run() and process() functions.
+        Values of those returned components will be passed to run() and process() functions.
         """
 
         pass
@@ -47,7 +47,7 @@ class Script:
 
         This function should return:
          - False if the script should not be shown in UI at all
-         - True if the script should be shown in UI if it's scelected in the scripts drowpdown
+         - True if the script should be shown in UI if it's selected in the scripts dropdown
          - script.AlwaysVisible if the script should be shown in UI at all times
          """
 
diff --git a/modules/sd_hijack_inpainting.py b/modules/sd_hijack_inpainting.py
index 938f9a58..d72f83fd 100644
--- a/modules/sd_hijack_inpainting.py
+++ b/modules/sd_hijack_inpainting.py
@@ -209,7 +209,7 @@ def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=F
         else:
             x_in = torch.cat([x] * 2)
             t_in = torch.cat([t] * 2)
-            
+
             if isinstance(c, dict):
                 assert isinstance(unconditional_conditioning, dict)
                 c_in = dict()
@@ -278,7 +278,7 @@ def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=F
     x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)
 
     return x_prev, pred_x0, e_t
-    
+
 # =================================================================================================
 # Monkey patch LatentInpaintDiffusion to load the checkpoint with a proper config.
 # Adapted from:
@@ -326,7 +326,7 @@ def do_inpainting_hijack():
     # most of this stuff seems to no longer be needed because it is already included into SD2.0
     # LatentInpaintDiffusion remains because SD2.0's LatentInpaintDiffusion can't be loaded without specifying a checkpoint
     # p_sample_plms is needed because PLMS can't work with dicts as conditionings
-    # this file should be cleaned up later if weverything tuens out to work fine
+    # this file should be cleaned up later if everything turns out to work fine
 
     # ldm.models.diffusion.ddpm.get_unconditional_conditioning = get_unconditional_conditioning
     ldm.models.diffusion.ddpm.LatentInpaintDiffusion = LatentInpaintDiffusion
diff --git a/modules/sd_hijack_unet.py b/modules/sd_hijack_unet.py
index 1b9d7757..18daf8c1 100644
--- a/modules/sd_hijack_unet.py
+++ b/modules/sd_hijack_unet.py
@@ -4,7 +4,7 @@ import torch
 class TorchHijackForUnet:
     """
     This is torch, but with cat that resizes tensors to appropriate dimensions if they do not match;
-    this makes it possible to create pictures with dimensions that are muliples of 8 rather than 64
+    this makes it possible to create pictures with dimensions that are multiples of 8 rather than 64
     """
 
     def __getattr__(self, item):
diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py
index 2dc64c3c..88d68c76 100644
--- a/modules/textual_inversion/dataset.py
+++ b/modules/textual_inversion/dataset.py
@@ -28,9 +28,9 @@ class DatasetEntry:
 
 
 class PersonalizedBase(Dataset):
-    def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, cond_model=None, device=None, template_file=None, include_cond=False, batch_size=1, gradient_step=1, shuffle_tags=False, tag_drop_out=0, latent_sampling_method='once'):        
+    def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, cond_model=None, device=None, template_file=None, include_cond=False, batch_size=1, gradient_step=1, shuffle_tags=False, tag_drop_out=0, latent_sampling_method='once'):
         re_word = re.compile(shared.opts.dataset_filename_word_regex) if len(shared.opts.dataset_filename_word_regex) > 0 else None
-        
+
         self.placeholder_token = placeholder_token
 
         self.width = width
@@ -50,14 +50,14 @@ class PersonalizedBase(Dataset):
 
         self.image_paths = [os.path.join(data_root, file_path) for file_path in os.listdir(data_root)]
 
-        
+
         self.shuffle_tags = shuffle_tags
         self.tag_drop_out = tag_drop_out
 
         print("Preparing dataset...")
         for path in tqdm.tqdm(self.image_paths):
             if shared.state.interrupted:
-                raise Exception("inturrupted")
+                raise Exception("interrupted")
             try:
                 image = Image.open(path).convert('RGB').resize((self.width, self.height), PIL.Image.BICUBIC)
             except Exception:
@@ -144,7 +144,7 @@ class PersonalizedDataLoader(DataLoader):
             self.collate_fn = collate_wrapper_random
         else:
             self.collate_fn = collate_wrapper
-        
+
 
 class BatchLoader:
     def __init__(self, data):
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index e28c357a..daf3997b 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -133,7 +133,7 @@ class EmbeddingDatabase:
 
                 process_file(fullfn, fn)
             except Exception:
-                print(f"Error loading emedding {fn}:", file=sys.stderr)
+                print(f"Error loading embedding {fn}:", file=sys.stderr)
                 print(traceback.format_exc(), file=sys.stderr)
                 continue
 
@@ -194,7 +194,7 @@ def write_loss(log_directory, filename, step, epoch_len, values):
             csv_writer.writeheader()
 
         epoch = (step - 1) // epoch_len
-        epoch_step = (step - 1) % epoch_len 
+        epoch_step = (step - 1) % epoch_len
 
         csv_writer.writerow({
             "step": step,
@@ -270,9 +270,9 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
    # dataset loading may take a while, so input validations and early returns should be done before this
     shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..."
     old_parallel_processing_allowed = shared.parallel_processing_allowed
-    
+
     pin_memory = shared.opts.pin_memory
-    
+
     ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, cond_model=shared.sd_model.cond_stage_model, device=devices.device, template_file=template_file, batch_size=batch_size, gradient_step=gradient_step, shuffle_tags=shuffle_tags, tag_drop_out=tag_drop_out, latent_sampling_method=latent_sampling_method)
 
     latent_sampling_method = ds.latent_sampling_method
@@ -295,12 +295,12 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
     loss_step = 0
     _loss_step = 0 #internal
 
-    
+
     last_saved_file = "<none>"
     last_saved_image = "<none>"
     forced_filename = "<none>"
     embedding_yet_to_be_embedded = False
-    
+
     pbar = tqdm.tqdm(total=steps - initial_step)
     try:
         for i in range((steps-initial_step) * gradient_step):
@@ -327,10 +327,10 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
                     c = shared.sd_model.cond_stage_model(batch.cond_text)
                     loss = shared.sd_model(x, c)[0] / gradient_step
                     del x
-                    
+
                     _loss_step += loss.item()
                 scaler.scale(loss).backward()
-                
+
                 # go back until we reach gradient accumulation steps
                 if (j + 1) % gradient_step != 0:
                     continue
diff --git a/scripts/prompt_matrix.py b/scripts/prompt_matrix.py
index c53ca28c..4c79eaef 100644
--- a/scripts/prompt_matrix.py
+++ b/scripts/prompt_matrix.py
@@ -18,7 +18,7 @@ def draw_xy_grid(xs, ys, x_label, y_label, cell):
     ver_texts = [[images.GridAnnotation(y_label(y))] for y in ys]
     hor_texts = [[images.GridAnnotation(x_label(x))] for x in xs]
 
-    first_pocessed = None
+    first_processed = None
 
     state.job_count = len(xs) * len(ys)
 
@@ -27,17 +27,17 @@ def draw_xy_grid(xs, ys, x_label, y_label, cell):
             state.job = f"{ix + iy * len(xs) + 1} out of {len(xs) * len(ys)}"
 
             processed = cell(x, y)
-            if first_pocessed is None:
-                first_pocessed = processed
+            if first_processed is None:
+                first_processed = processed
 
             res.append(processed.images[0])
 
     grid = images.image_grid(res, rows=len(ys))
     grid = images.draw_grid_annotations(grid, res[0].width, res[0].height, hor_texts, ver_texts)
 
-    first_pocessed.images = [grid]
+    first_processed.images = [grid]
 
-    return first_pocessed
+    return first_processed
 
 
 class Script(scripts.Script):
diff --git a/webui.py b/webui.py
index c2d0c6be..4b32e77d 100644
--- a/webui.py
+++ b/webui.py
@@ -153,8 +153,8 @@ def webui():
 
         # gradio uses a very open CORS policy via app.user_middleware, which makes it possible for
         # an attacker to trick the user into opening a malicious HTML page, which makes a request to the
-        # running web ui and do whatever the attcker wants, including installing an extension and
-        # runnnig its code. We disable this here. Suggested by RyotaK.
+        # running web ui and do whatever the attacker wants, including installing an extension and
+        # running its code. We disable this here. Suggested by RyotaK.
         app.user_middleware = [x for x in app.user_middleware if x.cls.__name__ != 'CORSMiddleware']
 
         setup_cors(app)
-- 
cgit v1.2.3


From 3bf5591efe9a9f219c6088be322a87adc4f48f95 Mon Sep 17 00:00:00 2001
From: Yuval Aboulafia <yuval.abou@gmail.com>
Date: Sat, 24 Dec 2022 21:35:29 +0200
Subject: fix F541 f-string without any placeholders

---
 extensions-builtin/LDSR/ldsr_model_arch.py     | 2 +-
 modules/codeformer/vqgan_arch.py               | 4 ++--
 modules/hypernetworks/hypernetwork.py          | 4 ++--
 modules/images.py                              | 2 +-
 modules/interrogate.py                         | 2 +-
 modules/safe.py                                | 8 ++++----
 modules/sd_models.py                           | 8 ++++----
 modules/sd_vae.py                              | 2 +-
 modules/textual_inversion/textual_inversion.py | 2 +-
 scripts/prompts_from_file.py                   | 2 +-
 10 files changed, 18 insertions(+), 18 deletions(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/extensions-builtin/LDSR/ldsr_model_arch.py b/extensions-builtin/LDSR/ldsr_model_arch.py
index f5bd8ae4..0ad49f4e 100644
--- a/extensions-builtin/LDSR/ldsr_model_arch.py
+++ b/extensions-builtin/LDSR/ldsr_model_arch.py
@@ -26,7 +26,7 @@ class LDSR:
         global cached_ldsr_model
 
         if shared.opts.ldsr_cached and cached_ldsr_model is not None:
-            print(f"Loading model from cache")
+            print("Loading model from cache")
             model: torch.nn.Module = cached_ldsr_model
         else:
             print(f"Loading model from {self.modelPath}")
diff --git a/modules/codeformer/vqgan_arch.py b/modules/codeformer/vqgan_arch.py
index c06c590c..e7293683 100644
--- a/modules/codeformer/vqgan_arch.py
+++ b/modules/codeformer/vqgan_arch.py
@@ -382,7 +382,7 @@ class VQAutoEncoder(nn.Module):
                 self.load_state_dict(torch.load(model_path, map_location='cpu')['params'])
                 logger.info(f'vqgan is loaded from: {model_path} [params]')
             else:
-                raise ValueError(f'Wrong params!')
+                raise ValueError('Wrong params!')
 
 
     def forward(self, x):
@@ -431,7 +431,7 @@ class VQGANDiscriminator(nn.Module):
             elif 'params' in chkpt:
                 self.load_state_dict(torch.load(model_path, map_location='cpu')['params'])
             else:
-                raise ValueError(f'Wrong params!')
+                raise ValueError('Wrong params!')
 
     def forward(self, x):
         return self.main(x)
\ No newline at end of file
diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py
index c406ffb3..9d3034ae 100644
--- a/modules/hypernetworks/hypernetwork.py
+++ b/modules/hypernetworks/hypernetwork.py
@@ -277,7 +277,7 @@ def load_hypernetwork(filename):
             print(traceback.format_exc(), file=sys.stderr)
     else:
         if shared.loaded_hypernetwork is not None:
-            print(f"Unloading hypernetwork")
+            print("Unloading hypernetwork")
 
         shared.loaded_hypernetwork = None
 
@@ -417,7 +417,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step,
 
     initial_step = hypernetwork.step or 0
     if initial_step >= steps:
-        shared.state.textinfo = f"Model has already been trained beyond specified max steps"
+        shared.state.textinfo = "Model has already been trained beyond specified max steps"
         return hypernetwork, filename
 
     scheduler = LearnRateScheduler(learn_rate, steps, initial_step)
diff --git a/modules/images.py b/modules/images.py
index 809ad9f7..31d4528d 100644
--- a/modules/images.py
+++ b/modules/images.py
@@ -599,7 +599,7 @@ def read_info_from_image(image):
 Negative prompt: {json_info["uc"]}
 Steps: {json_info["steps"]}, Sampler: {sampler}, CFG scale: {json_info["scale"]}, Seed: {json_info["seed"]}, Size: {image.width}x{image.height}, Clip skip: 2, ENSD: 31337"""
         except Exception:
-            print(f"Error parsing NovelAI image generation parameters:", file=sys.stderr)
+            print("Error parsing NovelAI image generation parameters:", file=sys.stderr)
             print(traceback.format_exc(), file=sys.stderr)
 
     return geninfo, items
diff --git a/modules/interrogate.py b/modules/interrogate.py
index 0068b81c..46935210 100644
--- a/modules/interrogate.py
+++ b/modules/interrogate.py
@@ -172,7 +172,7 @@ class InterrogateModels:
                             res += ", " + match
 
         except Exception:
-            print(f"Error interrogating", file=sys.stderr)
+            print("Error interrogating", file=sys.stderr)
             print(traceback.format_exc(), file=sys.stderr)
             res += "<error>"
 
diff --git a/modules/safe.py b/modules/safe.py
index 479c8b86..1d4c20b9 100644
--- a/modules/safe.py
+++ b/modules/safe.py
@@ -137,15 +137,15 @@ def load_with_extra(filename, extra_handler=None, *args, **kwargs):
     except pickle.UnpicklingError:
         print(f"Error verifying pickled file from {filename}:", file=sys.stderr)
         print(traceback.format_exc(), file=sys.stderr)
-        print(f"-----> !!!! The file is most likely corrupted !!!! <-----", file=sys.stderr)
-        print(f"You can skip this check with --disable-safe-unpickle commandline argument, but that is not going to help you.\n\n", file=sys.stderr)
+        print("-----> !!!! The file is most likely corrupted !!!! <-----", file=sys.stderr)
+        print("You can skip this check with --disable-safe-unpickle commandline argument, but that is not going to help you.\n\n", file=sys.stderr)
         return None
 
     except Exception:
         print(f"Error verifying pickled file from {filename}:", file=sys.stderr)
         print(traceback.format_exc(), file=sys.stderr)
-        print(f"\nThe file may be malicious, so the program is not going to read it.", file=sys.stderr)
-        print(f"You can skip this check with --disable-safe-unpickle commandline argument.\n\n", file=sys.stderr)
+        print("\nThe file may be malicious, so the program is not going to read it.", file=sys.stderr)
+        print("You can skip this check with --disable-safe-unpickle commandline argument.\n\n", file=sys.stderr)
         return None
 
     return unsafe_torch_load(filename, *args, **kwargs)
diff --git a/modules/sd_models.py b/modules/sd_models.py
index 6ca06211..ecdd91c5 100644
--- a/modules/sd_models.py
+++ b/modules/sd_models.py
@@ -117,13 +117,13 @@ def select_checkpoint():
         return checkpoint_info
 
     if len(checkpoints_list) == 0:
-        print(f"No checkpoints found. When searching for checkpoints, looked at:", file=sys.stderr)
+        print("No checkpoints found. When searching for checkpoints, looked at:", file=sys.stderr)
         if shared.cmd_opts.ckpt is not None:
             print(f" - file {os.path.abspath(shared.cmd_opts.ckpt)}", file=sys.stderr)
         print(f" - directory {model_path}", file=sys.stderr)
         if shared.cmd_opts.ckpt_dir is not None:
             print(f" - directory {os.path.abspath(shared.cmd_opts.ckpt_dir)}", file=sys.stderr)
-        print(f"Can't run without a checkpoint. Find and place a .ckpt file into any of those locations. The program will exit.", file=sys.stderr)
+        print("Can't run without a checkpoint. Find and place a .ckpt file into any of those locations. The program will exit.", file=sys.stderr)
         exit(1)
 
     checkpoint_info = next(iter(checkpoints_list.values()))
@@ -324,7 +324,7 @@ def load_model(checkpoint_info=None):
 
     script_callbacks.model_loaded_callback(sd_model)
 
-    print(f"Model loaded.")
+    print("Model loaded.")
     return sd_model
 
 
@@ -359,5 +359,5 @@ def reload_model_weights(sd_model=None, info=None):
     if not shared.cmd_opts.lowvram and not shared.cmd_opts.medvram:
         sd_model.to(devices.device)
 
-    print(f"Weights loaded.")
+    print("Weights loaded.")
     return sd_model
diff --git a/modules/sd_vae.py b/modules/sd_vae.py
index 25638a83..3856418e 100644
--- a/modules/sd_vae.py
+++ b/modules/sd_vae.py
@@ -208,5 +208,5 @@ def reload_vae_weights(sd_model=None, vae_file="auto"):
     if not shared.cmd_opts.lowvram and not shared.cmd_opts.medvram:
         sd_model.to(devices.device)
 
-    print(f"VAE Weights loaded.")
+    print("VAE Weights loaded.")
     return sd_model
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index daf3997b..f6112578 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -263,7 +263,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
 
     initial_step = embedding.step or 0
     if initial_step >= steps:
-        shared.state.textinfo = f"Model has already been trained beyond specified max steps"
+        shared.state.textinfo = "Model has already been trained beyond specified max steps"
         return embedding, filename
     scheduler = LearnRateScheduler(learn_rate, steps, initial_step)
 
diff --git a/scripts/prompts_from_file.py b/scripts/prompts_from_file.py
index 6e118ddb..e8386ed2 100644
--- a/scripts/prompts_from_file.py
+++ b/scripts/prompts_from_file.py
@@ -140,7 +140,7 @@ class Script(scripts.Script):
                 try:
                     args = cmdargs(line)
                 except Exception:
-                    print(f"Error parsing line [line] as commandline:", file=sys.stderr)
+                    print(f"Error parsing line {line} as commandline:", file=sys.stderr)
                     print(traceback.format_exc(), file=sys.stderr)
                     args = {"prompt": line}
             else:
-- 
cgit v1.2.3


From f55ac33d446185680604e872ceda2ae858821d5c Mon Sep 17 00:00:00 2001
From: Vladimir Mandic <mandic00@live.com>
Date: Sat, 31 Dec 2022 11:27:02 -0500
Subject: validate textual inversion embeddings

---
 modules/sd_models.py                           |  3 ++
 modules/textual_inversion/textual_inversion.py | 43 +++++++++++++++++++++++---
 modules/ui.py                                  |  2 --
 3 files changed, 41 insertions(+), 7 deletions(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/sd_models.py b/modules/sd_models.py
index ecdd91c5..ebd4dff7 100644
--- a/modules/sd_models.py
+++ b/modules/sd_models.py
@@ -325,6 +325,9 @@ def load_model(checkpoint_info=None):
     script_callbacks.model_loaded_callback(sd_model)
 
     print("Model loaded.")
+
+    sd_hijack.model_hijack.embedding_db.load_textual_inversion_embeddings(force_reload = True) # Reload embeddings after model load as they may or may not fit the model
+
     return sd_model
 
 
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index f6112578..103ace60 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -23,6 +23,8 @@ class Embedding:
         self.vec = vec
         self.name = name
         self.step = step
+        self.shape = None
+        self.vectors = 0
         self.cached_checksum = None
         self.sd_checkpoint = None
         self.sd_checkpoint_name = None
@@ -57,8 +59,10 @@ class EmbeddingDatabase:
     def __init__(self, embeddings_dir):
         self.ids_lookup = {}
         self.word_embeddings = {}
+        self.skipped_embeddings = []
         self.dir_mtime = None
         self.embeddings_dir = embeddings_dir
+        self.expected_shape = -1
 
     def register_embedding(self, embedding, model):
 
@@ -75,14 +79,35 @@ class EmbeddingDatabase:
 
         return embedding
 
-    def load_textual_inversion_embeddings(self):
+    def get_expected_shape(self):
+        expected_shape = -1 # initialize with unknown
+        idx = torch.tensor(0).to(shared.device)
+        if expected_shape == -1:
+            try: # matches sd15 signature
+                first_embedding = shared.sd_model.cond_stage_model.wrapped.transformer.text_model.embeddings.token_embedding.wrapped(idx)
+                expected_shape = first_embedding.shape[0]
+            except:
+                pass
+        if expected_shape == -1:
+            try: # matches sd20 signature
+                first_embedding = shared.sd_model.cond_stage_model.wrapped.model.token_embedding.wrapped(idx)
+                expected_shape = first_embedding.shape[0]
+            except:
+                pass
+        if expected_shape == -1:
+            print('Could not determine expected embeddings shape from model')
+        return expected_shape
+
+    def load_textual_inversion_embeddings(self, force_reload = False):
         mt = os.path.getmtime(self.embeddings_dir)
-        if self.dir_mtime is not None and mt <= self.dir_mtime:
+        if not force_reload and self.dir_mtime is not None and mt <= self.dir_mtime:
             return
 
         self.dir_mtime = mt
         self.ids_lookup.clear()
         self.word_embeddings.clear()
+        self.skipped_embeddings = []
+        self.expected_shape = self.get_expected_shape()
 
         def process_file(path, filename):
             name = os.path.splitext(filename)[0]
@@ -122,7 +147,14 @@ class EmbeddingDatabase:
             embedding.step = data.get('step', None)
             embedding.sd_checkpoint = data.get('sd_checkpoint', None)
             embedding.sd_checkpoint_name = data.get('sd_checkpoint_name', None)
-            self.register_embedding(embedding, shared.sd_model)
+            embedding.vectors = vec.shape[0]
+            embedding.shape = vec.shape[-1]
+
+            if (self.expected_shape == -1) or (self.expected_shape == embedding.shape):
+                self.register_embedding(embedding, shared.sd_model)
+            else:
+                self.skipped_embeddings.append(name)
+                # print('Skipping embedding {name}: shape was {shape} expected {expected}'.format(name = name, shape = embedding.shape, expected = self.expected_shape))
 
         for fn in os.listdir(self.embeddings_dir):
             try:
@@ -137,8 +169,9 @@ class EmbeddingDatabase:
                 print(traceback.format_exc(), file=sys.stderr)
                 continue
 
-        print(f"Loaded a total of {len(self.word_embeddings)} textual inversion embeddings.")
-        print("Embeddings:", ', '.join(self.word_embeddings.keys()))
+        print("Textual inversion embeddings {num} loaded: {val}".format(num = len(self.word_embeddings), val = ', '.join(self.word_embeddings.keys())))
+        if (len(self.skipped_embeddings) > 0):
+            print("Textual inversion embeddings {num} skipped: {val}".format(num = len(self.skipped_embeddings), val = ', '.join(self.skipped_embeddings)))
 
     def find_embedding_at_position(self, tokens, offset):
         token = tokens[offset]
diff --git a/modules/ui.py b/modules/ui.py
index 57ee0465..397dd804 100644
--- a/modules/ui.py
+++ b/modules/ui.py
@@ -1157,8 +1157,6 @@ def create_ui():
             with gr.Column(variant='panel'):
                 submit_result = gr.Textbox(elem_id="modelmerger_result", show_label=False)
 
-    sd_hijack.model_hijack.embedding_db.load_textual_inversion_embeddings()
-
     with gr.Blocks(analytics_enabled=False) as train_interface:
         with gr.Row().style(equal_height=False):
             gr.HTML(value="<p style='margin-bottom: 0.7em'>See <b><a href=\"https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Textual-Inversion\">wiki</a></b> for detailed explanation.</p>")
-- 
cgit v1.2.3


From bdbe09827b39be63c9c0b3636132ca58da38ebf6 Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Sat, 31 Dec 2022 22:49:09 +0300
Subject: changed embedding accepted shape detection to use existing code and
 support the new alt-diffusion model, and reformatted messages a bit #6149

---
 modules/textual_inversion/textual_inversion.py | 30 ++++++--------------------
 1 file changed, 6 insertions(+), 24 deletions(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index 103ace60..66f40367 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -80,23 +80,8 @@ class EmbeddingDatabase:
         return embedding
 
     def get_expected_shape(self):
-        expected_shape = -1 # initialize with unknown
-        idx = torch.tensor(0).to(shared.device)
-        if expected_shape == -1:
-            try: # matches sd15 signature
-                first_embedding = shared.sd_model.cond_stage_model.wrapped.transformer.text_model.embeddings.token_embedding.wrapped(idx)
-                expected_shape = first_embedding.shape[0]
-            except:
-                pass
-        if expected_shape == -1:
-            try: # matches sd20 signature
-                first_embedding = shared.sd_model.cond_stage_model.wrapped.model.token_embedding.wrapped(idx)
-                expected_shape = first_embedding.shape[0]
-            except:
-                pass
-        if expected_shape == -1:
-            print('Could not determine expected embeddings shape from model')
-        return expected_shape
+        vec = shared.sd_model.cond_stage_model.encode_embedding_init_text(",", 1)
+        return vec.shape[1]
 
     def load_textual_inversion_embeddings(self, force_reload = False):
         mt = os.path.getmtime(self.embeddings_dir)
@@ -112,8 +97,6 @@ class EmbeddingDatabase:
         def process_file(path, filename):
             name = os.path.splitext(filename)[0]
 
-            data = []
-
             if os.path.splitext(filename.upper())[-1] in ['.PNG', '.WEBP', '.JXL', '.AVIF']:
                 embed_image = Image.open(path)
                 if hasattr(embed_image, 'text') and 'sd-ti-embedding' in embed_image.text:
@@ -150,11 +133,10 @@ class EmbeddingDatabase:
             embedding.vectors = vec.shape[0]
             embedding.shape = vec.shape[-1]
 
-            if (self.expected_shape == -1) or (self.expected_shape == embedding.shape):
+            if self.expected_shape == -1 or self.expected_shape == embedding.shape:
                 self.register_embedding(embedding, shared.sd_model)
             else:
                 self.skipped_embeddings.append(name)
-                # print('Skipping embedding {name}: shape was {shape} expected {expected}'.format(name = name, shape = embedding.shape, expected = self.expected_shape))
 
         for fn in os.listdir(self.embeddings_dir):
             try:
@@ -169,9 +151,9 @@ class EmbeddingDatabase:
                 print(traceback.format_exc(), file=sys.stderr)
                 continue
 
-        print("Textual inversion embeddings {num} loaded: {val}".format(num = len(self.word_embeddings), val = ', '.join(self.word_embeddings.keys())))
-        if (len(self.skipped_embeddings) > 0):
-            print("Textual inversion embeddings {num} skipped: {val}".format(num = len(self.skipped_embeddings), val = ', '.join(self.skipped_embeddings)))
+        print(f"Textual inversion embeddings loaded({len(self.word_embeddings)}): {', '.join(self.word_embeddings.keys())}")
+        if len(self.skipped_embeddings) > 0:
+            print(f"Textual inversion embeddings skipped({len(self.skipped_embeddings)}): {', '.join(self.skipped_embeddings)}")
 
     def find_embedding_at_position(self, tokens, offset):
         token = tokens[offset]
-- 
cgit v1.2.3


From 311354c0bb8930ea939d6aa6b3edd50c69301320 Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Mon, 2 Jan 2023 00:38:09 +0300
Subject: fix the issue with training on SD2.0

---
 modules/sd_models.py                           | 2 ++
 modules/textual_inversion/textual_inversion.py | 3 +--
 2 files changed, 3 insertions(+), 2 deletions(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/sd_models.py b/modules/sd_models.py
index ebd4dff7..bff8d6c9 100644
--- a/modules/sd_models.py
+++ b/modules/sd_models.py
@@ -228,6 +228,8 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"):
     model.sd_model_checkpoint = checkpoint_file
     model.sd_checkpoint_info = checkpoint_info
 
+    model.logvar = model.logvar.to(devices.device)  # fix for training
+
     sd_vae.delete_base_vae()
     sd_vae.clear_loaded_vae()
     vae_file = sd_vae.resolve_vae(checkpoint_file, vae_file=vae_file)
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index 66f40367..1e5722e7 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -282,7 +282,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
         return embedding, filename
     scheduler = LearnRateScheduler(learn_rate, steps, initial_step)
 
-   # dataset loading may take a while, so input validations and early returns should be done before this
+    # dataset loading may take a while, so input validations and early returns should be done before this
     shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..."
     old_parallel_processing_allowed = shared.parallel_processing_allowed
 
@@ -310,7 +310,6 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
     loss_step = 0
     _loss_step = 0 #internal
 
-
     last_saved_file = "<none>"
     last_saved_image = "<none>"
     forced_filename = "<none>"
-- 
cgit v1.2.3


From c65909ad16a1962129114c6251de092f49479b06 Mon Sep 17 00:00:00 2001
From: Philpax <me@philpax.me>
Date: Mon, 2 Jan 2023 12:21:22 +1100
Subject: feat(api): return more data for embeddings

---
 modules/api/api.py                             | 17 +++++++++++++++--
 modules/api/models.py                          | 11 +++++++++--
 modules/textual_inversion/textual_inversion.py |  8 ++++----
 3 files changed, 28 insertions(+), 8 deletions(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/api/api.py b/modules/api/api.py
index 30bf3dac..9c670f00 100644
--- a/modules/api/api.py
+++ b/modules/api/api.py
@@ -330,9 +330,22 @@ class Api:
 
     def get_embeddings(self):
         db = sd_hijack.model_hijack.embedding_db
+
+        def convert_embedding(embedding):
+            return {
+                "step": embedding.step,
+                "sd_checkpoint": embedding.sd_checkpoint,
+                "sd_checkpoint_name": embedding.sd_checkpoint_name,
+                "shape": embedding.shape,
+                "vectors": embedding.vectors,
+            }
+
+        def convert_embeddings(embeddings):
+            return {embedding.name: convert_embedding(embedding) for embedding in embeddings.values()}
+
         return {
-            "loaded": sorted(db.word_embeddings.keys()),
-            "skipped": sorted(db.skipped_embeddings),
+            "loaded": convert_embeddings(db.word_embeddings),
+            "skipped": convert_embeddings(db.skipped_embeddings),
         }
 
     def refresh_checkpoints(self):
diff --git a/modules/api/models.py b/modules/api/models.py
index a8472dc9..4a632c68 100644
--- a/modules/api/models.py
+++ b/modules/api/models.py
@@ -249,6 +249,13 @@ class ArtistItem(BaseModel):
     score: float = Field(title="Score")
     category: str = Field(title="Category")
 
+class EmbeddingItem(BaseModel):
+    step: Optional[int] = Field(title="Step", description="The number of steps that were used to train this embedding, if available")
+    sd_checkpoint: Optional[str] = Field(title="SD Checkpoint", description="The hash of the checkpoint this embedding was trained on, if available")
+    sd_checkpoint_name: Optional[str] = Field(title="SD Checkpoint Name", description="The name of the checkpoint this embedding was trained on, if available. Note that this is the name that was used by the trainer; for a stable identifier, use `sd_checkpoint` instead")
+    shape: int = Field(title="Shape", description="The length of each individual vector in the embedding")
+    vectors: int = Field(title="Vectors", description="The number of vectors in the embedding")
+
 class EmbeddingsResponse(BaseModel):
-    loaded: List[str] = Field(title="Loaded", description="Embeddings loaded for the current model")
-    skipped: List[str] = Field(title="Skipped", description="Embeddings skipped for the current model (likely due to architecture incompatibility)")
\ No newline at end of file
+    loaded: Dict[str, EmbeddingItem] = Field(title="Loaded", description="Embeddings loaded for the current model")
+    skipped: Dict[str, EmbeddingItem] = Field(title="Skipped", description="Embeddings skipped for the current model (likely due to architecture incompatibility)")
\ No newline at end of file
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index 1e5722e7..fd253477 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -59,7 +59,7 @@ class EmbeddingDatabase:
     def __init__(self, embeddings_dir):
         self.ids_lookup = {}
         self.word_embeddings = {}
-        self.skipped_embeddings = []
+        self.skipped_embeddings = {}
         self.dir_mtime = None
         self.embeddings_dir = embeddings_dir
         self.expected_shape = -1
@@ -91,7 +91,7 @@ class EmbeddingDatabase:
         self.dir_mtime = mt
         self.ids_lookup.clear()
         self.word_embeddings.clear()
-        self.skipped_embeddings = []
+        self.skipped_embeddings.clear()
         self.expected_shape = self.get_expected_shape()
 
         def process_file(path, filename):
@@ -136,7 +136,7 @@ class EmbeddingDatabase:
             if self.expected_shape == -1 or self.expected_shape == embedding.shape:
                 self.register_embedding(embedding, shared.sd_model)
             else:
-                self.skipped_embeddings.append(name)
+                self.skipped_embeddings[name] = embedding
 
         for fn in os.listdir(self.embeddings_dir):
             try:
@@ -153,7 +153,7 @@ class EmbeddingDatabase:
 
         print(f"Textual inversion embeddings loaded({len(self.word_embeddings)}): {', '.join(self.word_embeddings.keys())}")
         if len(self.skipped_embeddings) > 0:
-            print(f"Textual inversion embeddings skipped({len(self.skipped_embeddings)}): {', '.join(self.skipped_embeddings)}")
+            print(f"Textual inversion embeddings skipped({len(self.skipped_embeddings)}): {', '.join(self.skipped_embeddings.keys())}")
 
     def find_embedding_at_position(self, tokens, offset):
         token = tokens[offset]
-- 
cgit v1.2.3


From bddebe09edeb6a18f2c06986d5658a7be3a563ea Mon Sep 17 00:00:00 2001
From: Shondoit <shondoit@gmail.com>
Date: Tue, 3 Jan 2023 10:26:37 +0100
Subject: Save Optimizer next to TI embedding

Also add check to load only .PT and .BIN files as embeddings. (since we add .optim files in the same directory)
---
 modules/shared.py                              |  2 +-
 modules/textual_inversion/textual_inversion.py | 40 ++++++++++++++++++++------
 2 files changed, 33 insertions(+), 9 deletions(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/shared.py b/modules/shared.py
index 23657a93..c541d18c 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -355,7 +355,7 @@ options_templates.update(options_section(('system', "System"), {
 options_templates.update(options_section(('training', "Training"), {
     "unload_models_when_training": OptionInfo(False, "Move VAE and CLIP to RAM when training if possible. Saves VRAM."),
     "pin_memory": OptionInfo(False, "Turn on pin_memory for DataLoader. Makes training slightly faster but can increase memory usage."),
-    "save_optimizer_state": OptionInfo(False, "Saves Optimizer state as separate *.optim file. Training can be resumed with HN itself and matching optim file."),
+    "save_optimizer_state": OptionInfo(False, "Saves Optimizer state as separate *.optim file. Training of embedding or HN can be resumed with the matching optim file."),
     "dataset_filename_word_regex": OptionInfo("", "Filename word regex"),
     "dataset_filename_join_string": OptionInfo(" ", "Filename join string"),
     "training_image_repeats_per_epoch": OptionInfo(1, "Number of repeats for a single input image per epoch; used only for displaying epoch number", gr.Number, {"precision": 0}),
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index fd253477..16176e90 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -28,6 +28,7 @@ class Embedding:
         self.cached_checksum = None
         self.sd_checkpoint = None
         self.sd_checkpoint_name = None
+        self.optimizer_state_dict = None
 
     def save(self, filename):
         embedding_data = {
@@ -41,6 +42,13 @@ class Embedding:
 
         torch.save(embedding_data, filename)
 
+        if shared.opts.save_optimizer_state and self.optimizer_state_dict is not None:
+            optimizer_saved_dict = {
+                'hash': self.checksum(),
+                'optimizer_state_dict': self.optimizer_state_dict,
+            }
+            torch.save(optimizer_saved_dict, filename + '.optim')
+
     def checksum(self):
         if self.cached_checksum is not None:
             return self.cached_checksum
@@ -95,9 +103,10 @@ class EmbeddingDatabase:
         self.expected_shape = self.get_expected_shape()
 
         def process_file(path, filename):
-            name = os.path.splitext(filename)[0]
+            name, ext = os.path.splitext(filename)
+            ext = ext.upper()
 
-            if os.path.splitext(filename.upper())[-1] in ['.PNG', '.WEBP', '.JXL', '.AVIF']:
+            if ext in ['.PNG', '.WEBP', '.JXL', '.AVIF']:
                 embed_image = Image.open(path)
                 if hasattr(embed_image, 'text') and 'sd-ti-embedding' in embed_image.text:
                     data = embedding_from_b64(embed_image.text['sd-ti-embedding'])
@@ -105,8 +114,10 @@ class EmbeddingDatabase:
                 else:
                     data = extract_image_data_embed(embed_image)
                     name = data.get('name', name)
-            else:
+            elif ext in ['.BIN', '.PT']:
                 data = torch.load(path, map_location="cpu")
+            else:
+                return
 
             # textual inversion embeddings
             if 'string_to_param' in data:
@@ -300,6 +311,20 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
 
     embedding.vec.requires_grad = True
     optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate, weight_decay=0.0)
+    if shared.opts.save_optimizer_state:
+        optimizer_state_dict = None
+        if os.path.exists(filename + '.optim'):
+            optimizer_saved_dict = torch.load(filename + '.optim', map_location='cpu')
+            if embedding.checksum() == optimizer_saved_dict.get('hash', None):
+                optimizer_state_dict = optimizer_saved_dict.get('optimizer_state_dict', None)
+    
+        if optimizer_state_dict is not None:
+            optimizer.load_state_dict(optimizer_state_dict)
+            print("Loaded existing optimizer from checkpoint")
+        else:
+            print("No saved optimizer exists in checkpoint")
+
+
     scaler = torch.cuda.amp.GradScaler()
 
     batch_size = ds.batch_size
@@ -366,9 +391,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
                     # Before saving, change name to match current checkpoint.
                     embedding_name_every = f'{embedding_name}-{steps_done}'
                     last_saved_file = os.path.join(embedding_dir, f'{embedding_name_every}.pt')
-                    #if shared.opts.save_optimizer_state:
-                        #embedding.optimizer_state_dict = optimizer.state_dict()
-                    save_embedding(embedding, checkpoint, embedding_name_every, last_saved_file, remove_cached_checksum=True)
+                    save_embedding(embedding, optimizer, checkpoint, embedding_name_every, last_saved_file, remove_cached_checksum=True)
                     embedding_yet_to_be_embedded = True
 
                 write_loss(log_directory, "textual_inversion_loss.csv", embedding.step, steps_per_epoch, {
@@ -458,7 +481,7 @@ Last saved image: {html.escape(last_saved_image)}<br/>
 </p>
 """
         filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt')
-        save_embedding(embedding, checkpoint, embedding_name, filename, remove_cached_checksum=True)
+        save_embedding(embedding, optimizer, checkpoint, embedding_name, filename, remove_cached_checksum=True)
     except Exception:
         print(traceback.format_exc(), file=sys.stderr)
         pass
@@ -470,7 +493,7 @@ Last saved image: {html.escape(last_saved_image)}<br/>
 
     return embedding, filename
 
-def save_embedding(embedding, checkpoint, embedding_name, filename, remove_cached_checksum=True):
+def save_embedding(embedding, optimizer, checkpoint, embedding_name, filename, remove_cached_checksum=True):
     old_embedding_name = embedding.name
     old_sd_checkpoint = embedding.sd_checkpoint if hasattr(embedding, "sd_checkpoint") else None
     old_sd_checkpoint_name = embedding.sd_checkpoint_name if hasattr(embedding, "sd_checkpoint_name") else None
@@ -481,6 +504,7 @@ def save_embedding(embedding, checkpoint, embedding_name, filename, remove_cache
         if remove_cached_checksum:
             embedding.cached_checksum = None
         embedding.name = embedding_name
+        embedding.optimizer_state_dict = optimizer.state_dict()
         embedding.save(filename)
     except:
         embedding.sd_checkpoint = old_sd_checkpoint
-- 
cgit v1.2.3


From 192ddc04d6de0d780f73aa5fbaa8c66cd4642e1c Mon Sep 17 00:00:00 2001
From: Vladimir Mandic <mandic00@live.com>
Date: Tue, 3 Jan 2023 10:34:51 -0500
Subject: add job info to modules

---
 modules/extras.py                              | 17 +++++++++++++----
 modules/hypernetworks/hypernetwork.py          |  1 +
 modules/textual_inversion/preprocess.py        |  1 +
 modules/textual_inversion/textual_inversion.py |  1 +
 4 files changed, 16 insertions(+), 4 deletions(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/extras.py b/modules/extras.py
index 7e222313..d665440a 100644
--- a/modules/extras.py
+++ b/modules/extras.py
@@ -58,6 +58,9 @@ cached_images: LruCache = LruCache(max_size=5)
 def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_dir, show_extras_results, gfpgan_visibility, codeformer_visibility, codeformer_weight, upscaling_resize, upscaling_resize_w, upscaling_resize_h, upscaling_crop, extras_upscaler_1, extras_upscaler_2, extras_upscaler_2_visibility, upscale_first: bool, save_output: bool = True):
     devices.torch_gc()
 
+    shared.state.begin()
+    shared.state.job = 'extras'
+
     imageArr = []
     # Also keep track of original file names
     imageNameArr = []
@@ -94,6 +97,7 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_
     # Extra operation definitions
 
     def run_gfpgan(image: Image.Image, info: str) -> Tuple[Image.Image, str]:
+        shared.state.job = 'extras-gfpgan'
         restored_img = modules.gfpgan_model.gfpgan_fix_faces(np.array(image, dtype=np.uint8))
         res = Image.fromarray(restored_img)
 
@@ -104,6 +108,7 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_
         return (res, info)
 
     def run_codeformer(image: Image.Image, info: str) -> Tuple[Image.Image, str]:
+        shared.state.job = 'extras-codeformer'
         restored_img = modules.codeformer_model.codeformer.restore(np.array(image, dtype=np.uint8), w=codeformer_weight)
         res = Image.fromarray(restored_img)
 
@@ -114,6 +119,7 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_
         return (res, info)
 
     def upscale(image, scaler_index, resize, mode, resize_w, resize_h, crop):
+        shared.state.job = 'extras-upscale'
         upscaler = shared.sd_upscalers[scaler_index]
         res = upscaler.scaler.upscale(image, resize, upscaler.data_path)
         if mode == 1 and crop:
@@ -180,6 +186,9 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_
     for image, image_name in zip(imageArr, imageNameArr):
         if image is None:
             return outputs, "Please select an input image.", ''
+
+        shared.state.textinfo = f'Processing image {image_name}'
+        
         existing_pnginfo = image.info or {}
 
         image = image.convert("RGB")
@@ -193,6 +202,10 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_
         else:
             basename = ''
 
+        if opts.enable_pnginfo: # append info before save
+            image.info = existing_pnginfo
+            image.info["extras"] = info
+
         if save_output:
             # Add upscaler name as a suffix.
             suffix = f"-{shared.sd_upscalers[extras_upscaler_1].name}" if shared.opts.use_upscaler_name_as_suffix else ""
@@ -203,10 +216,6 @@ def run_extras(extras_mode, resize_mode, image, image_folder, input_dir, output_
             images.save_image(image, path=outpath, basename=basename, seed=None, prompt=None, extension=opts.samples_format, info=info, short_filename=True,
                             no_prompt=True, grid=False, pnginfo_section_name="extras", existing_info=existing_pnginfo, forced_filename=None, suffix=suffix)
 
-        if opts.enable_pnginfo:
-            image.info = existing_pnginfo
-            image.info["extras"] = info
-
         if extras_mode != 2 or show_extras_results :
             outputs.append(image)
 
diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py
index 109e8078..450fecac 100644
--- a/modules/hypernetworks/hypernetwork.py
+++ b/modules/hypernetworks/hypernetwork.py
@@ -417,6 +417,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step,
     shared.loaded_hypernetwork = Hypernetwork()
     shared.loaded_hypernetwork.load(path)
 
+    shared.state.job = "train-hypernetwork"
     shared.state.textinfo = "Initializing hypernetwork training..."
     shared.state.job_count = steps
 
diff --git a/modules/textual_inversion/preprocess.py b/modules/textual_inversion/preprocess.py
index 56b9b2eb..feb876c6 100644
--- a/modules/textual_inversion/preprocess.py
+++ b/modules/textual_inversion/preprocess.py
@@ -124,6 +124,7 @@ def preprocess_work(process_src, process_dst, process_width, process_height, pre
 
     files = listfiles(src)
 
+    shared.state.job = "preprocess"
     shared.state.textinfo = "Preprocessing..."
     shared.state.job_count = len(files)
 
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index fd253477..2c1251d6 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -245,6 +245,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
     create_image_every = create_image_every or 0
     validate_train_inputs(embedding_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding")
 
+    shared.state.job = "train-embedding"
     shared.state.textinfo = "Initializing textual inversion training..."
     shared.state.job_count = steps
 
-- 
cgit v1.2.3


From 184e670126f5fc50ba56fa0fedcf0cf60e45ed7e Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Wed, 4 Jan 2023 17:45:01 +0300
Subject: fix the merge

---
 modules/textual_inversion/textual_inversion.py | 14 +++++---------
 1 file changed, 5 insertions(+), 9 deletions(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index 5421a758..8731ea5d 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -251,6 +251,7 @@ def validate_train_inputs(model_name, learn_rate, batch_size, gradient_step, dat
     if save_model_every or create_image_every:
         assert log_directory, "Log directory is empty"
 
+
 def create_dummy_mask(x, width=None, height=None):
     if shared.sd_model.model.conditioning_key in {'hybrid', 'concat'}:
 
@@ -380,17 +381,12 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
                     break
 
                 with devices.autocast():
-                    # c = stack_conds(batch.cond).to(devices.device)
-                    # mask = torch.tensor(batch.emb_index).to(devices.device, non_blocking=pin_memory)
-                    # print(mask)
-                    # c[:, 1:1+embedding.vec.shape[0]] = embedding.vec.to(devices.device, non_blocking=pin_memory)
-                    
-                    
-                    if img_c is None:
-                        img_c = create_dummy_mask(c, training_width, training_height)
-                    
                     x = batch.latent_sample.to(devices.device, non_blocking=pin_memory)
                     c = shared.sd_model.cond_stage_model(batch.cond_text)
+
+                    if img_c is None:
+                        img_c = create_dummy_mask(c, training_width, training_height)
+
                     cond = {"c_concat": [img_c], "c_crossattn": [c]}
                     loss = shared.sd_model(x, cond)[0] / gradient_step
                     del x
-- 
cgit v1.2.3


From 525cea924562afd676f55470095268a0f6fca59e Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Wed, 4 Jan 2023 17:58:07 +0300
Subject: use shared function from processing for creating dummy mask when
 training inpainting model

---
 modules/processing.py                          | 39 +++++++++++++-------------
 modules/textual_inversion/textual_inversion.py | 33 ++++++----------------
 2 files changed, 29 insertions(+), 43 deletions(-)

(limited to 'modules/textual_inversion/textual_inversion.py')

diff --git a/modules/processing.py b/modules/processing.py
index c03e77e7..c7264aff 100644
--- a/modules/processing.py
+++ b/modules/processing.py
@@ -76,6 +76,24 @@ def apply_overlay(image, paste_loc, index, overlays):
     return image
 
 
+def txt2img_image_conditioning(sd_model, x, width, height):
+    if sd_model.model.conditioning_key not in {'hybrid', 'concat'}:
+        # Dummy zero conditioning if we're not using inpainting model.
+        # Still takes up a bit of memory, but no encoder call.
+        # Pretty sure we can just make this a 1x1 image since its not going to be used besides its batch size.
+        return x.new_zeros(x.shape[0], 5, 1, 1, dtype=x.dtype, device=x.device)
+
+    # The "masked-image" in this case will just be all zeros since the entire image is masked.
+    image_conditioning = torch.zeros(x.shape[0], 3, height, width, device=x.device)
+    image_conditioning = sd_model.get_first_stage_encoding(sd_model.encode_first_stage(image_conditioning))
+
+    # Add the fake full 1s mask to the first dimension.
+    image_conditioning = torch.nn.functional.pad(image_conditioning, (0, 0, 0, 0, 1, 0), value=1.0)
+    image_conditioning = image_conditioning.to(x.dtype)
+
+    return image_conditioning
+
+
 class StableDiffusionProcessing():
     """
     The first set of paramaters: sd_models -> do_not_reload_embeddings represent the minimum required to create a StableDiffusionProcessing
@@ -139,26 +157,9 @@ class StableDiffusionProcessing():
         self.iteration = 0
 
     def txt2img_image_conditioning(self, x, width=None, height=None):
-        if self.sampler.conditioning_key not in {'hybrid', 'concat'}:
-            # Dummy zero conditioning if we're not using inpainting model.
-            # Still takes up a bit of memory, but no encoder call.
-            # Pretty sure we can just make this a 1x1 image since its not going to be used besides its batch size.
-            return x.new_zeros(x.shape[0], 5, 1, 1)
+        self.is_using_inpainting_conditioning = self.sd_model.model.conditioning_key in {'hybrid', 'concat'}
 
-        self.is_using_inpainting_conditioning = True
-
-        height = height or self.height
-        width = width or self.width
-
-        # The "masked-image" in this case will just be all zeros since the entire image is masked.
-        image_conditioning = torch.zeros(x.shape[0], 3, height, width, device=x.device)
-        image_conditioning = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(image_conditioning))
-
-        # Add the fake full 1s mask to the first dimension.
-        image_conditioning = torch.nn.functional.pad(image_conditioning, (0, 0, 0, 0, 1, 0), value=1.0)
-        image_conditioning = image_conditioning.to(x.dtype)
-
-        return image_conditioning
+        return txt2img_image_conditioning(self.sd_model, x, width or self.width, height or self.height)
 
     def depth2img_image_conditioning(self, source_image):
         # Use the AddMiDaS helper to Format our source image to suit the MiDaS model
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index 8731ea5d..2250e41b 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -252,26 +252,6 @@ def validate_train_inputs(model_name, learn_rate, batch_size, gradient_step, dat
         assert log_directory, "Log directory is empty"
 
 
-def create_dummy_mask(x, width=None, height=None):
-    if shared.sd_model.model.conditioning_key in {'hybrid', 'concat'}:
-
-        # The "masked-image" in this case will just be all zeros since the entire image is masked.
-        image_conditioning = torch.zeros(x.shape[0], 3, height, width, device=x.device)
-        image_conditioning = shared.sd_model.get_first_stage_encoding(shared.sd_model.encode_first_stage(image_conditioning))
-
-        # Add the fake full 1s mask to the first dimension.
-        image_conditioning = torch.nn.functional.pad(image_conditioning, (0, 0, 0, 0, 1, 0), value=1.0)
-        image_conditioning = image_conditioning.to(x.dtype)
-
-    else:
-        # Dummy zero conditioning if we're not using inpainting model.
-        # Still takes up a bit of memory, but no encoder call.
-        # Pretty sure we can just make this a 1x1 image since its not going to be used besides its batch size.
-        image_conditioning = torch.zeros(x.shape[0], 5, 1, 1, dtype=x.dtype, device=x.device)
-
-    return image_conditioning
-
-
 def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, steps, shuffle_tags, tag_drop_out, latent_sampling_method, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
     save_embedding_every = save_embedding_every or 0
     create_image_every = create_image_every or 0
@@ -346,7 +326,6 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
         else:
             print("No saved optimizer exists in checkpoint")
 
-
     scaler = torch.cuda.amp.GradScaler()
 
     batch_size = ds.batch_size
@@ -362,7 +341,9 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
     forced_filename = "<none>"
     embedding_yet_to_be_embedded = False
 
+    is_training_inpainting_model = shared.sd_model.model.conditioning_key in {'hybrid', 'concat'}
     img_c = None
+
     pbar = tqdm.tqdm(total=steps - initial_step)
     try:
         for i in range((steps-initial_step) * gradient_step):
@@ -384,10 +365,14 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
                     x = batch.latent_sample.to(devices.device, non_blocking=pin_memory)
                     c = shared.sd_model.cond_stage_model(batch.cond_text)
 
-                    if img_c is None:
-                        img_c = create_dummy_mask(c, training_width, training_height)
+                    if is_training_inpainting_model:
+                        if img_c is None:
+                            img_c = processing.txt2img_image_conditioning(shared.sd_model, c, training_width, training_height)
+
+                        cond = {"c_concat": [img_c], "c_crossattn": [c]}
+                    else:
+                        cond = c
 
-                    cond = {"c_concat": [img_c], "c_crossattn": [c]}
                     loss = shared.sd_model(x, cond)[0] / gradient_step
                     del x
 
-- 
cgit v1.2.3