From 820f1dc96b1979d7e92170c161db281ee8bd988b Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sun, 2 Oct 2022 15:03:39 +0300 Subject: initial support for training textual inversion --- modules/textual_inversion/textual_inversion.py | 258 +++++++++++++++++++++++++ 1 file changed, 258 insertions(+) create mode 100644 modules/textual_inversion/textual_inversion.py (limited to 'modules/textual_inversion/textual_inversion.py') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py new file mode 100644 index 00000000..c0baaace --- /dev/null +++ b/modules/textual_inversion/textual_inversion.py @@ -0,0 +1,258 @@ +import os +import sys +import traceback + +import torch +import tqdm +import html +import datetime + +from modules import shared, devices, sd_hijack, processing +import modules.textual_inversion.dataset + + +class Embedding: + def __init__(self, vec, name, step=None): + self.vec = vec + self.name = name + self.step = step + self.cached_checksum = None + + def save(self, filename): + embedding_data = { + "string_to_token": {"*": 265}, + "string_to_param": {"*": self.vec}, + "name": self.name, + "step": self.step, + } + + torch.save(embedding_data, filename) + + def checksum(self): + if self.cached_checksum is not None: + return self.cached_checksum + + def const_hash(a): + r = 0 + for v in a: + r = (r * 281 ^ int(v) * 997) & 0xFFFFFFFF + return r + + self.cached_checksum = f'{const_hash(self.vec.reshape(-1) * 100) & 0xffff:04x}' + return self.cached_checksum + +class EmbeddingDatabase: + def __init__(self, embeddings_dir): + self.ids_lookup = {} + self.word_embeddings = {} + self.dir_mtime = None + self.embeddings_dir = embeddings_dir + + def register_embedding(self, embedding, model): + + self.word_embeddings[embedding.name] = embedding + + ids = model.cond_stage_model.tokenizer([embedding.name], add_special_tokens=False)['input_ids'][0] + + first_id = ids[0] + if first_id not in self.ids_lookup: + self.ids_lookup[first_id] = [] + self.ids_lookup[first_id].append((ids, embedding)) + + return embedding + + def load_textual_inversion_embeddings(self): + mt = os.path.getmtime(self.embeddings_dir) + if self.dir_mtime is not None and mt <= self.dir_mtime: + return + + self.dir_mtime = mt + self.ids_lookup.clear() + self.word_embeddings.clear() + + def process_file(path, filename): + name = os.path.splitext(filename)[0] + + data = torch.load(path, map_location="cpu") + + # textual inversion embeddings + if 'string_to_param' in data: + param_dict = data['string_to_param'] + if hasattr(param_dict, '_parameters'): + param_dict = getattr(param_dict, '_parameters') # fix for torch 1.12.1 loading saved file from torch 1.11 + assert len(param_dict) == 1, 'embedding file has multiple terms in it' + emb = next(iter(param_dict.items()))[1] + # diffuser concepts + elif type(data) == dict and type(next(iter(data.values()))) == torch.Tensor: + assert len(data.keys()) == 1, 'embedding file has multiple terms in it' + + emb = next(iter(data.values())) + if len(emb.shape) == 1: + emb = emb.unsqueeze(0) + else: + raise Exception(f"Couldn't identify {filename} as neither textual inversion embedding nor diffuser concept.") + + vec = emb.detach().to(devices.device, dtype=torch.float32) + embedding = Embedding(vec, name) + embedding.step = data.get('step', None) + self.register_embedding(embedding, shared.sd_model) + + for fn in os.listdir(self.embeddings_dir): + try: + fullfn = os.path.join(self.embeddings_dir, fn) + + if os.stat(fullfn).st_size == 0: + continue + + process_file(fullfn, fn) + except Exception: + print(f"Error loading emedding {fn}:", file=sys.stderr) + print(traceback.format_exc(), file=sys.stderr) + continue + + print(f"Loaded a total of {len(self.word_embeddings)} textual inversion embeddings.") + + def find_embedding_at_position(self, tokens, offset): + token = tokens[offset] + possible_matches = self.ids_lookup.get(token, None) + + if possible_matches is None: + return None + + for ids, embedding in possible_matches: + if tokens[offset:offset + len(ids)] == ids: + return embedding + + return None + + + +def create_embedding(name, num_vectors_per_token): + init_text = '*' + + cond_model = shared.sd_model.cond_stage_model + embedding_layer = cond_model.wrapped.transformer.text_model.embeddings + + ids = cond_model.tokenizer(init_text, max_length=num_vectors_per_token, return_tensors="pt", add_special_tokens=False)["input_ids"] + embedded = embedding_layer(ids.to(devices.device)).squeeze(0) + vec = torch.zeros((num_vectors_per_token, embedded.shape[1]), device=devices.device) + + for i in range(num_vectors_per_token): + vec[i] = embedded[i * int(embedded.shape[0]) // num_vectors_per_token] + + fn = os.path.join(shared.cmd_opts.embeddings_dir, f"{name}.pt") + assert not os.path.exists(fn), f"file {fn} already exists" + + embedding = Embedding(vec, name) + embedding.step = 0 + embedding.save(fn) + + return fn + + +def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, create_image_every, save_embedding_every, template_file): + assert embedding_name, 'embedding not selected' + + shared.state.textinfo = "Initializing textual inversion training..." + shared.state.job_count = steps + + filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt') + + log_directory = os.path.join(log_directory, datetime.datetime.now().strftime("%Y-%d-%m"), embedding_name) + + if save_embedding_every > 0: + embedding_dir = os.path.join(log_directory, "embeddings") + os.makedirs(embedding_dir, exist_ok=True) + else: + embedding_dir = None + + if create_image_every > 0: + images_dir = os.path.join(log_directory, "images") + os.makedirs(images_dir, exist_ok=True) + else: + images_dir = None + + cond_model = shared.sd_model.cond_stage_model + + 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, size=512, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file) + + hijack = sd_hijack.model_hijack + + embedding = hijack.embedding_db.word_embeddings[embedding_name] + embedding.vec.requires_grad = True + + optimizer = torch.optim.AdamW([embedding.vec], lr=learn_rate) + + losses = torch.zeros((32,)) + + last_saved_file = "" + last_saved_image = "" + + ititial_step = embedding.step or 0 + if ititial_step > steps: + return embedding, filename + + pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) + for i, (x, text) in pbar: + embedding.step = i + ititial_step + + if embedding.step > steps: + break + + if shared.state.interrupted: + break + + with torch.autocast("cuda"): + c = cond_model([text]) + loss = shared.sd_model(x.unsqueeze(0), c)[0] + + losses[embedding.step % losses.shape[0]] = loss.item() + + optimizer.zero_grad() + loss.backward() + optimizer.step() + + pbar.set_description(f"loss: {losses.mean():.7f}") + + if embedding.step > 0 and embedding_dir is not None and embedding.step % save_embedding_every == 0: + last_saved_file = os.path.join(embedding_dir, f'{embedding_name}-{embedding.step}.pt') + embedding.save(last_saved_file) + + if embedding.step > 0 and images_dir is not None and embedding.step % create_image_every == 0: + last_saved_image = os.path.join(images_dir, f'{embedding_name}-{embedding.step}.png') + + p = processing.StableDiffusionProcessingTxt2Img( + sd_model=shared.sd_model, + prompt=text, + steps=20, + do_not_save_grid=True, + do_not_save_samples=True, + ) + + processed = processing.process_images(p) + image = processed.images[0] + + shared.state.current_image = image + image.save(last_saved_image) + + last_saved_image += f", prompt: {text}" + + shared.state.job_no = embedding.step + + shared.state.textinfo = f""" +

+Loss: {losses.mean():.7f}
+Step: {embedding.step}
+Last prompt: {html.escape(text)}
+Last saved embedding: {html.escape(last_saved_file)}
+Last saved image: {html.escape(last_saved_image)}
+

+""" + + embedding.cached_checksum = None + embedding.save(filename) + + return embedding, filename + -- cgit v1.2.3 From 88ec0cf5571883d84abd09196652b3679e359f2e Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sun, 2 Oct 2022 19:40:51 +0300 Subject: fix for incorrect embedding token length calculation (will break seeds that use embeddings, you're welcome!) add option to input initialization text for embeddings --- modules/textual_inversion/textual_inversion.py | 13 +++++-------- 1 file changed, 5 insertions(+), 8 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 c0baaace..0c50161d 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -117,24 +117,21 @@ class EmbeddingDatabase: possible_matches = self.ids_lookup.get(token, None) if possible_matches is None: - return None + return None, None for ids, embedding in possible_matches: if tokens[offset:offset + len(ids)] == ids: - return embedding + return embedding, len(ids) - return None + return None, None - -def create_embedding(name, num_vectors_per_token): - init_text = '*' - +def create_embedding(name, num_vectors_per_token, init_text='*'): cond_model = shared.sd_model.cond_stage_model embedding_layer = cond_model.wrapped.transformer.text_model.embeddings ids = cond_model.tokenizer(init_text, max_length=num_vectors_per_token, return_tensors="pt", add_special_tokens=False)["input_ids"] - embedded = embedding_layer(ids.to(devices.device)).squeeze(0) + embedded = embedding_layer.token_embedding.wrapped(ids.to(devices.device)).squeeze(0) vec = torch.zeros((num_vectors_per_token, embedded.shape[1]), device=devices.device) for i in range(num_vectors_per_token): -- cgit v1.2.3 From 71fe7fa49f5eb1a2c89932a9d217ed153c12fc8b Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sun, 2 Oct 2022 19:56:37 +0300 Subject: fix using aaaa-100 embedding when the prompt has aaaa-10000 and you have both aaaa-100 and aaaa-10000 in the directory with embeddings. --- modules/textual_inversion/textual_inversion.py | 3 ++- 1 file changed, 2 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 0c50161d..9d2241ce 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -57,7 +57,8 @@ class EmbeddingDatabase: first_id = ids[0] if first_id not in self.ids_lookup: self.ids_lookup[first_id] = [] - self.ids_lookup[first_id].append((ids, embedding)) + + self.ids_lookup[first_id] = sorted(self.ids_lookup[first_id] + [(ids, embedding)], key=lambda x: len(x[0]), reverse=True) return embedding -- cgit v1.2.3 From 4ec4af6e0b7addeee5221a03f32d117ccdc875d9 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sun, 2 Oct 2022 20:15:25 +0300 Subject: add checkpoint info to saved embeddings --- modules/textual_inversion/textual_inversion.py | 13 ++++++++++++- 1 file changed, 12 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 9d2241ce..1183aab7 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -7,7 +7,7 @@ import tqdm import html import datetime -from modules import shared, devices, sd_hijack, processing +from modules import shared, devices, sd_hijack, processing, sd_models import modules.textual_inversion.dataset @@ -17,6 +17,8 @@ class Embedding: self.name = name self.step = step self.cached_checksum = None + self.sd_checkpoint = None + self.sd_checkpoint_name = None def save(self, filename): embedding_data = { @@ -24,6 +26,8 @@ class Embedding: "string_to_param": {"*": self.vec}, "name": self.name, "step": self.step, + "sd_checkpoint": self.sd_checkpoint, + "sd_checkpoint_name": self.sd_checkpoint_name, } torch.save(embedding_data, filename) @@ -41,6 +45,7 @@ class Embedding: self.cached_checksum = f'{const_hash(self.vec.reshape(-1) * 100) & 0xffff:04x}' return self.cached_checksum + class EmbeddingDatabase: def __init__(self, embeddings_dir): self.ids_lookup = {} @@ -96,6 +101,8 @@ class EmbeddingDatabase: vec = emb.detach().to(devices.device, dtype=torch.float32) embedding = Embedding(vec, name) embedding.step = data.get('step', None) + embedding.sd_checkpoint = data.get('hash', None) + embedding.sd_checkpoint_name = data.get('sd_checkpoint_name', None) self.register_embedding(embedding, shared.sd_model) for fn in os.listdir(self.embeddings_dir): @@ -249,6 +256,10 @@ Last saved image: {html.escape(last_saved_image)}

""" + checkpoint = sd_models.select_checkpoint() + + embedding.sd_checkpoint = checkpoint.hash + embedding.sd_checkpoint_name = checkpoint.model_name embedding.cached_checksum = None embedding.save(filename) -- cgit v1.2.3 From c7543d4940da672d970124ae8f2fec9de7bdc1da Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sun, 2 Oct 2022 22:41:21 +0300 Subject: preprocessing for textual inversion added --- modules/textual_inversion/textual_inversion.py | 1 + 1 file changed, 1 insertion(+) (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 1183aab7..d4e250d8 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -7,6 +7,7 @@ import tqdm import html import datetime + from modules import shared, devices, sd_hijack, processing, sd_models import modules.textual_inversion.dataset -- cgit v1.2.3 From 6785331e22d6a488fbf5905fab56d7fec867e038 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sun, 2 Oct 2022 22:59:01 +0300 Subject: keep textual inversion dataset latents in CPU memory to save a bit of VRAM --- modules/textual_inversion/textual_inversion.py | 3 +++ 1 file changed, 3 insertions(+) (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 d4e250d8..8686f534 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -212,7 +212,10 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, with torch.autocast("cuda"): c = cond_model([text]) + + x = x.to(devices.device) loss = shared.sd_model(x.unsqueeze(0), c)[0] + del x losses[embedding.step % losses.shape[0]] = loss.item() -- cgit v1.2.3 From 2865ef4b9ab16d56326cc805541bebcf01d099bc Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Mon, 3 Oct 2022 13:10:03 +0300 Subject: fix broken date in TI --- 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 8686f534..cd9f3498 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -164,7 +164,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt') - log_directory = os.path.join(log_directory, datetime.datetime.now().strftime("%Y-%d-%m"), embedding_name) + log_directory = os.path.join(log_directory, datetime.datetime.now().strftime("%Y-%m-%d"), embedding_name) if save_embedding_every > 0: embedding_dir = os.path.join(log_directory, "embeddings") -- cgit v1.2.3 From 5841990b0df04906da7321beef6f7f7902b7d57b Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Sun, 9 Oct 2022 05:38:38 +0100 Subject: Update textual_inversion.py --- modules/textual_inversion/textual_inversion.py | 25 ++++++++++++++++++++++--- 1 file changed, 22 insertions(+), 3 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 cd9f3498..f6316020 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -7,6 +7,9 @@ import tqdm import html import datetime +from PIL import Image, PngImagePlugin +import base64 +from io import BytesIO from modules import shared, devices, sd_hijack, processing, sd_models import modules.textual_inversion.dataset @@ -80,7 +83,15 @@ class EmbeddingDatabase: def process_file(path, filename): name = os.path.splitext(filename)[0] - data = torch.load(path, map_location="cpu") + data = [] + + if filename.upper().endswith('.PNG'): + embed_image = Image.open(path) + if 'sd-embedding' in embed_image.text: + embeddingData = base64.b64decode(embed_image.text['sd-embedding']) + data = torch.load(BytesIO(embeddingData), map_location="cpu") + else: + data = torch.load(path, map_location="cpu") # textual inversion embeddings if 'string_to_param' in data: @@ -156,7 +167,7 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): return fn -def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, create_image_every, save_embedding_every, template_file): +def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding): assert embedding_name, 'embedding not selected' shared.state.textinfo = "Initializing textual inversion training..." @@ -244,7 +255,15 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, image = processed.images[0] shared.state.current_image = image - image.save(last_saved_image) + + if save_image_with_stored_embedding: + info = PngImagePlugin.PngInfo() + info.add_text("sd-embedding", base64.b64encode(open(last_saved_file,'rb').read())) + image.save(last_saved_image, "PNG", pnginfo=info) + else: + image.save(last_saved_image) + + last_saved_image += f", prompt: {text}" -- cgit v1.2.3 From 03694e1f9915e34cf7d9a31073f1a1a9def2909f Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Sun, 9 Oct 2022 21:58:14 +0100 Subject: add embedding load and save from b64 json --- modules/textual_inversion/textual_inversion.py | 30 ++++++++++++++++++-------- 1 file changed, 21 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 f6316020..1b7f8906 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -7,9 +7,11 @@ import tqdm import html import datetime -from PIL import Image, PngImagePlugin +from PIL import Image,PngImagePlugin +from ..images import captionImge +import numpy as np import base64 -from io import BytesIO +import json from modules import shared, devices, sd_hijack, processing, sd_models import modules.textual_inversion.dataset @@ -87,9 +89,9 @@ class EmbeddingDatabase: if filename.upper().endswith('.PNG'): embed_image = Image.open(path) - if 'sd-embedding' in embed_image.text: - embeddingData = base64.b64decode(embed_image.text['sd-embedding']) - data = torch.load(BytesIO(embeddingData), map_location="cpu") + if 'sd-ti-embedding' in embed_image.text: + data = embeddingFromB64(embed_image.text['sd-ti-embedding']) + name = data.get('name',name) else: data = torch.load(path, map_location="cpu") @@ -258,13 +260,23 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, if save_image_with_stored_embedding: info = PngImagePlugin.PngInfo() - info.add_text("sd-embedding", base64.b64encode(open(last_saved_file,'rb').read())) - image.save(last_saved_image, "PNG", pnginfo=info) + data = torch.load(last_saved_file) + info.add_text("sd-ti-embedding", embeddingToB64(data)) + + pre_lines = [((255, 207, 175),"<{}>".format(data.get('name','???')))] + + caption_checkpoint_hash = data.get('sd_checkpoint','UNK') + caption_checkpoint_hash = caption_checkpoint_hash.upper() if caption_checkpoint_hash else 'UNK' + caption_stepcount = data.get('step',0) + caption_stepcount = caption_stepcount if caption_stepcount else 0 + + post_lines = [((240, 223, 175),"Trained against checkpoint [{}] for {} steps".format(caption_checkpoint_hash, + caption_stepcount))] + captioned_image = captionImge(image,prelines=pre_lines,postlines=post_lines) + captioned_image.save(last_saved_image, "PNG", pnginfo=info) else: image.save(last_saved_image) - - last_saved_image += f", prompt: {text}" shared.state.job_no = embedding.step -- cgit v1.2.3 From 969bd8256e5b4f1007d3cc653723d4ad50a92528 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Sun, 9 Oct 2022 22:02:28 +0100 Subject: add alternate checkpoint hash source --- modules/textual_inversion/textual_inversion.py | 7 +++++-- 1 file changed, 5 insertions(+), 2 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 1b7f8906..d7813084 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -265,8 +265,11 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, pre_lines = [((255, 207, 175),"<{}>".format(data.get('name','???')))] - caption_checkpoint_hash = data.get('sd_checkpoint','UNK') - caption_checkpoint_hash = caption_checkpoint_hash.upper() if caption_checkpoint_hash else 'UNK' + caption_checkpoint_hash = data.get('sd_checkpoint') + if caption_checkpoint_hash is None: + caption_checkpoint_hash = data.get('hash') + caption_checkpoint_hash = caption_checkpoint_hash.upper() if caption_checkpoint_hash else 'UNKNOWN' + caption_stepcount = data.get('step',0) caption_stepcount = caption_stepcount if caption_stepcount else 0 -- cgit v1.2.3 From 5d12ec82d3e13f5ff4c55db2930e4e10aed7015a Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Sun, 9 Oct 2022 22:05:09 +0100 Subject: add encoder and decoder classes --- modules/textual_inversion/textual_inversion.py | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+) (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 d7813084..44d4e08b 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -16,6 +16,27 @@ import json from modules import shared, devices, sd_hijack, processing, sd_models import modules.textual_inversion.dataset +class EmbeddingEncoder(json.JSONEncoder): + def default(self, obj): + if isinstance(obj, torch.Tensor): + return {'EMBEDDINGTENSOR':obj.cpu().detach().numpy().tolist()} + return json.JSONEncoder.default(self, o) + +class EmbeddingDecoder(json.JSONDecoder): + def __init__(self, *args, **kwargs): + json.JSONDecoder.__init__(self, object_hook=self.object_hook, *args, **kwargs) + def object_hook(self, d): + if 'EMBEDDINGTENSOR' in d: + return torch.from_numpy(np.array(d['EMBEDDINGTENSOR'])) + return d + +def embeddingToB64(data): + d = json.dumps(data,cls=EmbeddingEncoder) + return base64.b64encode(d.encode()) + +def EmbeddingFromB64(data): + d = base64.b64decode(data) + return json.loads(d,cls=EmbeddingDecoder) class Embedding: def __init__(self, vec, name, step=None): -- cgit v1.2.3 From d0184b8f76ce492da699f1926f34b57cd095242e Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Sun, 9 Oct 2022 22:06:12 +0100 Subject: change json tensor key name --- modules/textual_inversion/textual_inversion.py | 6 +++--- 1 file changed, 3 insertions(+), 3 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 44d4e08b..ae8d207d 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -19,15 +19,15 @@ import modules.textual_inversion.dataset class EmbeddingEncoder(json.JSONEncoder): def default(self, obj): if isinstance(obj, torch.Tensor): - return {'EMBEDDINGTENSOR':obj.cpu().detach().numpy().tolist()} + return {'TORCHTENSOR':obj.cpu().detach().numpy().tolist()} return json.JSONEncoder.default(self, o) class EmbeddingDecoder(json.JSONDecoder): def __init__(self, *args, **kwargs): json.JSONDecoder.__init__(self, object_hook=self.object_hook, *args, **kwargs) def object_hook(self, d): - if 'EMBEDDINGTENSOR' in d: - return torch.from_numpy(np.array(d['EMBEDDINGTENSOR'])) + if 'TORCHTENSOR' in d: + return torch.from_numpy(np.array(d['TORCHTENSOR'])) return d def embeddingToB64(data): -- cgit v1.2.3 From 66846105103cfc282434d0dc2102910160b7a633 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Sun, 9 Oct 2022 22:06:42 +0100 Subject: correct case on embeddingFromB64 --- 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 ae8d207d..d2b95fa3 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -34,7 +34,7 @@ def embeddingToB64(data): d = json.dumps(data,cls=EmbeddingEncoder) return base64.b64encode(d.encode()) -def EmbeddingFromB64(data): +def embeddingFromB64(data): d = base64.b64decode(data) return json.loads(d,cls=EmbeddingDecoder) -- cgit v1.2.3 From 96f1e6be59316ec640cab2435fa95b3688194906 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Sun, 9 Oct 2022 22:14:50 +0100 Subject: source checkpoint hash from current checkpoint --- modules/textual_inversion/textual_inversion.py | 6 ++---- 1 file changed, 2 insertions(+), 4 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 d2b95fa3..b16fa84e 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -286,10 +286,8 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, pre_lines = [((255, 207, 175),"<{}>".format(data.get('name','???')))] - caption_checkpoint_hash = data.get('sd_checkpoint') - if caption_checkpoint_hash is None: - caption_checkpoint_hash = data.get('hash') - caption_checkpoint_hash = caption_checkpoint_hash.upper() if caption_checkpoint_hash else 'UNKNOWN' + checkpoint = sd_models.select_checkpoint() + caption_checkpoint_hash = checkpoint.hash caption_stepcount = data.get('step',0) caption_stepcount = caption_stepcount if caption_stepcount else 0 -- cgit v1.2.3 From 01fd9cf0d28d8b71a113ab1aa62accfe7f0d9c51 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Sun, 9 Oct 2022 22:17:02 +0100 Subject: change source of step count --- modules/textual_inversion/textual_inversion.py | 10 ++-------- 1 file changed, 2 insertions(+), 8 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 b16fa84e..e4f339b8 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -285,15 +285,9 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, info.add_text("sd-ti-embedding", embeddingToB64(data)) pre_lines = [((255, 207, 175),"<{}>".format(data.get('name','???')))] - checkpoint = sd_models.select_checkpoint() - caption_checkpoint_hash = checkpoint.hash - - caption_stepcount = data.get('step',0) - caption_stepcount = caption_stepcount if caption_stepcount else 0 - - post_lines = [((240, 223, 175),"Trained against checkpoint [{}] for {} steps".format(caption_checkpoint_hash, - caption_stepcount))] + post_lines = [((240, 223, 175),"Trained against checkpoint [{}] for {} steps".format(checkpoint.hash, + embedding.step))] captioned_image = captionImge(image,prelines=pre_lines,postlines=post_lines) captioned_image.save(last_saved_image, "PNG", pnginfo=info) else: -- cgit v1.2.3 From d6a599ef9ba18a66ae79b50f2945af5788fdda8f Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Mon, 10 Oct 2022 00:07:52 +0100 Subject: change caption method --- modules/textual_inversion/textual_inversion.py | 30 ++++++++++++++++++-------- 1 file changed, 21 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 e4f339b8..21596e78 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -8,7 +8,7 @@ import html import datetime from PIL import Image,PngImagePlugin -from ..images import captionImge +from ..images import captionImageOverlay import numpy as np import base64 import json @@ -212,6 +212,12 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, else: images_dir = None + if create_image_every > 0 and save_image_with_stored_embedding: + images_embeds_dir = os.path.join(log_directory, "image_embeddings") + os.makedirs(images_embeds_dir, exist_ok=True) + else: + images_embeds_dir = None + cond_model = shared.sd_model.cond_stage_model shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..." @@ -279,19 +285,25 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, shared.state.current_image = image - if save_image_with_stored_embedding: + if save_image_with_stored_embedding and os.path.exists(last_saved_file): + + last_saved_image_chunks = os.path.join(images_embeds_dir, f'{embedding_name}-{embedding.step}.png') + info = PngImagePlugin.PngInfo() data = torch.load(last_saved_file) info.add_text("sd-ti-embedding", embeddingToB64(data)) - pre_lines = [((255, 207, 175),"<{}>".format(data.get('name','???')))] + title = "<{}>".format(data.get('name','???')) checkpoint = sd_models.select_checkpoint() - post_lines = [((240, 223, 175),"Trained against checkpoint [{}] for {} steps".format(checkpoint.hash, - embedding.step))] - captioned_image = captionImge(image,prelines=pre_lines,postlines=post_lines) - captioned_image.save(last_saved_image, "PNG", pnginfo=info) - else: - image.save(last_saved_image) + footer_left = checkpoint.model_name + footer_mid = '[{}]'.format(checkpoint.hash) + footer_right = '[{}]'.format(embedding.step) + + captioned_image = captionImageOverlay(image,title,footer_left,footer_mid,footer_right) + + captioned_image.save(last_saved_image_chunks, "PNG", pnginfo=info) + + image.save(last_saved_image) last_saved_image += f", prompt: {text}" -- cgit v1.2.3 From e2c2925eb4d634b186de2c76798162ec56e2f869 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Mon, 10 Oct 2022 00:12:53 +0100 Subject: remove braces from steps --- 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 21596e78..9a18ee5c 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -297,7 +297,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, checkpoint = sd_models.select_checkpoint() footer_left = checkpoint.model_name footer_mid = '[{}]'.format(checkpoint.hash) - footer_right = '[{}]'.format(embedding.step) + footer_right = '{}'.format(embedding.step) captioned_image = captionImageOverlay(image,title,footer_left,footer_mid,footer_right) -- cgit v1.2.3 From 3110f895b2718a3a25aae419fdf5c87c177ec9f4 Mon Sep 17 00:00:00 2001 From: alg-wiki Date: Mon, 10 Oct 2022 17:07:46 +0900 Subject: Textual Inversion: Added custom training image size and number of repeats per input image in a single epoch --- modules/textual_inversion/textual_inversion.py | 15 ++++++++++++--- 1 file changed, 12 insertions(+), 3 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 cd9f3498..e34dc2e8 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -6,6 +6,7 @@ import torch import tqdm import html import datetime +import math from modules import shared, devices, sd_hijack, processing, sd_models @@ -156,7 +157,7 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): return fn -def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, create_image_every, save_embedding_every, template_file): +def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_size, steps, num_repeats, create_image_every, save_embedding_every, template_file): assert embedding_name, 'embedding not selected' shared.state.textinfo = "Initializing textual inversion training..." @@ -182,7 +183,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, 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, size=512, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file) + ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, size=training_size, repeats=num_repeats, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file) hijack = sd_hijack.model_hijack @@ -200,6 +201,9 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, if ititial_step > steps: return embedding, filename + tr_img_len = len([os.path.join(data_root, file_path) for file_path in os.listdir(data_root)]) + epoch_len = (tr_img_len * num_repeats) + tr_img_len + pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) for i, (x, text) in pbar: embedding.step = i + ititial_step @@ -223,7 +227,10 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, loss.backward() optimizer.step() - pbar.set_description(f"loss: {losses.mean():.7f}") + epoch_num = math.floor(embedding.step / epoch_len) + epoch_step = embedding.step - (epoch_num * epoch_len) + + pbar.set_description(f"[Epoch {epoch_num}: {epoch_step}/{epoch_len}]loss: {losses.mean():.7f}") if embedding.step > 0 and embedding_dir is not None and embedding.step % save_embedding_every == 0: last_saved_file = os.path.join(embedding_dir, f'{embedding_name}-{embedding.step}.pt') @@ -236,6 +243,8 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, sd_model=shared.sd_model, prompt=text, steps=20, + height=training_size, + width=training_size, do_not_save_grid=True, do_not_save_samples=True, ) -- cgit v1.2.3 From 4ee7519fc2e459ce8eff1f61f1655afba393357c Mon Sep 17 00:00:00 2001 From: alg-wiki Date: Mon, 10 Oct 2022 17:31:33 +0900 Subject: Fixed progress bar output for epoch --- 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 e34dc2e8..769682ea 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -228,7 +228,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini optimizer.step() epoch_num = math.floor(embedding.step / epoch_len) - epoch_step = embedding.step - (epoch_num * epoch_len) + epoch_step = embedding.step - (epoch_num * epoch_len) + 1 pbar.set_description(f"[Epoch {epoch_num}: {epoch_step}/{epoch_len}]loss: {losses.mean():.7f}") -- cgit v1.2.3 From 04c745ea4f81518999927fee5f78500560c25e29 Mon Sep 17 00:00:00 2001 From: alg-wiki Date: Mon, 10 Oct 2022 22:35:35 +0900 Subject: Custom Width and Height --- modules/textual_inversion/textual_inversion.py | 11 +++++------ 1 file changed, 5 insertions(+), 6 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 769682ea..5965c5a0 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -6,7 +6,6 @@ import torch import tqdm import html import datetime -import math from modules import shared, devices, sd_hijack, processing, sd_models @@ -157,7 +156,7 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): return fn -def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_size, steps, num_repeats, create_image_every, save_embedding_every, template_file): +def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, num_repeats, create_image_every, save_embedding_every, template_file): assert embedding_name, 'embedding not selected' shared.state.textinfo = "Initializing textual inversion training..." @@ -183,7 +182,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini 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, size=training_size, repeats=num_repeats, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file) + ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=num_repeats, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file) hijack = sd_hijack.model_hijack @@ -227,7 +226,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini loss.backward() optimizer.step() - epoch_num = math.floor(embedding.step / epoch_len) + epoch_num = embedding.step // epoch_len epoch_step = embedding.step - (epoch_num * epoch_len) + 1 pbar.set_description(f"[Epoch {epoch_num}: {epoch_step}/{epoch_len}]loss: {losses.mean():.7f}") @@ -243,8 +242,8 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini sd_model=shared.sd_model, prompt=text, steps=20, - height=training_size, - width=training_size, + height=training_height, + width=training_width, do_not_save_grid=True, do_not_save_samples=True, ) -- cgit v1.2.3 From ea00c1624bbb0dcb5be07f59c9509061baddf5b1 Mon Sep 17 00:00:00 2001 From: alg-wiki Date: Mon, 10 Oct 2022 17:07:46 +0900 Subject: Textual Inversion: Added custom training image size and number of repeats per input image in a single epoch --- modules/textual_inversion/textual_inversion.py | 15 ++++++++++++--- 1 file changed, 12 insertions(+), 3 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 cd9f3498..e34dc2e8 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -6,6 +6,7 @@ import torch import tqdm import html import datetime +import math from modules import shared, devices, sd_hijack, processing, sd_models @@ -156,7 +157,7 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): return fn -def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, create_image_every, save_embedding_every, template_file): +def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_size, steps, num_repeats, create_image_every, save_embedding_every, template_file): assert embedding_name, 'embedding not selected' shared.state.textinfo = "Initializing textual inversion training..." @@ -182,7 +183,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, 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, size=512, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file) + ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, size=training_size, repeats=num_repeats, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file) hijack = sd_hijack.model_hijack @@ -200,6 +201,9 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, if ititial_step > steps: return embedding, filename + tr_img_len = len([os.path.join(data_root, file_path) for file_path in os.listdir(data_root)]) + epoch_len = (tr_img_len * num_repeats) + tr_img_len + pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) for i, (x, text) in pbar: embedding.step = i + ititial_step @@ -223,7 +227,10 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, loss.backward() optimizer.step() - pbar.set_description(f"loss: {losses.mean():.7f}") + epoch_num = math.floor(embedding.step / epoch_len) + epoch_step = embedding.step - (epoch_num * epoch_len) + + pbar.set_description(f"[Epoch {epoch_num}: {epoch_step}/{epoch_len}]loss: {losses.mean():.7f}") if embedding.step > 0 and embedding_dir is not None and embedding.step % save_embedding_every == 0: last_saved_file = os.path.join(embedding_dir, f'{embedding_name}-{embedding.step}.pt') @@ -236,6 +243,8 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, sd_model=shared.sd_model, prompt=text, steps=20, + height=training_size, + width=training_size, do_not_save_grid=True, do_not_save_samples=True, ) -- cgit v1.2.3 From 6ad3a53e368d36535de1a4fca73b3bb78fd40654 Mon Sep 17 00:00:00 2001 From: alg-wiki Date: Mon, 10 Oct 2022 17:31:33 +0900 Subject: Fixed progress bar output for epoch --- 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 e34dc2e8..769682ea 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -228,7 +228,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini optimizer.step() epoch_num = math.floor(embedding.step / epoch_len) - epoch_step = embedding.step - (epoch_num * epoch_len) + epoch_step = embedding.step - (epoch_num * epoch_len) + 1 pbar.set_description(f"[Epoch {epoch_num}: {epoch_step}/{epoch_len}]loss: {losses.mean():.7f}") -- cgit v1.2.3 From 7a20f914eddfdf09c0ccced157ec108205bc3d0f Mon Sep 17 00:00:00 2001 From: alg-wiki Date: Mon, 10 Oct 2022 22:35:35 +0900 Subject: Custom Width and Height --- modules/textual_inversion/textual_inversion.py | 11 +++++------ 1 file changed, 5 insertions(+), 6 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 769682ea..5965c5a0 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -6,7 +6,6 @@ import torch import tqdm import html import datetime -import math from modules import shared, devices, sd_hijack, processing, sd_models @@ -157,7 +156,7 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): return fn -def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_size, steps, num_repeats, create_image_every, save_embedding_every, template_file): +def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, num_repeats, create_image_every, save_embedding_every, template_file): assert embedding_name, 'embedding not selected' shared.state.textinfo = "Initializing textual inversion training..." @@ -183,7 +182,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini 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, size=training_size, repeats=num_repeats, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file) + ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=num_repeats, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file) hijack = sd_hijack.model_hijack @@ -227,7 +226,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini loss.backward() optimizer.step() - epoch_num = math.floor(embedding.step / epoch_len) + epoch_num = embedding.step // epoch_len epoch_step = embedding.step - (epoch_num * epoch_len) + 1 pbar.set_description(f"[Epoch {epoch_num}: {epoch_step}/{epoch_len}]loss: {losses.mean():.7f}") @@ -243,8 +242,8 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini sd_model=shared.sd_model, prompt=text, steps=20, - height=training_size, - width=training_size, + height=training_height, + width=training_width, do_not_save_grid=True, do_not_save_samples=True, ) -- cgit v1.2.3 From 707a431100362645e914042bb344d08439f48ac8 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Mon, 10 Oct 2022 15:34:49 +0100 Subject: add pixel data footer --- modules/textual_inversion/textual_inversion.py | 48 ++++++++++++++++++++++++-- 1 file changed, 46 insertions(+), 2 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 7a24192e..6fb64691 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -12,6 +12,7 @@ from ..images import captionImageOverlay import numpy as np import base64 import json +import zlib from modules import shared, devices, sd_hijack, processing, sd_models import modules.textual_inversion.dataset @@ -20,7 +21,7 @@ class EmbeddingEncoder(json.JSONEncoder): def default(self, obj): if isinstance(obj, torch.Tensor): return {'TORCHTENSOR':obj.cpu().detach().numpy().tolist()} - return json.JSONEncoder.default(self, o) + return json.JSONEncoder.default(self, obj) class EmbeddingDecoder(json.JSONDecoder): def __init__(self, *args, **kwargs): @@ -38,6 +39,45 @@ def embeddingFromB64(data): d = base64.b64decode(data) return json.loads(d,cls=EmbeddingDecoder) +def appendImageDataFooter(image,data): + d = 3 + data_compressed = zlib.compress( json.dumps(data,cls=EmbeddingEncoder).encode(),level=9) + dnp = np.frombuffer(data_compressed,np.uint8).copy() + w = image.size[0] + next_size = dnp.shape[0] + (w-(dnp.shape[0]%w)) + next_size = next_size + ((w*d)-(next_size%(w*d))) + dnp.resize(next_size) + dnp = dnp.reshape((-1,w,d)) + print(dnp.shape) + im = Image.fromarray(dnp,mode='RGB') + background = Image.new('RGB',(image.size[0],image.size[1]+im.size[1]+1),(0,0,0)) + background.paste(image,(0,0)) + background.paste(im,(0,image.size[1]+1)) + return background + +def crop_black(img,tol=0): + mask = (img>tol).all(2) + mask0,mask1 = mask.any(0),mask.any(1) + col_start,col_end = mask0.argmax(),mask.shape[1]-mask0[::-1].argmax() + row_start,row_end = mask1.argmax(),mask.shape[0]-mask1[::-1].argmax() + return img[row_start:row_end,col_start:col_end] + +def extractImageDataFooter(image): + d=3 + outarr = crop_black(np.array(image.getdata()).reshape(image.size[1],image.size[0],d ).astype(np.uint8) ) + lastRow = np.where( np.sum(outarr, axis=(1,2))==0) + if lastRow[0].shape[0] == 0: + print('Image data block not found.') + return None + lastRow = lastRow[0] + + lastRow = lastRow.max() + + dataBlock = outarr[lastRow+1::].astype(np.uint8).flatten().tobytes() + print(lastRow) + data = zlib.decompress(dataBlock) + return json.loads(data,cls=EmbeddingDecoder) + class Embedding: def __init__(self, vec, name, step=None): self.vec = vec @@ -113,6 +153,9 @@ class EmbeddingDatabase: if 'sd-ti-embedding' in embed_image.text: data = embeddingFromB64(embed_image.text['sd-ti-embedding']) name = data.get('name',name) + else: + data = extractImageDataFooter(embed_image) + name = data.get('name',name) else: data = torch.load(path, map_location="cpu") @@ -190,7 +233,7 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): return fn -def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, num_repeats, create_image_every, save_embedding_every, template_file): +def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, num_repeats, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding): assert embedding_name, 'embedding not selected' shared.state.textinfo = "Initializing textual inversion training..." @@ -308,6 +351,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini footer_right = '{}'.format(embedding.step) captioned_image = captionImageOverlay(image,title,footer_left,footer_mid,footer_right) + captioned_image = appendImageDataFooter(captioned_image,data) captioned_image.save(last_saved_image_chunks, "PNG", pnginfo=info) -- cgit v1.2.3 From df6d0d9286279c41c4c67460c3158fa268697524 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Mon, 10 Oct 2022 15:43:09 +0100 Subject: convert back to rgb as some hosts add alpha --- 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 6fb64691..667a7cf2 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -64,7 +64,7 @@ def crop_black(img,tol=0): def extractImageDataFooter(image): d=3 - outarr = crop_black(np.array(image.getdata()).reshape(image.size[1],image.size[0],d ).astype(np.uint8) ) + outarr = crop_black(np.array(image.convert('RGB').getdata()).reshape(image.size[1],image.size[0],d ).astype(np.uint8) ) lastRow = np.where( np.sum(outarr, axis=(1,2))==0) if lastRow[0].shape[0] == 0: print('Image data block not found.') -- cgit v1.2.3 From bc3e183b739913e7be91213a256f038b10eb71e9 Mon Sep 17 00:00:00 2001 From: alg-wiki Date: Tue, 11 Oct 2022 04:30:13 +0900 Subject: Textual Inversion: Preprocess and Training will only pick-up image files --- modules/textual_inversion/textual_inversion.py | 3 ++- 1 file changed, 2 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 5965c5a0..45397be9 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -161,6 +161,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini shared.state.textinfo = "Initializing textual inversion training..." shared.state.job_count = steps + extns = [".jpg",".jpeg",".png"] filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt') @@ -200,7 +201,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini if ititial_step > steps: return embedding, filename - tr_img_len = len([os.path.join(data_root, file_path) for file_path in os.listdir(data_root)]) + tr_img_len = len([os.path.join(data_root, file_path) for file_path in os.listdir(data_root) if os.path.splitext(file_path.casefold())[1] in extns]) epoch_len = (tr_img_len * num_repeats) + tr_img_len pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) -- cgit v1.2.3 From 2536ecbb1790da2af0d61b6a26f38732cba665cd Mon Sep 17 00:00:00 2001 From: Fampai <> Date: Mon, 10 Oct 2022 17:10:29 -0400 Subject: Refactored learning rate code --- modules/textual_inversion/textual_inversion.py | 51 ++++++++++++++++++++++++-- 1 file changed, 47 insertions(+), 4 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 5965c5a0..c64a4598 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -189,8 +189,6 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini embedding = hijack.embedding_db.word_embeddings[embedding_name] embedding.vec.requires_grad = True - optimizer = torch.optim.AdamW([embedding.vec], lr=learn_rate) - losses = torch.zeros((32,)) last_saved_file = "" @@ -203,12 +201,24 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini tr_img_len = len([os.path.join(data_root, file_path) for file_path in os.listdir(data_root)]) epoch_len = (tr_img_len * num_repeats) + tr_img_len + scheduleIter = iter(LearnSchedule(learn_rate, steps, ititial_step)) + (learn_rate, end_step) = next(scheduleIter) + print(f'Training at rate of {learn_rate} until step {end_step}') + + optimizer = torch.optim.AdamW([embedding.vec], lr=learn_rate) + pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) for i, (x, text) in pbar: embedding.step = i + ititial_step - if embedding.step > steps: - break + if embedding.step > end_step: + try: + (learn_rate, end_step) = next(scheduleIter) + except: + break + tqdm.tqdm.write(f'Training at rate of {learn_rate} until step {end_step}') + for pg in optimizer.param_groups: + pg['lr'] = learn_rate if shared.state.interrupted: break @@ -277,3 +287,36 @@ Last saved image: {html.escape(last_saved_image)}
return embedding, filename +class LearnSchedule: + def __init__(self, learn_rate, max_steps, cur_step=0): + pairs = learn_rate.split(',') + self.rates = [] + self.it = 0 + self.maxit = 0 + for i, pair in enumerate(pairs): + tmp = pair.split(':') + if len(tmp) == 2: + step = int(tmp[1]) + if step > cur_step: + self.rates.append((float(tmp[0]), min(step, max_steps))) + self.maxit += 1 + if step > max_steps: + return + elif step == -1: + self.rates.append((float(tmp[0]), max_steps)) + self.maxit += 1 + return + else: + self.rates.append((float(tmp[0]), max_steps)) + self.maxit += 1 + return + + def __iter__(self): + return self + + def __next__(self): + if self.it < self.maxit: + self.it += 1 + return self.rates[self.it - 1] + else: + raise StopIteration -- cgit v1.2.3 From 907a88b2d0be320575c2129d8d6a1d4f3a68f9eb Mon Sep 17 00:00:00 2001 From: alg-wiki Date: Tue, 11 Oct 2022 06:33:08 +0900 Subject: Added .webp .bmp --- 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 a03b299c..33c923d1 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -161,7 +161,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini shared.state.textinfo = "Initializing textual inversion training..." shared.state.job_count = steps - extns = [".jpg",".jpeg",".png"] + extns = [".jpg",".jpeg",".png",".webp",".bmp"] filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt') -- cgit v1.2.3 From 315d5a8ed975c88f670bc484f40a23fbf3a77b63 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Mon, 10 Oct 2022 23:14:44 +0100 Subject: update data dis[play style --- modules/textual_inversion/textual_inversion.py | 88 +++++++++++++++++++------- 1 file changed, 65 insertions(+), 23 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 667a7cf2..95eebea7 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -39,20 +39,59 @@ def embeddingFromB64(data): d = base64.b64decode(data) return json.loads(d,cls=EmbeddingDecoder) -def appendImageDataFooter(image,data): +def xorBlock(block): + return np.bitwise_xor(block.astype(np.uint8), + ((np.random.RandomState(0xDEADBEEF).random(block.shape)*255).astype(np.uint8)) & 0x0F ) + +def styleBlock(block,sequence): + im = Image.new('RGB',(block.shape[1],block.shape[0])) + draw = ImageDraw.Draw(im) + i=0 + for x in range(-6,im.size[0],8): + for yi,y in enumerate(range(-6,im.size[1],8)): + offset=0 + if yi%2==0: + offset=4 + shade = sequence[i%len(sequence)] + i+=1 + draw.ellipse((x+offset, y, x+6+offset, y+6), fill =(shade,shade,shade) ) + + fg = np.array(im).astype(np.uint8) & 0xF0 + return block ^ fg + +def insertImageDataEmbed(image,data): d = 3 data_compressed = zlib.compress( json.dumps(data,cls=EmbeddingEncoder).encode(),level=9) dnp = np.frombuffer(data_compressed,np.uint8).copy() - w = image.size[0] - next_size = dnp.shape[0] + (w-(dnp.shape[0]%w)) - next_size = next_size + ((w*d)-(next_size%(w*d))) - dnp.resize(next_size) - dnp = dnp.reshape((-1,w,d)) - print(dnp.shape) - im = Image.fromarray(dnp,mode='RGB') - background = Image.new('RGB',(image.size[0],image.size[1]+im.size[1]+1),(0,0,0)) - background.paste(image,(0,0)) - background.paste(im,(0,image.size[1]+1)) + dnphigh = dnp >> 4 + dnplow = dnp & 0x0F + + h = image.size[1] + next_size = dnplow.shape[0] + (h-(dnplow.shape[0]%h)) + next_size = next_size + ((h*d)-(next_size%(h*d))) + + dnplow.resize(next_size) + dnplow = dnplow.reshape((h,-1,d)) + + dnphigh.resize(next_size) + dnphigh = dnphigh.reshape((h,-1,d)) + + edgeStyleWeights = list(data['string_to_param'].values())[0].cpu().detach().numpy().tolist()[0][:1024] + edgeStyleWeights = (np.abs(edgeStyleWeights)/np.max(np.abs(edgeStyleWeights))*255).astype(np.uint8) + + dnplow = styleBlock(dnplow,sequence=edgeStyleWeights) + dnplow = xorBlock(dnplow) + dnphigh = styleBlock(dnphigh,sequence=edgeStyleWeights[::-1]) + dnphigh = xorBlock(dnphigh) + + imlow = Image.fromarray(dnplow,mode='RGB') + imhigh = Image.fromarray(dnphigh,mode='RGB') + + background = Image.new('RGB',(image.size[0]+imlow.size[0]+imhigh.size[0]+2,image.size[1]),(0,0,0)) + background.paste(imlow,(0,0)) + background.paste(image,(imlow.size[0]+1,0)) + background.paste(imhigh,(imlow.size[0]+1+image.size[0]+1,0)) + return background def crop_black(img,tol=0): @@ -62,19 +101,22 @@ def crop_black(img,tol=0): row_start,row_end = mask1.argmax(),mask.shape[0]-mask1[::-1].argmax() return img[row_start:row_end,col_start:col_end] -def extractImageDataFooter(image): +def extractImageDataEmbed(image): d=3 - outarr = crop_black(np.array(image.convert('RGB').getdata()).reshape(image.size[1],image.size[0],d ).astype(np.uint8) ) - lastRow = np.where( np.sum(outarr, axis=(1,2))==0) - if lastRow[0].shape[0] == 0: - print('Image data block not found.') + outarr = crop_black(np.array(image.getdata()).reshape(image.size[1],image.size[0],d ).astype(np.uint8) ) & 0x0F + blackCols = np.where( np.sum(outarr, axis=(0,2))==0) + if blackCols[0].shape[0] < 2: + print('No Image data blocks found.') return None - lastRow = lastRow[0] - - lastRow = lastRow.max() - dataBlock = outarr[lastRow+1::].astype(np.uint8).flatten().tobytes() - print(lastRow) + dataBlocklower = outarr[:,:blackCols[0].min(),:].astype(np.uint8) + dataBlockupper = outarr[:,blackCols[0].max()+1:,:].astype(np.uint8) + + dataBlocklower = xorBlock(dataBlocklower) + dataBlockupper = xorBlock(dataBlockupper) + + dataBlock = (dataBlockupper << 4) | (dataBlocklower) + dataBlock = dataBlock.flatten().tobytes() data = zlib.decompress(dataBlock) return json.loads(data,cls=EmbeddingDecoder) @@ -154,7 +196,7 @@ class EmbeddingDatabase: data = embeddingFromB64(embed_image.text['sd-ti-embedding']) name = data.get('name',name) else: - data = extractImageDataFooter(embed_image) + data = extractImageDataEmbed(embed_image) name = data.get('name',name) else: data = torch.load(path, map_location="cpu") @@ -351,7 +393,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini footer_right = '{}'.format(embedding.step) captioned_image = captionImageOverlay(image,title,footer_left,footer_mid,footer_right) - captioned_image = appendImageDataFooter(captioned_image,data) + captioned_image = insertImageDataEmbed(captioned_image,data) captioned_image.save(last_saved_image_chunks, "PNG", pnginfo=info) -- cgit v1.2.3 From 767202a4c324f9b49f63ab4dabbb5736fe9df6e5 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Mon, 10 Oct 2022 23:20:52 +0100 Subject: add dependency --- 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 95eebea7..f3cacaa0 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -7,7 +7,7 @@ import tqdm import html import datetime -from PIL import Image,PngImagePlugin +from PIL import Image,PngImagePlugin,ImageDraw from ..images import captionImageOverlay import numpy as np import base64 -- cgit v1.2.3 From e0fbe6d27e7b4505766c8cb5a4264e1114cf3721 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Mon, 10 Oct 2022 23:26:24 +0100 Subject: colour depth conversion fix --- 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 f3cacaa0..ae807268 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -103,7 +103,7 @@ def crop_black(img,tol=0): def extractImageDataEmbed(image): d=3 - outarr = crop_black(np.array(image.getdata()).reshape(image.size[1],image.size[0],d ).astype(np.uint8) ) & 0x0F + outarr = crop_black(np.array(image.convert('RGB').getdata()).reshape(image.size[1],image.size[0],d ).astype(np.uint8) ) & 0x0F blackCols = np.where( np.sum(outarr, axis=(0,2))==0) if blackCols[0].shape[0] < 2: print('No Image data blocks found.') -- cgit v1.2.3 From 7aa8fcac1e45c3ad9c6a40df0e44a346afcd5032 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Tue, 11 Oct 2022 04:17:36 +0100 Subject: use simple lcg in xor --- modules/textual_inversion/textual_inversion.py | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 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 ae807268..13416a08 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -39,9 +39,15 @@ def embeddingFromB64(data): d = base64.b64decode(data) return json.loads(d,cls=EmbeddingDecoder) +def lcg(m=2**32, a=1664525, c=1013904223, seed=0): + while True: + seed = (a * seed + c) % m + yield seed + def xorBlock(block): - return np.bitwise_xor(block.astype(np.uint8), - ((np.random.RandomState(0xDEADBEEF).random(block.shape)*255).astype(np.uint8)) & 0x0F ) + g = lcg() + randblock = np.array([next(g) for _ in range(np.product(block.shape))]).astype(np.uint8).reshape(block.shape) + return np.bitwise_xor(block.astype(np.uint8),randblock & 0x0F) def styleBlock(block,sequence): im = Image.new('RGB',(block.shape[1],block.shape[0])) -- cgit v1.2.3 From b2368a3bce663f19a7209d9cb38617e635ca6e3c Mon Sep 17 00:00:00 2001 From: alg-wiki Date: Tue, 11 Oct 2022 17:32:46 +0900 Subject: Switched to exception handling --- modules/textual_inversion/textual_inversion.py | 18 ++++++++---------- 1 file changed, 8 insertions(+), 10 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 33c923d1..91cde04b 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -161,7 +161,6 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini shared.state.textinfo = "Initializing textual inversion training..." shared.state.job_count = steps - extns = [".jpg",".jpeg",".png",".webp",".bmp"] filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt') @@ -201,10 +200,6 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini if ititial_step > steps: return embedding, filename - tr_img_len = len([os.path.join(data_root, file_path) for file_path in os.listdir(data_root) if os.path.splitext(file_path.casefold())[1] in extns]) - - epoch_len = (tr_img_len * num_repeats) + tr_img_len - pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) for i, (x, text) in pbar: embedding.step = i + ititial_step @@ -228,10 +223,10 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini loss.backward() optimizer.step() - epoch_num = embedding.step // epoch_len - epoch_step = embedding.step - (epoch_num * epoch_len) + 1 + epoch_num = embedding.step // len(ds) + epoch_step = embedding.step - (epoch_num * len(ds)) + 1 - pbar.set_description(f"[Epoch {epoch_num}: {epoch_step}/{epoch_len}]loss: {losses.mean():.7f}") + pbar.set_description(f"[Epoch {epoch_num}: {epoch_step}/{len(ds)}]loss: {losses.mean():.7f}") if embedding.step > 0 and embedding_dir is not None and embedding.step % save_embedding_every == 0: last_saved_file = os.path.join(embedding_dir, f'{embedding_name}-{embedding.step}.pt') @@ -243,9 +238,12 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini p = processing.StableDiffusionProcessingTxt2Img( sd_model=shared.sd_model, prompt=text, - steps=20, - height=training_height, + steps=28, + height=768, width=training_width, + negative_prompt="lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts,signature, watermark, username, blurry, artist name", + cfg_scale=7.0, + sampler_index=0, do_not_save_grid=True, do_not_save_samples=True, ) -- cgit v1.2.3 From 8bacbca0a1ab9aabcb0ad0cbf070e0006991e98a Mon Sep 17 00:00:00 2001 From: alg-wiki Date: Tue, 11 Oct 2022 17:35:09 +0900 Subject: Removed my local edits to checkpoint image generation --- modules/textual_inversion/textual_inversion.py | 7 ++----- 1 file changed, 2 insertions(+), 5 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 91cde04b..e9ff80c2 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -238,12 +238,9 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini p = processing.StableDiffusionProcessingTxt2Img( sd_model=shared.sd_model, prompt=text, - steps=28, - height=768, + steps=20, + height=training_height, width=training_width, - negative_prompt="lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts,signature, watermark, username, blurry, artist name", - cfg_scale=7.0, - sampler_index=0, do_not_save_grid=True, do_not_save_samples=True, ) -- cgit v1.2.3 From 530103b586109c11fd068eb70ef09503ec6a4caf Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 11 Oct 2022 14:53:02 +0300 Subject: fixes related to merge --- modules/textual_inversion/textual_inversion.py | 12 +++++++----- 1 file changed, 7 insertions(+), 5 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 5965c5a0..d6977950 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -156,7 +156,7 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): return fn -def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, num_repeats, create_image_every, save_embedding_every, template_file): +def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, num_repeats, create_image_every, save_embedding_every, template_file, preview_image_prompt): assert embedding_name, 'embedding not selected' shared.state.textinfo = "Initializing textual inversion training..." @@ -238,12 +238,14 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini if embedding.step > 0 and images_dir is not None and embedding.step % create_image_every == 0: last_saved_image = os.path.join(images_dir, f'{embedding_name}-{embedding.step}.png') + preview_text = text if preview_image_prompt == "" else preview_image_prompt + p = processing.StableDiffusionProcessingTxt2Img( sd_model=shared.sd_model, - prompt=text, + prompt=preview_text, steps=20, - height=training_height, - width=training_width, + height=training_height, + width=training_width, do_not_save_grid=True, do_not_save_samples=True, ) @@ -254,7 +256,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini shared.state.current_image = image image.save(last_saved_image) - last_saved_image += f", prompt: {text}" + last_saved_image += f", prompt: {preview_text}" shared.state.job_no = embedding.step -- cgit v1.2.3 From d4ea5f4d8631f778d11efcde397e4a5b8801d43b Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 11 Oct 2022 19:03:08 +0300 Subject: add an option to unload models during hypernetwork training to save VRAM --- 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 bb05cdc6..35f4bd9e 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -201,7 +201,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini return embedding, filename pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) - for i, (x, text) in pbar: + for i, (x, text, _) in pbar: embedding.step = i + ititial_step if embedding.step > steps: -- cgit v1.2.3 From 61788c0538415fa9ca1dd1b306519c116b18bd2c Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Tue, 11 Oct 2022 19:50:50 +0100 Subject: shift embedding logic out of textual_inversion --- modules/textual_inversion/textual_inversion.py | 125 ++----------------------- 1 file changed, 6 insertions(+), 119 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 8c66aeb5..22b4ae7f 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -7,124 +7,11 @@ import tqdm import html import datetime -from PIL import Image,PngImagePlugin,ImageDraw -from ..images import captionImageOverlay -import numpy as np -import base64 -import json -import zlib +from PIL import Image,PngImagePlugin from modules import shared, devices, sd_hijack, processing, sd_models import modules.textual_inversion.dataset -class EmbeddingEncoder(json.JSONEncoder): - def default(self, obj): - if isinstance(obj, torch.Tensor): - return {'TORCHTENSOR':obj.cpu().detach().numpy().tolist()} - return json.JSONEncoder.default(self, obj) - -class EmbeddingDecoder(json.JSONDecoder): - def __init__(self, *args, **kwargs): - json.JSONDecoder.__init__(self, object_hook=self.object_hook, *args, **kwargs) - def object_hook(self, d): - if 'TORCHTENSOR' in d: - return torch.from_numpy(np.array(d['TORCHTENSOR'])) - return d - -def embeddingToB64(data): - d = json.dumps(data,cls=EmbeddingEncoder) - return base64.b64encode(d.encode()) - -def embeddingFromB64(data): - d = base64.b64decode(data) - return json.loads(d,cls=EmbeddingDecoder) - -def lcg(m=2**32, a=1664525, c=1013904223, seed=0): - while True: - seed = (a * seed + c) % m - yield seed - -def xorBlock(block): - g = lcg() - randblock = np.array([next(g) for _ in range(np.product(block.shape))]).astype(np.uint8).reshape(block.shape) - return np.bitwise_xor(block.astype(np.uint8),randblock & 0x0F) - -def styleBlock(block,sequence): - im = Image.new('RGB',(block.shape[1],block.shape[0])) - draw = ImageDraw.Draw(im) - i=0 - for x in range(-6,im.size[0],8): - for yi,y in enumerate(range(-6,im.size[1],8)): - offset=0 - if yi%2==0: - offset=4 - shade = sequence[i%len(sequence)] - i+=1 - draw.ellipse((x+offset, y, x+6+offset, y+6), fill =(shade,shade,shade) ) - - fg = np.array(im).astype(np.uint8) & 0xF0 - return block ^ fg - -def insertImageDataEmbed(image,data): - d = 3 - data_compressed = zlib.compress( json.dumps(data,cls=EmbeddingEncoder).encode(),level=9) - dnp = np.frombuffer(data_compressed,np.uint8).copy() - dnphigh = dnp >> 4 - dnplow = dnp & 0x0F - - h = image.size[1] - next_size = dnplow.shape[0] + (h-(dnplow.shape[0]%h)) - next_size = next_size + ((h*d)-(next_size%(h*d))) - - dnplow.resize(next_size) - dnplow = dnplow.reshape((h,-1,d)) - - dnphigh.resize(next_size) - dnphigh = dnphigh.reshape((h,-1,d)) - - edgeStyleWeights = list(data['string_to_param'].values())[0].cpu().detach().numpy().tolist()[0][:1024] - edgeStyleWeights = (np.abs(edgeStyleWeights)/np.max(np.abs(edgeStyleWeights))*255).astype(np.uint8) - - dnplow = styleBlock(dnplow,sequence=edgeStyleWeights) - dnplow = xorBlock(dnplow) - dnphigh = styleBlock(dnphigh,sequence=edgeStyleWeights[::-1]) - dnphigh = xorBlock(dnphigh) - - imlow = Image.fromarray(dnplow,mode='RGB') - imhigh = Image.fromarray(dnphigh,mode='RGB') - - background = Image.new('RGB',(image.size[0]+imlow.size[0]+imhigh.size[0]+2,image.size[1]),(0,0,0)) - background.paste(imlow,(0,0)) - background.paste(image,(imlow.size[0]+1,0)) - background.paste(imhigh,(imlow.size[0]+1+image.size[0]+1,0)) - - return background - -def crop_black(img,tol=0): - mask = (img>tol).all(2) - mask0,mask1 = mask.any(0),mask.any(1) - col_start,col_end = mask0.argmax(),mask.shape[1]-mask0[::-1].argmax() - row_start,row_end = mask1.argmax(),mask.shape[0]-mask1[::-1].argmax() - return img[row_start:row_end,col_start:col_end] - -def extractImageDataEmbed(image): - d=3 - outarr = crop_black(np.array(image.convert('RGB').getdata()).reshape(image.size[1],image.size[0],d ).astype(np.uint8) ) & 0x0F - blackCols = np.where( np.sum(outarr, axis=(0,2))==0) - if blackCols[0].shape[0] < 2: - print('No Image data blocks found.') - return None - - dataBlocklower = outarr[:,:blackCols[0].min(),:].astype(np.uint8) - dataBlockupper = outarr[:,blackCols[0].max()+1:,:].astype(np.uint8) - - dataBlocklower = xorBlock(dataBlocklower) - dataBlockupper = xorBlock(dataBlockupper) - - dataBlock = (dataBlockupper << 4) | (dataBlocklower) - dataBlock = dataBlock.flatten().tobytes() - data = zlib.decompress(dataBlock) - return json.loads(data,cls=EmbeddingDecoder) class Embedding: def __init__(self, vec, name, step=None): @@ -199,10 +86,10 @@ class EmbeddingDatabase: if filename.upper().endswith('.PNG'): embed_image = Image.open(path) if 'sd-ti-embedding' in embed_image.text: - data = embeddingFromB64(embed_image.text['sd-ti-embedding']) + data = embedding_from_b64(embed_image.text['sd-ti-embedding']) name = data.get('name',name) else: - data = extractImageDataEmbed(embed_image) + data = extract_image_data_embed(embed_image) name = data.get('name',name) else: data = torch.load(path, map_location="cpu") @@ -393,7 +280,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini info = PngImagePlugin.PngInfo() data = torch.load(last_saved_file) - info.add_text("sd-ti-embedding", embeddingToB64(data)) + info.add_text("sd-ti-embedding", embedding_to_b64(data)) title = "<{}>".format(data.get('name','???')) checkpoint = sd_models.select_checkpoint() @@ -401,8 +288,8 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini footer_mid = '[{}]'.format(checkpoint.hash) footer_right = '{}'.format(embedding.step) - captioned_image = captionImageOverlay(image,title,footer_left,footer_mid,footer_right) - captioned_image = insertImageDataEmbed(captioned_image,data) + 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) -- cgit v1.2.3 From d6fcc6b87bc00fcdecea276fe5b7c7945f7a8b14 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 11 Oct 2022 22:03:05 +0300 Subject: apply lr schedule to hypernets --- modules/textual_inversion/textual_inversion.py | 44 +++----------------------- 1 file changed, 4 insertions(+), 40 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 47a27faf..7717837d 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -10,6 +10,7 @@ import datetime from modules import shared, devices, sd_hijack, processing, sd_models import modules.textual_inversion.dataset +from modules.textual_inversion.learn_schedule import LearnSchedule class Embedding: @@ -198,11 +199,8 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini if ititial_step > steps: return embedding, filename - tr_img_len = len([os.path.join(data_root, file_path) for file_path in os.listdir(data_root)]) - epoch_len = (tr_img_len * num_repeats) + tr_img_len - - scheduleIter = iter(LearnSchedule(learn_rate, steps, ititial_step)) - (learn_rate, end_step) = next(scheduleIter) + schedules = iter(LearnSchedule(learn_rate, steps, ititial_step)) + (learn_rate, end_step) = next(schedules) print(f'Training at rate of {learn_rate} until step {end_step}') optimizer = torch.optim.AdamW([embedding.vec], lr=learn_rate) @@ -213,7 +211,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini if embedding.step > end_step: try: - (learn_rate, end_step) = next(scheduleIter) + (learn_rate, end_step) = next(schedules) except: break tqdm.tqdm.write(f'Training at rate of {learn_rate} until step {end_step}') @@ -288,37 +286,3 @@ Last saved image: {html.escape(last_saved_image)}
embedding.save(filename) return embedding, filename - -class LearnSchedule: - def __init__(self, learn_rate, max_steps, cur_step=0): - pairs = learn_rate.split(',') - self.rates = [] - self.it = 0 - self.maxit = 0 - for i, pair in enumerate(pairs): - tmp = pair.split(':') - if len(tmp) == 2: - step = int(tmp[1]) - if step > cur_step: - self.rates.append((float(tmp[0]), min(step, max_steps))) - self.maxit += 1 - if step > max_steps: - return - elif step == -1: - self.rates.append((float(tmp[0]), max_steps)) - self.maxit += 1 - return - else: - self.rates.append((float(tmp[0]), max_steps)) - self.maxit += 1 - return - - def __iter__(self): - return self - - def __next__(self): - if self.it < self.maxit: - self.it += 1 - return self.rates[self.it - 1] - else: - raise StopIteration -- cgit v1.2.3 From aa75d5cfe8c84768b0f5d16f977ddba298677379 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Tue, 11 Oct 2022 20:06:13 +0100 Subject: correct conflict resolution typo --- 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 22b4ae7f..789383ce 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -169,7 +169,7 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): -def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, num_repeats, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_image_prompt) +def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, num_repeats, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_image_prompt): assert embedding_name, 'embedding not selected' shared.state.textinfo = "Initializing textual inversion training..." -- cgit v1.2.3 From 91d7ee0d097a7ea203d261b570cd2b834837d9e2 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Tue, 11 Oct 2022 20:09:10 +0100 Subject: update imports --- modules/textual_inversion/textual_inversion.py | 3 +++ 1 file changed, 3 insertions(+) (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 789383ce..ff0a62b3 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -12,6 +12,9 @@ from PIL import Image,PngImagePlugin from modules import shared, devices, sd_hijack, processing, sd_models import modules.textual_inversion.dataset +from modules.textual_inversion.image_embedding import( embedding_to_b64,embedding_from_b64, + insert_image_data_embed,extract_image_data_embed, + caption_image_overlay ) class Embedding: def __init__(self, vec, name, step=None): -- cgit v1.2.3 From 5f3317376bb7952bc5145f05f16c1bbd466efc85 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Tue, 11 Oct 2022 20:09:49 +0100 Subject: spacing --- 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 ff0a62b3..485ef46c 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -12,7 +12,7 @@ from PIL import Image,PngImagePlugin from modules import shared, devices, sd_hijack, processing, sd_models import modules.textual_inversion.dataset -from modules.textual_inversion.image_embedding import( embedding_to_b64,embedding_from_b64, +from modules.textual_inversion.image_embedding import (embedding_to_b64,embedding_from_b64, insert_image_data_embed,extract_image_data_embed, caption_image_overlay ) -- cgit v1.2.3 From 10a2de644f8ea4cfade88e85d768da3480f4c9f0 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Wed, 12 Oct 2022 13:15:35 +0100 Subject: formatting --- modules/textual_inversion/textual_inversion.py | 22 +++++++++++----------- 1 file changed, 11 insertions(+), 11 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 485ef46c..b072d745 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -7,14 +7,14 @@ import tqdm import html import datetime -from PIL import Image,PngImagePlugin +from PIL import Image, PngImagePlugin from modules import shared, devices, sd_hijack, processing, sd_models import modules.textual_inversion.dataset -from modules.textual_inversion.image_embedding import (embedding_to_b64,embedding_from_b64, - insert_image_data_embed,extract_image_data_embed, - caption_image_overlay ) +from modules.textual_inversion.image_embedding import (embedding_to_b64, embedding_from_b64, + insert_image_data_embed, extract_image_data_embed, + caption_image_overlay) class Embedding: def __init__(self, vec, name, step=None): @@ -90,10 +90,10 @@ class EmbeddingDatabase: embed_image = Image.open(path) if 'sd-ti-embedding' in embed_image.text: data = embedding_from_b64(embed_image.text['sd-ti-embedding']) - name = data.get('name',name) + name = data.get('name', name) else: data = extract_image_data_embed(embed_image) - name = data.get('name',name) + name = data.get('name', name) else: data = torch.load(path, map_location="cpu") @@ -278,24 +278,24 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini shared.state.current_image = image if save_image_with_stored_embedding and os.path.exists(last_saved_file): - + last_saved_image_chunks = os.path.join(images_embeds_dir, f'{embedding_name}-{embedding.step}.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','???')) + title = "<{}>".format(data.get('name', '???')) checkpoint = sd_models.select_checkpoint() footer_left = checkpoint.model_name footer_mid = '[{}]'.format(checkpoint.hash) footer_right = '{}'.format(embedding.step) - captioned_image = caption_image_overlay(image,title,footer_left,footer_mid,footer_right) - captioned_image = insert_image_data_embed(captioned_image,data) + 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) - + image.save(last_saved_image) last_saved_image += f", prompt: {preview_text}" -- cgit v1.2.3 From c3c8eef9fd5a0c8b26319e32ca4a19b56204e6df Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Wed, 12 Oct 2022 20:49:47 +0300 Subject: train: change filename processing to be more simple and configurable train: make it possible to make text files with prompts train: rework scheduler so that there's less repeating code in textual inversion and hypernets train: move epochs setting to options --- modules/textual_inversion/textual_inversion.py | 35 ++++++++++---------------- 1 file changed, 13 insertions(+), 22 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 c5153e4a..fa0e33a2 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -11,7 +11,7 @@ from PIL import Image, PngImagePlugin from modules import shared, devices, sd_hijack, processing, sd_models import modules.textual_inversion.dataset -from modules.textual_inversion.learn_schedule import LearnSchedule +from modules.textual_inversion.learn_schedule import LearnRateScheduler from modules.textual_inversion.image_embedding import (embedding_to_b64, embedding_from_b64, insert_image_data_embed, extract_image_data_embed, @@ -172,8 +172,7 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): return fn - -def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, num_repeats, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_image_prompt): +def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_image_prompt): assert embedding_name, 'embedding not selected' shared.state.textinfo = "Initializing textual inversion training..." @@ -205,7 +204,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini 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=num_repeats, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file) + 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) hijack = sd_hijack.model_hijack @@ -221,32 +220,24 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini if ititial_step > steps: return embedding, filename - schedules = iter(LearnSchedule(learn_rate, steps, ititial_step)) - (learn_rate, end_step) = next(schedules) - print(f'Training at rate of {learn_rate} until step {end_step}') - - optimizer = torch.optim.AdamW([embedding.vec], lr=learn_rate) + scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) + optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate) pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) - for i, (x, text, _) in pbar: + for i, entry in pbar: embedding.step = i + ititial_step - if embedding.step > end_step: - try: - (learn_rate, end_step) = next(schedules) - except: - break - tqdm.tqdm.write(f'Training at rate of {learn_rate} until step {end_step}') - for pg in optimizer.param_groups: - pg['lr'] = learn_rate + scheduler.apply(optimizer, embedding.step) + if scheduler.finished: + break if shared.state.interrupted: break with torch.autocast("cuda"): - c = cond_model([text]) + c = cond_model([entry.cond_text]) - x = x.to(devices.device) + x = entry.latent.to(devices.device) loss = shared.sd_model(x.unsqueeze(0), c)[0] del x @@ -268,7 +259,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini if embedding.step > 0 and images_dir is not None and embedding.step % create_image_every == 0: last_saved_image = os.path.join(images_dir, f'{embedding_name}-{embedding.step}.png') - preview_text = text if preview_image_prompt == "" else preview_image_prompt + preview_text = entry.cond_text if preview_image_prompt == "" else preview_image_prompt p = processing.StableDiffusionProcessingTxt2Img( sd_model=shared.sd_model, @@ -314,7 +305,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini

Loss: {losses.mean():.7f}
Step: {embedding.step}
-Last prompt: {html.escape(text)}
+Last prompt: {html.escape(entry.cond_text)}
Last saved embedding: {html.escape(last_saved_file)}
Last saved image: {html.escape(last_saved_image)}

-- cgit v1.2.3 From 1cfc2a18981ee56bdb69a2de7b463a11ad05e329 Mon Sep 17 00:00:00 2001 From: Melan Date: Wed, 12 Oct 2022 23:36:29 +0200 Subject: Save a csv containing the loss while training --- modules/textual_inversion/textual_inversion.py | 17 ++++++++++++++++- 1 file changed, 16 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 fa0e33a2..25038a89 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -6,6 +6,7 @@ import torch import tqdm import html import datetime +import csv from PIL import Image, PngImagePlugin @@ -172,7 +173,7 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): return fn -def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_image_prompt): +def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, write_csv_every, template_file, save_image_with_stored_embedding, preview_image_prompt): assert embedding_name, 'embedding not selected' shared.state.textinfo = "Initializing textual inversion training..." @@ -256,6 +257,20 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini last_saved_file = os.path.join(embedding_dir, f'{embedding_name}-{embedding.step}.pt') embedding.save(last_saved_file) + if write_csv_every > 0 and log_directory is not None and embedding.step % write_csv_every == 0: + write_csv_header = False if os.path.exists(os.path.join(log_directory, "textual_inversion_loss.csv")) else True + + with open(os.path.join(log_directory, "textual_inversion_loss.csv"), "a+") as fout: + + csv_writer = csv.DictWriter(fout, fieldnames=["epoch", "epoch_step", "loss"]) + + if write_csv_header: + csv_writer.writeheader() + + csv_writer.writerow({"epoch": epoch_num + 1, + "epoch_step": epoch_step - 1, + "loss": f"{losses.mean():.7f}"}) + if embedding.step > 0 and images_dir is not None and embedding.step % create_image_every == 0: last_saved_image = os.path.join(images_dir, f'{embedding_name}-{embedding.step}.png') -- cgit v1.2.3 From 8636b50aea83f9c743f005722d9f3f8ee9303e00 Mon Sep 17 00:00:00 2001 From: Melan Date: Thu, 13 Oct 2022 12:37:58 +0200 Subject: Add learn_rate to csv and removed a left-over debug statement --- modules/textual_inversion/textual_inversion.py | 5 +++-- 1 file changed, 3 insertions(+), 2 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 25038a89..b83df079 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -262,14 +262,15 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini with open(os.path.join(log_directory, "textual_inversion_loss.csv"), "a+") as fout: - csv_writer = csv.DictWriter(fout, fieldnames=["epoch", "epoch_step", "loss"]) + csv_writer = csv.DictWriter(fout, fieldnames=["epoch", "epoch_step", "loss", "learn_rate"]) if write_csv_header: csv_writer.writeheader() csv_writer.writerow({"epoch": epoch_num + 1, "epoch_step": epoch_step - 1, - "loss": f"{losses.mean():.7f}"}) + "loss": f"{losses.mean():.7f}", + "learn_rate": scheduler.learn_rate}) if embedding.step > 0 and images_dir is not None and embedding.step % create_image_every == 0: last_saved_image = os.path.join(images_dir, f'{embedding_name}-{embedding.step}.png') -- cgit v1.2.3 From bb57f30c2de46cfca5419ad01738a41705f96cc3 Mon Sep 17 00:00:00 2001 From: MalumaDev Date: Fri, 14 Oct 2022 10:56:41 +0200 Subject: init --- modules/textual_inversion/textual_inversion.py | 35 ++++++++++++++++++-------- 1 file changed, 25 insertions(+), 10 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 fa0e33a2..b12a8e6d 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -172,7 +172,15 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): return fn -def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_image_prompt): +def batched(dataset, total, n=1): + for ndx in range(0, total, n): + yield [dataset.__getitem__(i) for i in range(ndx, min(ndx + n, total))] + + +def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, + create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, + preview_image_prompt, batch_size=1, + gradient_accumulation=1): assert embedding_name, 'embedding not selected' shared.state.textinfo = "Initializing textual inversion training..." @@ -204,7 +212,11 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini 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) + 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) hijack = sd_hijack.model_hijack @@ -223,7 +235,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate) - pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) + pbar = tqdm.tqdm(enumerate(batched(ds, steps - ititial_step, batch_size)), total=steps - ititial_step) for i, entry in pbar: embedding.step = i + ititial_step @@ -235,17 +247,20 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini break with torch.autocast("cuda"): - c = cond_model([entry.cond_text]) + c = cond_model([e.cond_text for e in entry]) + + x = torch.stack([e.latent for e in entry]).to(devices.device) + loss = shared.sd_model(x, c)[0] - x = entry.latent.to(devices.device) - loss = shared.sd_model(x.unsqueeze(0), c)[0] del x losses[embedding.step % losses.shape[0]] = loss.item() - optimizer.zero_grad() loss.backward() - optimizer.step() + if ((i + 1) % gradient_accumulation == 0) or (i + 1 == steps - ititial_step): + optimizer.step() + optimizer.zero_grad() + epoch_num = embedding.step // len(ds) epoch_step = embedding.step - (epoch_num * len(ds)) + 1 @@ -259,7 +274,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini if embedding.step > 0 and images_dir is not None and embedding.step % create_image_every == 0: last_saved_image = os.path.join(images_dir, f'{embedding_name}-{embedding.step}.png') - preview_text = entry.cond_text if preview_image_prompt == "" else preview_image_prompt + preview_text = entry[0].cond_text if preview_image_prompt == "" else preview_image_prompt p = processing.StableDiffusionProcessingTxt2Img( sd_model=shared.sd_model, @@ -305,7 +320,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini

Loss: {losses.mean():.7f}
Step: {embedding.step}
-Last prompt: {html.escape(entry.cond_text)}
+Last prompt: {html.escape(entry[-1].cond_text)}
Last saved embedding: {html.escape(last_saved_file)}
Last saved image: {html.escape(last_saved_image)}

-- cgit v1.2.3 From c344ba3b325459abbf9b0df2c1b18f7bf99805b2 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 14 Oct 2022 20:31:49 +0300 Subject: add option to read generation params for learning previews from txt2img --- modules/textual_inversion/textual_inversion.py | 25 ++++++++++++++++++------- 1 file changed, 18 insertions(+), 7 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 fa0e33a2..3d835358 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -172,7 +172,7 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): return fn -def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_image_prompt): +def train_embedding(embedding_name, learn_rate, 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): assert embedding_name, 'embedding not selected' shared.state.textinfo = "Initializing textual inversion training..." @@ -259,18 +259,29 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini if embedding.step > 0 and images_dir is not None and embedding.step % create_image_every == 0: last_saved_image = os.path.join(images_dir, f'{embedding_name}-{embedding.step}.png') - preview_text = entry.cond_text if preview_image_prompt == "" else preview_image_prompt - p = processing.StableDiffusionProcessingTxt2Img( sd_model=shared.sd_model, - prompt=preview_text, - steps=20, - height=training_height, - width=training_width, 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_index = preview_sampler_index + p.cfg_scale = preview_cfg_scale + p.seed = preview_seed + p.width = preview_width + p.height = preview_height + else: + p.prompt = entry.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] -- cgit v1.2.3 From 03d62538aebeff51713619fe808c953bdb70193d Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 14 Oct 2022 22:43:55 +0300 Subject: remove duplicate code for log loss, add step, make it read from options rather than gradio input --- modules/textual_inversion/textual_inversion.py | 44 ++++++++++++++++++-------- 1 file changed, 30 insertions(+), 14 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 1f5ace6f..da0d77a0 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -173,6 +173,32 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): return fn +def write_loss(log_directory, filename, step, epoch_len, values): + if shared.opts.training_write_csv_every == 0: + return + + 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 + + with open(os.path.join(log_directory, filename), "a+", newline='') as fout: + csv_writer = csv.DictWriter(fout, fieldnames=["step", "epoch", "epoch_step", *(values.keys())]) + + if write_csv_header: + csv_writer.writeheader() + + epoch = step // epoch_len + epoch_step = step - epoch * epoch_len + + csv_writer.writerow({ + "step": step + 1, + "epoch": epoch + 1, + "epoch_step": epoch_step + 1, + **values, + }) + + def train_embedding(embedding_name, learn_rate, 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): assert embedding_name, 'embedding not selected' @@ -257,20 +283,10 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini last_saved_file = os.path.join(embedding_dir, f'{embedding_name}-{embedding.step}.pt') embedding.save(last_saved_file) - if write_csv_every > 0 and log_directory is not None and embedding.step % write_csv_every == 0: - write_csv_header = False if os.path.exists(os.path.join(log_directory, "textual_inversion_loss.csv")) else True - - with open(os.path.join(log_directory, "textual_inversion_loss.csv"), "a+") as fout: - - csv_writer = csv.DictWriter(fout, fieldnames=["epoch", "epoch_step", "loss", "learn_rate"]) - - if write_csv_header: - csv_writer.writeheader() - - csv_writer.writerow({"epoch": epoch_num + 1, - "epoch_step": epoch_step - 1, - "loss": f"{losses.mean():.7f}", - "learn_rate": scheduler.learn_rate}) + write_loss(log_directory, "textual_inversion_loss.csv", embedding.step, len(ds), { + "loss": f"{losses.mean():.7f}", + "learn_rate": scheduler.learn_rate + }) if embedding.step > 0 and images_dir is not None and embedding.step % create_image_every == 0: last_saved_image = os.path.join(images_dir, f'{embedding_name}-{embedding.step}.png') -- cgit v1.2.3 From c7a86f7fe9c0b8967a87e8d709f507d2f44400d8 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sat, 15 Oct 2022 09:24:59 +0300 Subject: add option to use batch size for training --- modules/textual_inversion/textual_inversion.py | 17 ++++++++--------- 1 file changed, 8 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 da0d77a0..e754747e 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -199,7 +199,7 @@ def write_loss(log_directory, filename, step, epoch_len, values): }) -def train_embedding(embedding_name, learn_rate, 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, 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): assert embedding_name, 'embedding not selected' shared.state.textinfo = "Initializing textual inversion training..." @@ -231,7 +231,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini 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) + 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) hijack = sd_hijack.model_hijack @@ -251,7 +251,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate) pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) - for i, entry in pbar: + for i, entries in pbar: embedding.step = i + ititial_step scheduler.apply(optimizer, embedding.step) @@ -262,10 +262,9 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini break with torch.autocast("cuda"): - c = cond_model([entry.cond_text]) - - x = entry.latent.to(devices.device) - loss = shared.sd_model(x.unsqueeze(0), c)[0] + 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() @@ -307,7 +306,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini p.width = preview_width p.height = preview_height else: - p.prompt = entry.cond_text + p.prompt = entries[0].cond_text p.steps = 20 p.width = training_width p.height = training_height @@ -348,7 +347,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini

Loss: {losses.mean():.7f}
Step: {embedding.step}
-Last prompt: {html.escape(entry.cond_text)}
+Last prompt: {html.escape(entries[0].cond_text)}
Last saved embedding: {html.escape(last_saved_file)}
Last saved image: {html.escape(last_saved_image)}

-- cgit v1.2.3 From 9e846083b702a498fdb60accd72f075fa26701d9 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Fri, 14 Oct 2022 14:50:25 +0100 Subject: add vector size to embed text --- modules/textual_inversion/textual_inversion.py | 8 +++++++- 1 file changed, 7 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 e754747e..6f549d62 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -327,10 +327,16 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc 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 = '{}'.format(embedding.step) + footer_right = 'v{} {}s'.format(vectorSize, embedding.step) captioned_image = caption_image_overlay(image, title, footer_left, footer_mid, footer_right) captioned_image = insert_image_data_embed(captioned_image, data) -- cgit v1.2.3 From 939f16529a72fe48c2ce3ef31bdaba785925a33c Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Fri, 14 Oct 2022 14:55:05 +0100 Subject: only save 1 image per embedding --- modules/textual_inversion/textual_inversion.py | 5 ++++- 1 file changed, 4 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 6f549d62..1d697c90 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -242,6 +242,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc last_saved_file = "" last_saved_image = "" + embedding_yet_to_be_embedded = False ititial_step = embedding.step or 0 if ititial_step > steps: @@ -281,6 +282,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc if embedding.step > 0 and embedding_dir is not None and embedding.step % save_embedding_every == 0: last_saved_file = os.path.join(embedding_dir, f'{embedding_name}-{embedding.step}.pt') embedding.save(last_saved_file) + embedding_yet_to_be_embedded = True write_loss(log_directory, "textual_inversion_loss.csv", embedding.step, len(ds), { "loss": f"{losses.mean():.7f}", @@ -318,7 +320,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc shared.state.current_image = image - if save_image_with_stored_embedding and os.path.exists(last_saved_file): + 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}-{embedding.step}.png') @@ -342,6 +344,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc 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 image.save(last_saved_image) -- cgit v1.2.3 From 9a1dcd78edbf9caf68b9e6286d7b5ca81500e243 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Fri, 14 Oct 2022 18:14:02 +0100 Subject: add webp for embed load --- modules/textual_inversion/textual_inversion.py | 4 ++++ 1 file changed, 4 insertions(+) (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 1d697c90..c07bffc3 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -96,6 +96,10 @@ class EmbeddingDatabase: else: data = extract_image_data_embed(embed_image) name = data.get('name', name) + elif filename.upper().endswith('.WEBP'): + embed_image = Image.open(path) + data = extract_image_data_embed(embed_image) + name = data.get('name', name) else: data = torch.load(path, map_location="cpu") -- cgit v1.2.3 From ddf6899df0cf87d4da77cb2ce223061f4a5edf18 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Fri, 14 Oct 2022 18:23:20 +0100 Subject: generalise to popular lossless formats --- modules/textual_inversion/textual_inversion.py | 8 ++------ 1 file changed, 2 insertions(+), 6 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 c07bffc3..b99df3b1 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -88,18 +88,14 @@ class EmbeddingDatabase: data = [] - if filename.upper().endswith('.PNG'): + if os.path.splitext(filename.upper())[-1] in ['.PNG', '.WEBP', '.JXL', '.AVIF']: embed_image = Image.open(path) - if 'sd-ti-embedding' in embed_image.text: + if hasattr(embed_image, 'text') and 'sd-ti-embedding' in embed_image.text: data = embedding_from_b64(embed_image.text['sd-ti-embedding']) name = data.get('name', name) else: data = extract_image_data_embed(embed_image) name = data.get('name', name) - elif filename.upper().endswith('.WEBP'): - embed_image = Image.open(path) - data = extract_image_data_embed(embed_image) - name = data.get('name', name) else: data = torch.load(path, map_location="cpu") -- cgit v1.2.3 From b6e3b96dab94a00f51725f9cc977eebc6b4072ab Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Sat, 15 Oct 2022 15:17:21 +0100 Subject: Change vector size footer label --- 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 b99df3b1..2ed345b1 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -338,7 +338,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc checkpoint = sd_models.select_checkpoint() footer_left = checkpoint.model_name footer_mid = '[{}]'.format(checkpoint.hash) - footer_right = 'v{} {}s'.format(vectorSize, embedding.step) + footer_right = '{}v {}s'.format(vectorSize, embedding.step) captioned_image = caption_image_overlay(image, title, footer_left, footer_mid, footer_right) captioned_image = insert_image_data_embed(captioned_image, data) -- cgit v1.2.3 From 0c5fa9a681672508adadbe1e10fc16d7fe0ed6dd Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sun, 16 Oct 2022 08:51:24 +0300 Subject: do not reload embeddings from disk when doing textual inversion --- modules/textual_inversion/textual_inversion.py | 1 + 1 file changed, 1 insertion(+) (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 2ed345b1..7ec75018 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -296,6 +296,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc sd_model=shared.sd_model, do_not_save_grid=True, do_not_save_samples=True, + do_not_reload_embeddings=True, ) if preview_from_txt2img: -- cgit v1.2.3 From 62edfae257e8982cd620d03862c7bdd44159d18f Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Sun, 16 Oct 2022 20:28:15 +0100 Subject: print list of embeddings on reload --- modules/textual_inversion/textual_inversion.py | 1 + 1 file changed, 1 insertion(+) (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 7ec75018..3be69562 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -137,6 +137,7 @@ class EmbeddingDatabase: continue print(f"Loaded a total of {len(self.word_embeddings)} textual inversion embeddings.") + print("Embeddings:", ', '.join(self.word_embeddings.keys())) def find_embedding_at_position(self, tokens, offset): token = tokens[offset] -- cgit v1.2.3 From 0087079c2d487b67b06ffc30f36ce486a74e6318 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Thu, 20 Oct 2022 00:10:59 +0100 Subject: allow overwrite old embedding --- modules/textual_inversion/textual_inversion.py | 5 +++-- 1 file changed, 3 insertions(+), 2 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 3be69562..5776778b 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -153,7 +153,7 @@ class EmbeddingDatabase: return None, None -def create_embedding(name, num_vectors_per_token, init_text='*'): +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 @@ -165,7 +165,8 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): vec[i] = embedded[i * int(embedded.shape[0]) // num_vectors_per_token] fn = os.path.join(shared.cmd_opts.embeddings_dir, f"{name}.pt") - assert not os.path.exists(fn), f"file {fn} already exists" + if not overwrite_old: + assert not os.path.exists(fn), f"file {fn} already exists" embedding = Embedding(vec, name) embedding.step = 0 -- cgit v1.2.3 From cbb857b675cf0f169b21515c29da492b513cc8c4 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Wed, 26 Oct 2022 09:44:02 +0300 Subject: enable creating embedding with --medvram --- modules/textual_inversion/textual_inversion.py | 3 +++ 1 file changed, 3 insertions(+) (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 529ed3e2..647ffe3e 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -157,6 +157,9 @@ 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) vec = torch.zeros((num_vectors_per_token, embedded.shape[1]), device=devices.device) -- cgit v1.2.3 From c2dc9bfa89070b8e1d857f8773a790b752f1b709 Mon Sep 17 00:00:00 2001 From: timntorres Date: Mon, 24 Oct 2022 23:22:58 -0700 Subject: Implement PR #3189 but for embeddings. --- modules/textual_inversion/textual_inversion.py | 10 +++++----- 1 file changed, 5 insertions(+), 5 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 647ffe3e..22c7b54b 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 +from modules import shared, devices, sd_hijack, processing, sd_models, images import modules.textual_inversion.dataset from modules.textual_inversion.learn_schedule import LearnRateScheduler @@ -247,6 +247,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc last_saved_file = "" last_saved_image = "" + forced_filename = "" embedding_yet_to_be_embedded = False ititial_step = embedding.step or 0 @@ -296,8 +297,8 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc }) if embedding.step > 0 and images_dir is not None and embedding.step % create_image_every == 0: - last_saved_image = os.path.join(images_dir, f'{embedding_name}-{embedding.step}.png') - + forced_filename = f'{embedding_name}-{embedding.step}' + last_saved_image = os.path.join(images_dir, forced_filename) p = processing.StableDiffusionProcessingTxt2Img( sd_model=shared.sd_model, do_not_save_grid=True, @@ -353,8 +354,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc captioned_image.save(last_saved_image_chunks, "PNG", pnginfo=info) embedding_yet_to_be_embedded = False - image.save(last_saved_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) last_saved_image += f", prompt: {preview_text}" shared.state.job_no = embedding.step -- cgit v1.2.3 From 4875a6c217df5cc06ee2bf11fb645b172c7156a8 Mon Sep 17 00:00:00 2001 From: timntorres Date: Mon, 24 Oct 2022 23:38:07 -0700 Subject: Implement PR #3309 but for embeddings. --- modules/textual_inversion/textual_inversion.py | 9 ++++++++- 1 file changed, 8 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 22c7b54b..4921bd01 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -167,6 +167,8 @@ def create_embedding(name, num_vectors_per_token, overwrite_old, init_text='*'): for i in range(num_vectors_per_token): vec[i] = embedded[i * int(embedded.shape[0]) // num_vectors_per_token] + # Remove illegal characters from name. + name = "".join( x for x in name if (x.isalnum() or x in "._- ")) fn = os.path.join(shared.cmd_opts.embeddings_dir, f"{name}.pt") if not overwrite_old: assert not os.path.exists(fn), f"file {fn} already exists" @@ -287,7 +289,9 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc pbar.set_description(f"[Epoch {epoch_num}: {epoch_step}/{len(ds)}]loss: {losses.mean():.7f}") if embedding.step > 0 and embedding_dir is not None and embedding.step % save_embedding_every == 0: - last_saved_file = os.path.join(embedding_dir, f'{embedding_name}-{embedding.step}.pt') + # Before saving, change name to match current checkpoint. + embedding.name = f'{embedding_name}-{embedding.step}' + last_saved_file = os.path.join(embedding_dir, f'{embedding.name}.pt') embedding.save(last_saved_file) embedding_yet_to_be_embedded = True @@ -374,6 +378,9 @@ Last saved image: {html.escape(last_saved_image)}
embedding.sd_checkpoint = checkpoint.hash embedding.sd_checkpoint_name = checkpoint.model_name embedding.cached_checksum = None + # Before saving for the last time, change name back to base name (as opposed to the save_embedding_every step-suffixed naming convention). + embedding.name = embedding_name + filename = os.path.join(shared.cmd_opts.embedding_dir, f'{embedding.name}.pt') embedding.save(filename) return embedding, filename -- cgit v1.2.3 From f4e14642173a04723200b131deb417c6c79cab17 Mon Sep 17 00:00:00 2001 From: timntorres Date: Tue, 25 Oct 2022 00:04:25 -0700 Subject: Implement PR #3625 but for embeddings. --- 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 4921bd01..4fcebe74 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -358,7 +358,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc 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) + 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 -- cgit v1.2.3 From 737eb28faca8be2bb996ee0930ec77d1f7ebd939 Mon Sep 17 00:00:00 2001 From: DepFA <35278260+dfaker@users.noreply.github.com> Date: Wed, 26 Oct 2022 14:45:33 +0100 Subject: typo: cmd_opts.embedding_dir to cmd_opts.embeddings_dir --- 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 4fcebe74..ff002d3e 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -380,7 +380,7 @@ Last saved image: {html.escape(last_saved_image)}
embedding.cached_checksum = None # Before saving for the last time, change name back to base name (as opposed to the save_embedding_every step-suffixed naming convention). embedding.name = embedding_name - filename = os.path.join(shared.cmd_opts.embedding_dir, f'{embedding.name}.pt') + filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding.name}.pt') embedding.save(filename) return embedding, filename -- cgit v1.2.3 From 9ceef81f77ecce89f0c8f412c4d849210d852e82 Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Fri, 28 Oct 2022 20:48:08 +0700 Subject: Fix log off by 1 --- modules/textual_inversion/textual_inversion.py | 24 ++++++++++++------------ 1 file changed, 12 insertions(+), 12 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 ff002d3e..17dfb223 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -184,9 +184,8 @@ def write_loss(log_directory, filename, step, epoch_len, values): if shared.opts.training_write_csv_every == 0: return - if step % shared.opts.training_write_csv_every != 0: + if (step + 1) % shared.opts.training_write_csv_every != 0: return - write_csv_header = False if os.path.exists(os.path.join(log_directory, filename)) else True with open(os.path.join(log_directory, filename), "a+", newline='') as fout: @@ -196,11 +195,11 @@ def write_loss(log_directory, filename, step, epoch_len, values): csv_writer.writeheader() epoch = step // epoch_len - epoch_step = step - epoch * epoch_len + epoch_step = step % epoch_len csv_writer.writerow({ "step": step + 1, - "epoch": epoch + 1, + "epoch": epoch, "epoch_step": epoch_step + 1, **values, }) @@ -282,15 +281,16 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc loss.backward() optimizer.step() + steps_done = embedding.step + 1 epoch_num = embedding.step // len(ds) - epoch_step = embedding.step - (epoch_num * len(ds)) + 1 + epoch_step = embedding.step % len(ds) - pbar.set_description(f"[Epoch {epoch_num}: {epoch_step}/{len(ds)}]loss: {losses.mean():.7f}") + pbar.set_description(f"[Epoch {epoch_num}: {epoch_step+1}/{len(ds)}]loss: {losses.mean():.7f}") - if embedding.step > 0 and embedding_dir is not None and embedding.step % save_embedding_every == 0: + if embedding_dir is not None and steps_done % save_embedding_every == 0: # Before saving, change name to match current checkpoint. - embedding.name = f'{embedding_name}-{embedding.step}' + embedding.name = f'{embedding_name}-{steps_done}' last_saved_file = os.path.join(embedding_dir, f'{embedding.name}.pt') embedding.save(last_saved_file) embedding_yet_to_be_embedded = True @@ -300,8 +300,8 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc "learn_rate": scheduler.learn_rate }) - if embedding.step > 0 and images_dir is not None and embedding.step % create_image_every == 0: - forced_filename = f'{embedding_name}-{embedding.step}' + 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) p = processing.StableDiffusionProcessingTxt2Img( sd_model=shared.sd_model, @@ -334,7 +334,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc 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}-{embedding.step}.png') + 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) @@ -350,7 +350,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc checkpoint = sd_models.select_checkpoint() footer_left = checkpoint.model_name footer_mid = '[{}]'.format(checkpoint.hash) - footer_right = '{}v {}s'.format(vectorSize, embedding.step) + 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) -- cgit v1.2.3 From ab27c111d06ec920791c73eea25ad9a61671852e Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Sat, 29 Oct 2022 18:09:17 +0700 Subject: Add input validations before loading dataset for training --- modules/textual_inversion/textual_inversion.py | 48 +++++++++++++++++++------- 1 file changed, 36 insertions(+), 12 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 17dfb223..44f06443 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -204,9 +204,30 @@ def write_loss(log_directory, filename, step, epoch_len, values): **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"): + 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 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" + assert template_file, "Prompt template file is empty" + assert os.path.isfile(template_file), "Prompt template file doesn't exist" + assert steps, "Max steps is empty or 0" + assert isinstance(steps, int), "Max steps must be integer" + assert steps > 0 , "Max steps must be positive" + assert isinstance(save_model_every, int), "Save {name} must be integer" + assert save_model_every >= 0 , "Save {name} must be positive or 0" + assert isinstance(create_image_every, int), "Create image must be integer" + assert create_image_every >= 0 , "Create image must be positive or 0" + 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): - assert embedding_name, 'embedding not selected' + 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") shared.state.textinfo = "Initializing textual inversion training..." shared.state.job_count = steps @@ -232,17 +253,27 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc os.makedirs(images_embeds_dir, exist_ok=True) 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] + + ititial_step = embedding.step or 0 + if ititial_step > steps: + shared.state.textinfo = f"Model has already been trained beyond specified max steps" + return embedding, filename + + scheduler = LearnRateScheduler(learn_rate, steps, ititial_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=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file, batch_size=batch_size) - hijack = sd_hijack.model_hijack - - embedding = hijack.embedding_db.word_embeddings[embedding_name] embedding.vec.requires_grad = True + optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate) losses = torch.zeros((32,)) @@ -251,13 +282,6 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc forced_filename = "" embedding_yet_to_be_embedded = False - ititial_step = embedding.step or 0 - if ititial_step > steps: - return embedding, filename - - scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) - optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate) - pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) for i, entries in pbar: embedding.step = i + ititial_step -- cgit v1.2.3 From 3ce2bfdf95bd5f26d0f6e250e67338ada91980d1 Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Sat, 29 Oct 2022 19:43:21 +0700 Subject: Add cleanup after training --- modules/textual_inversion/textual_inversion.py | 185 +++++++++++++------------ 1 file changed, 95 insertions(+), 90 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 44f06443..fd7f0897 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -283,111 +283,113 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc 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 = f'{embedding_name}-{steps_done}' - last_saved_file = os.path.join(embedding_dir, f'{embedding.name}.pt') - embedding.save(last_saved_file) - 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) - 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_index = preview_sampler_index - 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 + try: + 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 = f'{embedding_name}-{steps_done}' + last_saved_file = os.path.join(embedding_dir, f'{embedding.name}.pt') + embedding.save(last_saved_file) + 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) + 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_index = preview_sampler_index + 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 + preview_text = p.prompt - processed = processing.process_images(p) - image = processed.images[0] + processed = processing.process_images(p) + image = processed.images[0] - shared.state.current_image = image + shared.state.current_image = image - if save_image_with_stored_embedding and os.path.exists(last_saved_file) and embedding_yet_to_be_embedded: + 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') + 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)) + info = PngImagePlugin.PngInfo() + data = torch.load(last_saved_file) + info.add_text("sd-ti-embedding", embedding_to_b64(data)) - title = "<{}>".format(data.get('name', '???')) + title = "<{}>".format(data.get('name', '???')) - try: - vectorSize = list(data['string_to_param'].values())[0].shape[0] - except Exception as e: - vectorSize = '?' + 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) + 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 = 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 + 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}" + 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.job_no = embedding.step - shared.state.textinfo = f""" + shared.state.textinfo = f"""

Loss: {losses.mean():.7f}
Step: {embedding.step}
@@ -396,6 +398,9 @@ Last saved embedding: {html.escape(last_saved_file)}
Last saved image: {html.escape(last_saved_image)}

""" + finally: + if embedding and embedding.vec is not None: + embedding.vec.requires_grad = False checkpoint = sd_models.select_checkpoint() -- cgit v1.2.3 From ab05a74ead9fabb45dd099990e34061c7eb02ca3 Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Sun, 30 Oct 2022 00:32:02 +0700 Subject: Revert "Add cleanup after training" This reverts commit 3ce2bfdf95bd5f26d0f6e250e67338ada91980d1. --- modules/textual_inversion/textual_inversion.py | 185 ++++++++++++------------- 1 file changed, 90 insertions(+), 95 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 fd7f0897..44f06443 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -283,113 +283,111 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc 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 - try: - 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 = f'{embedding_name}-{steps_done}' - last_saved_file = os.path.join(embedding_dir, f'{embedding.name}.pt') - embedding.save(last_saved_file) - 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) - 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_index = preview_sampler_index - 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 + 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 = f'{embedding_name}-{steps_done}' + last_saved_file = os.path.join(embedding_dir, f'{embedding.name}.pt') + embedding.save(last_saved_file) + 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) + 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_index = preview_sampler_index + 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 + preview_text = p.prompt - processed = processing.process_images(p) - image = processed.images[0] + processed = processing.process_images(p) + image = processed.images[0] - shared.state.current_image = image + shared.state.current_image = image - if save_image_with_stored_embedding and os.path.exists(last_saved_file) and embedding_yet_to_be_embedded: + 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') + 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)) + info = PngImagePlugin.PngInfo() + data = torch.load(last_saved_file) + info.add_text("sd-ti-embedding", embedding_to_b64(data)) - title = "<{}>".format(data.get('name', '???')) + title = "<{}>".format(data.get('name', '???')) - try: - vectorSize = list(data['string_to_param'].values())[0].shape[0] - except Exception as e: - vectorSize = '?' + 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) + 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 = 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 + 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}" + 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.job_no = embedding.step - shared.state.textinfo = f""" + shared.state.textinfo = f"""

Loss: {losses.mean():.7f}
Step: {embedding.step}
@@ -398,9 +396,6 @@ Last saved embedding: {html.escape(last_saved_file)}
Last saved image: {html.escape(last_saved_image)}

""" - finally: - if embedding and embedding.vec is not None: - embedding.vec.requires_grad = False checkpoint = sd_models.select_checkpoint() -- cgit v1.2.3 From a07f054c86f33360ff620d6a3fffdee366ab2d99 Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Sun, 30 Oct 2022 00:49:29 +0700 Subject: Add missing info on hypernetwork/embedding model log Mentioned here: https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/1528#discussioncomment-3991513 Also group the saving into one --- modules/textual_inversion/textual_inversion.py | 39 +++++++++++++++++--------- 1 file changed, 26 insertions(+), 13 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 44f06443..ee9917ce 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -119,7 +119,7 @@ class EmbeddingDatabase: vec = emb.detach().to(devices.device, dtype=torch.float32) embedding = Embedding(vec, name) embedding.step = data.get('step', None) - embedding.sd_checkpoint = data.get('hash', 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) @@ -259,6 +259,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc 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: @@ -314,9 +315,9 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc if embedding_dir is not None and steps_done % save_embedding_every == 0: # Before saving, change name to match current checkpoint. - embedding.name = f'{embedding_name}-{steps_done}' - last_saved_file = os.path.join(embedding_dir, f'{embedding.name}.pt') - embedding.save(last_saved_file) + 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), { @@ -397,14 +398,26 @@ Last saved image: {html.escape(last_saved_image)}

""" - checkpoint = sd_models.select_checkpoint() - - embedding.sd_checkpoint = checkpoint.hash - embedding.sd_checkpoint_name = checkpoint.model_name - embedding.cached_checksum = None - # Before saving for the last time, change name back to base name (as opposed to the save_embedding_every step-suffixed naming convention). - embedding.name = embedding_name - filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding.name}.pt') - embedding.save(filename) + filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt') + save_embedding(embedding, checkpoint, embedding_name, filename, remove_cached_checksum=True) return embedding, filename + +def save_embedding(embedding, 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 + old_cached_checksum = embedding.cached_checksum if hasattr(embedding, "cached_checksum") else None + try: + embedding.sd_checkpoint = checkpoint.hash + embedding.sd_checkpoint_name = checkpoint.model_name + if remove_cached_checksum: + embedding.cached_checksum = None + embedding.name = embedding_name + embedding.save(filename) + except: + embedding.sd_checkpoint = old_sd_checkpoint + embedding.sd_checkpoint_name = old_sd_checkpoint_name + embedding.name = old_embedding_name + embedding.cached_checksum = old_cached_checksum + raise -- cgit v1.2.3 From 3d58510f214c645ce5cdb261aa47df6573b239e9 Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Sun, 30 Oct 2022 00:54:59 +0700 Subject: Fix dataset still being loaded even when training will be skipped --- 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 ee9917ce..e0babb46 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -262,7 +262,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc checkpoint = sd_models.select_checkpoint() ititial_step = embedding.step or 0 - if ititial_step > steps: + if ititial_step >= steps: shared.state.textinfo = f"Model has already been trained beyond specified max steps" return embedding, filename -- cgit v1.2.3 From 006756f9cd6258eae418e9209cfc13f940ec53e1 Mon Sep 17 00:00:00 2001 From: Fampai <> Date: Mon, 31 Oct 2022 07:26:08 -0400 Subject: Added TI training optimizations option to use xattention optimizations when training option to unload vae when training --- modules/textual_inversion/textual_inversion.py | 9 +++++++++ 1 file changed, 9 insertions(+) (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 17dfb223..b0a1d26b 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -214,6 +214,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt') log_directory = os.path.join(log_directory, datetime.datetime.now().strftime("%Y-%m-%d"), embedding_name) + unload = shared.opts.unload_models_when_training if save_embedding_every > 0: embedding_dir = os.path.join(log_directory, "embeddings") @@ -238,6 +239,8 @@ 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) + if unload: + shared.sd_model.first_stage_model.to(devices.cpu) hijack = sd_hijack.model_hijack @@ -303,6 +306,9 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc 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, @@ -330,6 +336,9 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc 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: -- cgit v1.2.3 From 890e68aaf75ae80d5eb2fa95b4bf1adf78b96881 Mon Sep 17 00:00:00 2001 From: Fampai <> Date: Mon, 31 Oct 2022 10:07:12 -0400 Subject: Fixed minor bug when unloading vae during TI training, generating images after training will error out --- modules/textual_inversion/textual_inversion.py | 1 + 1 file changed, 1 insertion(+) (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 54a734f1..0aeb0459 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -409,6 +409,7 @@ Last saved image: {html.escape(last_saved_image)}
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) return embedding, filename -- cgit v1.2.3 From cffc240a7327ae60671ff533469fc4ed4bf605de Mon Sep 17 00:00:00 2001 From: Nerogar 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 = "" 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/textual_inversion/textual_inversion.py | 5 +++++ 1 file changed, 5 insertions(+) (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..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)}
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/textual_inversion/textual_inversion.py | 4 ++-- 1 file changed, 2 insertions(+), 2 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 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 -- cgit v1.2.3 From bd68e35de3b7cf7547ed97d8bdf60147402133cc Mon Sep 17 00:00:00 2001 From: flamelaw Date: Sun, 20 Nov 2022 12:35:26 +0900 Subject: Gradient accumulation, autocast fix, new latent sampling method, etc --- modules/textual_inversion/textual_inversion.py | 320 ++++++++++++++----------- 1 file changed, 183 insertions(+), 137 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 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 = "" last_saved_image = "" forced_filename = "" 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"""

-Loss: {losses.mean():.7f}
+Loss: {loss_step:.7f}
Step: {embedding.step}
-Last prompt: {html.escape(entries[0].cond_text)}
+Last prompt: {html.escape(batch.cond_text[0])}
Last saved embedding: {html.escape(last_saved_file)}
Last saved image: {html.escape(last_saved_image)}

""" - - 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 -- cgit v1.2.3 From 5b57f61ba47f8b11d19a5b46e7fb5a52458abae5 Mon Sep 17 00:00:00 2001 From: flamelaw Date: Mon, 21 Nov 2022 10:15:46 +0900 Subject: fix pin_memory with different latent sampling method --- modules/textual_inversion/textual_inversion.py | 7 +------ 1 file changed, 1 insertion(+), 6 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 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 Date: Wed, 23 Nov 2022 02:49:01 +0900 Subject: small fixes --- 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 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"""

Loss: {loss_step:.7f}
-Step: {embedding.step}
+Step: {steps_done}
Last prompt: {html.escape(batch.cond_text[0])}
Last saved embedding: {html.escape(last_saved_file)}
Last saved image: {html.escape(last_saved_image)}
-- 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 --- modules/textual_inversion/textual_inversion.py | 7 +++---- 1 file changed, 3 insertions(+), 4 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 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): -- cgit v1.2.3 From 755df94b2aa62eabd96f900e0dd7ddc83c2f692c Mon Sep 17 00:00:00 2001 From: flamelaw 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 Date: Mon, 28 Nov 2022 21:36:35 -0500 Subject: Use devices.autocast instead of torch.autocast --- 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 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 Date: Wed, 14 Dec 2022 21:01:32 -0500 Subject: Fix various typos --- modules/textual_inversion/textual_inversion.py | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 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 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 = "" last_saved_image = "" forced_filename = "" 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 -- cgit v1.2.3 From 3bf5591efe9a9f219c6088be322a87adc4f48f95 Mon Sep 17 00:00:00 2001 From: Yuval Aboulafia Date: Sat, 24 Dec 2022 21:35:29 +0200 Subject: fix F541 f-string without any placeholders --- 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 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) -- cgit v1.2.3 From f55ac33d446185680604e872ceda2ae858821d5c Mon Sep 17 00:00:00 2001 From: Vladimir Mandic Date: Sat, 31 Dec 2022 11:27:02 -0500 Subject: validate textual inversion embeddings --- modules/textual_inversion/textual_inversion.py | 43 +++++++++++++++++++++++--- 1 file changed, 38 insertions(+), 5 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 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] -- 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/textual_inversion/textual_inversion.py | 3 +-- 1 file changed, 1 insertion(+), 2 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 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 = "" last_saved_image = "" forced_filename = "" -- cgit v1.2.3 From c65909ad16a1962129114c6251de092f49479b06 Mon Sep 17 00:00:00 2001 From: Philpax Date: Mon, 2 Jan 2023 12:21:22 +1100 Subject: feat(api): return more data for embeddings --- modules/textual_inversion/textual_inversion.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 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 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 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/textual_inversion/textual_inversion.py | 40 ++++++++++++++++++++------ 1 file changed, 32 insertions(+), 8 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 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)}

""" 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)}
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 Date: Tue, 3 Jan 2023 10:34:51 -0500 Subject: add job info to modules --- modules/textual_inversion/textual_inversion.py | 1 + 1 file changed, 1 insertion(+) (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 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/textual_inversion/textual_inversion.py | 33 +++++++------------------- 1 file changed, 9 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 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 = "" 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