From 29e74d6e71826da9a3fe3c5790fed1329fc4d1e8 Mon Sep 17 00:00:00 2001 From: Melan Date: Thu, 20 Oct 2022 16:26:16 +0200 Subject: Add support for Tensorboard for training embeddings --- modules/textual_inversion/textual_inversion.py | 31 +++++++++++++++++++++++++- 1 file changed, 30 insertions(+), 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 3be69562..c57d3ace 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 import csv +import numpy as np +import torchvision.transforms from PIL import Image, PngImagePlugin - +from torch.utils.tensorboard import SummaryWriter from modules import shared, devices, sd_hijack, processing, sd_models import modules.textual_inversion.dataset from modules.textual_inversion.learn_schedule import LearnRateScheduler @@ -199,6 +201,19 @@ def write_loss(log_directory, filename, step, epoch_len, values): **values, }) +def tensorboard_add_scaler(tensorboard_writer, tag, value, step): + if shared.opts.training_enable_tensorboard: + tensorboard_writer.add_scalar(tag=tag, + scalar_value=value, global_step=step) + +def tensorboard_add_image(tensorboard_writer, tag, pil_image, step): + if shared.opts.training_enable_tensorboard: + # Convert a pil image to a torch tensor + img_tensor = torch.as_tensor(np.array(pil_image, copy=True)) + img_tensor = img_tensor.view(pil_image.size[1], pil_image.size[0], len(pil_image.getbands())) + img_tensor = img_tensor.permute((2, 0, 1)) + + tensorboard_writer.add_image(tag, img_tensor, global_step=step) 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' @@ -252,6 +267,12 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate) + if shared.opts.training_enable_tensorboard: + os.makedirs(os.path.join(log_directory, "tensorboard"), exist_ok=True) + tensorboard_writer = SummaryWriter( + log_dir=os.path.join(log_directory, "tensorboard"), + flush_secs=shared.opts.training_tensorboard_flush_every) + pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) for i, entries in pbar: embedding.step = i + ititial_step @@ -270,6 +291,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc del x losses[embedding.step % losses.shape[0]] = loss.item() + optimizer.zero_grad() loss.backward() @@ -285,6 +307,12 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc embedding.save(last_saved_file) embedding_yet_to_be_embedded = True + if shared.opts.training_enable_tensorboard: + tensorboard_add_scaler(tensorboard_writer, "Loss/train", losses.mean(), embedding.step) + tensorboard_add_scaler(tensorboard_writer, f"Loss/train/epoch-{epoch_num}", losses.mean(), epoch_step) + tensorboard_add_scaler(tensorboard_writer, "Learn rate/train", scheduler.learn_rate, embedding.step) + tensorboard_add_scaler(tensorboard_writer, f"Learn rate/train/epoch-{epoch_num}", scheduler.learn_rate, epoch_step) + write_loss(log_directory, "textual_inversion_loss.csv", embedding.step, len(ds), { "loss": f"{losses.mean():.7f}", "learn_rate": scheduler.learn_rate @@ -349,6 +377,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc embedding_yet_to_be_embedded = False image.save(last_saved_image) + tensorboard_add_image(tensorboard_writer, f"Validation at epoch {epoch_num}", image, embedding.step) last_saved_image += f", prompt: {preview_text}" -- cgit v1.2.3 From a6d593a6b51dc6a8443f2aa5c24caa391a04cd56 Mon Sep 17 00:00:00 2001 From: Melan Date: Thu, 20 Oct 2022 19:43:21 +0200 Subject: Fixed a typo in a variable --- modules/textual_inversion/textual_inversion.py | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index c57d3ace..ec8176bf 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -260,11 +260,11 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc last_saved_image = "" embedding_yet_to_be_embedded = False - ititial_step = embedding.step or 0 - if ititial_step > steps: + initial_step = embedding.step or 0 + if initial_step > steps: return embedding, filename - scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) + scheduler = LearnRateScheduler(learn_rate, steps, initial_step) optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate) if shared.opts.training_enable_tensorboard: @@ -273,9 +273,9 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc log_dir=os.path.join(log_directory, "tensorboard"), flush_secs=shared.opts.training_tensorboard_flush_every) - pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) + pbar = tqdm.tqdm(enumerate(ds), total=steps-initial_step) for i, entries in pbar: - embedding.step = i + ititial_step + embedding.step = i + initial_step scheduler.apply(optimizer, embedding.step) if scheduler.finished: -- cgit v1.2.3 From 8f5912984794c4c69e429c4636e984854d911b6a Mon Sep 17 00:00:00 2001 From: Melan Date: Thu, 20 Oct 2022 22:37:16 +0200 Subject: Some changes to the tensorboard code and hypernetwork support --- modules/textual_inversion/textual_inversion.py | 45 +++++++++++++++----------- 1 file changed, 27 insertions(+), 18 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index ec8176bf..b1dc2596 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -201,19 +201,30 @@ def write_loss(log_directory, filename, step, epoch_len, values): **values, }) +def tensorboard_setup(log_directory): + os.makedirs(os.path.join(log_directory, "tensorboard"), exist_ok=True) + return SummaryWriter( + log_dir=os.path.join(log_directory, "tensorboard"), + flush_secs=shared.opts.training_tensorboard_flush_every) + +def tensorboard_add(tensorboard_writer, loss, global_step, step, learn_rate, epoch_num): + tensorboard_add_scaler(tensorboard_writer, "Loss/train", loss, global_step) + tensorboard_add_scaler(tensorboard_writer, f"Loss/train/epoch-{epoch_num}", loss, step) + tensorboard_add_scaler(tensorboard_writer, "Learn rate/train", learn_rate, global_step) + tensorboard_add_scaler(tensorboard_writer, f"Learn rate/train/epoch-{epoch_num}", learn_rate, step) + def tensorboard_add_scaler(tensorboard_writer, tag, value, step): - if shared.opts.training_enable_tensorboard: - tensorboard_writer.add_scalar(tag=tag, - scalar_value=value, global_step=step) + tensorboard_writer.add_scalar(tag=tag, + scalar_value=value, global_step=step) def tensorboard_add_image(tensorboard_writer, tag, pil_image, step): - if shared.opts.training_enable_tensorboard: - # Convert a pil image to a torch tensor - img_tensor = torch.as_tensor(np.array(pil_image, copy=True)) - img_tensor = img_tensor.view(pil_image.size[1], pil_image.size[0], len(pil_image.getbands())) - img_tensor = img_tensor.permute((2, 0, 1)) + # Convert a pil image to a torch tensor + img_tensor = torch.as_tensor(np.array(pil_image, copy=True)) + img_tensor = img_tensor.view(pil_image.size[1], pil_image.size[0], + len(pil_image.getbands())) + img_tensor = img_tensor.permute((2, 0, 1)) - tensorboard_writer.add_image(tag, img_tensor, global_step=step) + tensorboard_writer.add_image(tag, img_tensor, global_step=step) 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' @@ -268,10 +279,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate) if shared.opts.training_enable_tensorboard: - os.makedirs(os.path.join(log_directory, "tensorboard"), exist_ok=True) - tensorboard_writer = SummaryWriter( - log_dir=os.path.join(log_directory, "tensorboard"), - flush_secs=shared.opts.training_tensorboard_flush_every) + tensorboard_writer = tensorboard_setup(log_directory) pbar = tqdm.tqdm(enumerate(ds), total=steps-initial_step) for i, entries in pbar: @@ -308,10 +316,8 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc embedding_yet_to_be_embedded = True if shared.opts.training_enable_tensorboard: - tensorboard_add_scaler(tensorboard_writer, "Loss/train", losses.mean(), embedding.step) - tensorboard_add_scaler(tensorboard_writer, f"Loss/train/epoch-{epoch_num}", losses.mean(), epoch_step) - tensorboard_add_scaler(tensorboard_writer, "Learn rate/train", scheduler.learn_rate, embedding.step) - tensorboard_add_scaler(tensorboard_writer, f"Learn rate/train/epoch-{epoch_num}", scheduler.learn_rate, epoch_step) + tensorboard_add(tensorboard_writer, loss=losses.mean(), global_step=embedding.step, + step=epoch_step, learn_rate=scheduler.learn_rate, epoch_num=epoch_num) write_loss(log_directory, "textual_inversion_loss.csv", embedding.step, len(ds), { "loss": f"{losses.mean():.7f}", @@ -377,7 +383,10 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc embedding_yet_to_be_embedded = False image.save(last_saved_image) - tensorboard_add_image(tensorboard_writer, f"Validation at epoch {epoch_num}", image, embedding.step) + + if shared.opts.training_enable_tensorboard and shared.opts.training_tensorboard_save_images: + tensorboard_add_image(tensorboard_writer, f"Validation at epoch {epoch_num}", + image, embedding.step) last_saved_image += f", prompt: {preview_text}" -- cgit v1.2.3 From 18f86e41f6f289042c075bff1498e620ab997b8c Mon Sep 17 00:00:00 2001 From: Melan Date: Mon, 24 Oct 2022 17:21:18 +0200 Subject: Removed two unused imports --- modules/textual_inversion/textual_inversion.py | 1 - 1 file changed, 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index b1dc2596..589314fe 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -9,7 +9,6 @@ import datetime import csv import numpy as np -import torchvision.transforms from PIL import Image, PngImagePlugin from torch.utils.tensorboard import SummaryWriter from modules import shared, devices, sd_hijack, processing, sd_models -- cgit v1.2.3 From 1618df41bad092e068c61bf510b1e20856821ad5 Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Fri, 28 Oct 2022 10:31:27 +0700 Subject: Gradient clipping for textual embedding --- modules/textual_inversion/textual_inversion.py | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index ff002d3e..7bad73a6 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -206,7 +206,7 @@ def write_loss(log_directory, filename, step, epoch_len, values): }) -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, data_root, log_directory, training_width, training_height, steps, clip_grad_mode, clip_grad_value, 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..." @@ -256,6 +256,9 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc if ititial_step > steps: return embedding, filename + clip_grad_mode_value = clip_grad_mode == "value" + clip_grad_mode_norm = clip_grad_mode == "norm" + scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate) @@ -280,6 +283,12 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc optimizer.zero_grad() loss.backward() + + if clip_grad_mode_value: + torch.nn.utils.clip_grad_value_(embedding.vec, clip_value=clip_grad_value) + elif clip_grad_mode_norm: + torch.nn.utils.clip_grad_norm_(embedding.vec, max_norm=clip_grad_value) + optimizer.step() -- cgit v1.2.3 From 16451ca573220e49f2eaaab97580b6b91287c8c4 Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Fri, 28 Oct 2022 17:16:23 +0700 Subject: Learning rate sched syntax support for grad clipping --- modules/textual_inversion/learn_schedule.py | 11 ++++++++--- modules/textual_inversion/textual_inversion.py | 12 +++++++++--- 2 files changed, 17 insertions(+), 6 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/learn_schedule.py b/modules/textual_inversion/learn_schedule.py index 2062726a..ffec3e1b 100644 --- a/modules/textual_inversion/learn_schedule.py +++ b/modules/textual_inversion/learn_schedule.py @@ -51,14 +51,19 @@ class LearnRateScheduler: self.finished = False - def apply(self, optimizer, step_number): + def step(self, step_number): if step_number <= self.end_step: - return + return False try: (self.learn_rate, self.end_step) = next(self.schedules) - except Exception: + except StopIteration: self.finished = True + return False + return True + + def apply(self, optimizer, step_number): + if not self.step(step_number): return if self.verbose: diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 7bad73a6..6b00c6a1 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -255,9 +255,12 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc ititial_step = embedding.step or 0 if ititial_step > steps: return embedding, filename - + clip_grad_mode_value = clip_grad_mode == "value" clip_grad_mode_norm = clip_grad_mode == "norm" + clip_grad_enabled = clip_grad_mode_value or clip_grad_mode_norm + if clip_grad_enabled: + clip_grad_sched = LearnRateScheduler(clip_grad_value, steps, ititial_step, verbose=False) scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate) @@ -273,6 +276,9 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc if shared.state.interrupted: break + if clip_grad_enabled: + clip_grad_sched.step(embedding.step) + 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) @@ -285,9 +291,9 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc loss.backward() if clip_grad_mode_value: - torch.nn.utils.clip_grad_value_(embedding.vec, clip_value=clip_grad_value) + torch.nn.utils.clip_grad_value_(embedding.vec, clip_value=clip_grad_sched.learn_rate) elif clip_grad_mode_norm: - torch.nn.utils.clip_grad_norm_(embedding.vec, max_norm=clip_grad_value) + torch.nn.utils.clip_grad_norm_(embedding.vec, max_norm=clip_grad_sched.learn_rate) optimizer.step() -- 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') 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 bb832d7725187f8a8ab44faa6ee1b38cb5f600aa Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Sat, 5 Nov 2022 11:48:38 +0700 Subject: Simplify grad clip --- modules/textual_inversion/textual_inversion.py | 16 +++++++--------- 1 file changed, 7 insertions(+), 9 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index c567ec3f..687d97bb 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -269,10 +269,10 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) - clip_grad_mode_value = clip_grad_mode == "value" - clip_grad_mode_norm = clip_grad_mode == "norm" - clip_grad_enabled = clip_grad_mode_value or clip_grad_mode_norm - if clip_grad_enabled: + clip_grad = torch.nn.utils.clip_grad_value_ if clip_grad_mode == "value" else \ + torch.nn.utils.clip_grad_norm_ if clip_grad_mode == "norm" else \ + None + if clip_grad: clip_grad_sched = LearnRateScheduler(clip_grad_value, steps, ititial_step, verbose=False) # 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)}..." @@ -302,7 +302,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc if shared.state.interrupted: break - if clip_grad_enabled: + if clip_grad: clip_grad_sched.step(embedding.step) with torch.autocast("cuda"): @@ -316,10 +316,8 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc optimizer.zero_grad() loss.backward() - if clip_grad_mode_value: - torch.nn.utils.clip_grad_value_(embedding.vec, clip_value=clip_grad_sched.learn_rate) - elif clip_grad_mode_norm: - torch.nn.utils.clip_grad_norm_(embedding.vec, max_norm=clip_grad_sched.learn_rate) + if clip_grad: + clip_grad(embedding.vec, clip_grad_sched.learn_rate) optimizer.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') 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') 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') 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') 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') 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/preprocess.py | 1 + modules/textual_inversion/textual_inversion.py | 1 + 2 files changed, 2 insertions(+) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/preprocess.py b/modules/textual_inversion/preprocess.py index 56b9b2eb..feb876c6 100644 --- a/modules/textual_inversion/preprocess.py +++ b/modules/textual_inversion/preprocess.py @@ -124,6 +124,7 @@ def preprocess_work(process_src, process_dst, process_width, process_height, pre files = listfiles(src) + shared.state.job = "preprocess" shared.state.textinfo = "Preprocessing..." shared.state.job_count = len(files) diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index fd253477..2c1251d6 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -245,6 +245,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ create_image_every = create_image_every or 0 validate_train_inputs(embedding_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding") + shared.state.job = "train-embedding" shared.state.textinfo = "Initializing textual inversion training..." shared.state.job_count = steps -- cgit v1.2.3 From 184e670126f5fc50ba56fa0fedcf0cf60e45ed7e Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Wed, 4 Jan 2023 17:45:01 +0300 Subject: fix the merge --- modules/textual_inversion/textual_inversion.py | 14 +++++--------- 1 file changed, 5 insertions(+), 9 deletions(-) (limited to 'modules/textual_inversion') 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') 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 From eea8fc40e16664ddc8a9aec77206da704a35dde0 Mon Sep 17 00:00:00 2001 From: timntorres Date: Thu, 5 Jan 2023 07:24:22 -0800 Subject: Add option to save ti settings to file. --- modules/textual_inversion/textual_inversion.py | 30 +++++++++++++++++++++++--- 1 file changed, 27 insertions(+), 3 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 71e07bcc..2bed2ecb 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -1,6 +1,7 @@ import os import sys import traceback +import inspect import torch import tqdm @@ -229,6 +230,28 @@ def write_loss(log_directory, filename, step, epoch_len, values): **values, }) +def save_settings_to_file(initial_step, num_of_dataset_images, embedding_name, vectors_per_token, 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): + checkpoint = sd_models.select_checkpoint() + model_name = checkpoint.model_name + model_hash = '[{}]'.format(checkpoint.hash) + + # Get a list of the argument names. + arg_names = inspect.getfullargspec(save_settings_to_file).args + + # Create a list of the argument names to include in the settings string. + names = arg_names[:16] # Include all arguments up until the preview-related ones. + if preview_from_txt2img: + names.extend(arg_names[16:]) # Include all remaining arguments if `preview_from_txt2img` is True. + + # Build the settings string. + settings_str = "datetime : " + datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S") + "\n" + for name in names: + value = locals()[name] + settings_str += f"{name}: {value}\n" + + with open(os.path.join(log_directory, 'settings.txt'), "a+") as fout: + fout.write(settings_str + "\n\n") + 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" @@ -292,13 +315,13 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ if initial_step >= steps: shared.state.textinfo = "Model has already been trained beyond specified max steps" return embedding, filename + scheduler = LearnRateScheduler(learn_rate, steps, initial_step) - clip_grad = torch.nn.utils.clip_grad_value_ if clip_grad_mode == "value" else \ torch.nn.utils.clip_grad_norm_ if clip_grad_mode == "norm" else \ None if clip_grad: - clip_grad_sched = LearnRateScheduler(clip_grad_value, steps, ititial_step, verbose=False) + clip_grad_sched = LearnRateScheduler(clip_grad_value, steps, initial_step, verbose=False) # 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 @@ -306,7 +329,8 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ 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) - + if shared.opts.save_train_settings_to_txt: + save_settings_to_file(initial_step , len(ds) , embedding_name, len(embedding.vec) , 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) latent_sampling_method = ds.latent_sampling_method dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, latent_sampling_method=latent_sampling_method, batch_size=ds.batch_size, pin_memory=pin_memory) -- cgit v1.2.3 From b85c2b5cf4a6809bc871718cf4680d49c3e95e94 Mon Sep 17 00:00:00 2001 From: timntorres Date: Thu, 5 Jan 2023 08:14:38 -0800 Subject: Clean up ti, add same behavior to hypernetwork. --- modules/textual_inversion/textual_inversion.py | 14 +++++++++----- 1 file changed, 9 insertions(+), 5 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 2bed2ecb..68648550 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -230,18 +230,20 @@ def write_loss(log_directory, filename, step, epoch_len, values): **values, }) +# Note: hypernetwork.py has a nearly identical function of the same name. def save_settings_to_file(initial_step, num_of_dataset_images, embedding_name, vectors_per_token, 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): checkpoint = sd_models.select_checkpoint() model_name = checkpoint.model_name model_hash = '[{}]'.format(checkpoint.hash) - + # Starting index of preview-related arguments. + border_index = 16 # Get a list of the argument names. arg_names = inspect.getfullargspec(save_settings_to_file).args # Create a list of the argument names to include in the settings string. - names = arg_names[:16] # Include all arguments up until the preview-related ones. + names = arg_names[:border_index] # Include all arguments up until the preview-related ones. if preview_from_txt2img: - names.extend(arg_names[16:]) # Include all remaining arguments if `preview_from_txt2img` is True. + names.extend(arg_names[border_index:]) # Include all remaining arguments if `preview_from_txt2img` is True. # Build the settings string. settings_str = "datetime : " + datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S") + "\n" @@ -329,8 +331,10 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ 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) - if shared.opts.save_train_settings_to_txt: - save_settings_to_file(initial_step , len(ds) , embedding_name, len(embedding.vec) , 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) + + if shared.opts.save_training_settings_to_txt: + save_settings_to_file(initial_step, len(ds), embedding_name, len(embedding.vec), 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) + latent_sampling_method = ds.latent_sampling_method dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, latent_sampling_method=latent_sampling_method, batch_size=ds.batch_size, pin_memory=pin_memory) -- cgit v1.2.3 From b6bab2f052b32c0ffebe6aecc1819ccf20cf8c5d Mon Sep 17 00:00:00 2001 From: timntorres Date: Thu, 5 Jan 2023 09:14:56 -0800 Subject: Include model in log file. Exclude directory. --- modules/textual_inversion/textual_inversion.py | 22 +++++++++------------- 1 file changed, 9 insertions(+), 13 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 68648550..ce7e4f5d 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -231,26 +231,22 @@ def write_loss(log_directory, filename, step, epoch_len, values): }) # Note: hypernetwork.py has a nearly identical function of the same name. -def save_settings_to_file(initial_step, num_of_dataset_images, embedding_name, vectors_per_token, 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): - checkpoint = sd_models.select_checkpoint() - model_name = checkpoint.model_name - model_hash = '[{}]'.format(checkpoint.hash) +def save_settings_to_file(model_name, model_hash, initial_step, num_of_dataset_images, embedding_name, vectors_per_token, 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): # Starting index of preview-related arguments. - border_index = 16 + border_index = 18 # Get a list of the argument names. - arg_names = inspect.getfullargspec(save_settings_to_file).args - + arg_names = inspect.getfullargspec(save_settings_to_file).args # Create a list of the argument names to include in the settings string. names = arg_names[:border_index] # Include all arguments up until the preview-related ones. if preview_from_txt2img: - names.extend(arg_names[border_index:]) # Include all remaining arguments if `preview_from_txt2img` is True. - + names.extend(arg_names[border_index:]) # Include preview-related arguments if applicable. # Build the settings string. settings_str = "datetime : " + datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S") + "\n" for name in names: - value = locals()[name] - settings_str += f"{name}: {value}\n" - + if name != 'log_directory': # It's useless and redundant to save log_directory. + value = locals()[name] + settings_str += f"{name}: {value}\n" + # Create or append to the file. with open(os.path.join(log_directory, 'settings.txt'), "a+") as fout: fout.write(settings_str + "\n\n") @@ -333,7 +329,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ 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) if shared.opts.save_training_settings_to_txt: - save_settings_to_file(initial_step, len(ds), embedding_name, len(embedding.vec), 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_settings_to_file(checkpoint.model_name, '[{}]'.format(checkpoint.hash), initial_step, len(ds), embedding_name, len(embedding.vec), 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) latent_sampling_method = ds.latent_sampling_method -- cgit v1.2.3 From fda04e620d529031e2134520e74756d0efa30464 Mon Sep 17 00:00:00 2001 From: Kuma <36082288+KumiIT@users.noreply.github.com> Date: Thu, 5 Jan 2023 18:44:19 +0100 Subject: typo in TI --- modules/textual_inversion/textual_inversion.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 71e07bcc..24b43045 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -298,7 +298,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ torch.nn.utils.clip_grad_norm_ if clip_grad_mode == "norm" else \ None if clip_grad: - clip_grad_sched = LearnRateScheduler(clip_grad_value, steps, ititial_step, verbose=False) + clip_grad_sched = LearnRateScheduler(clip_grad_value, steps, initial_step, verbose=False) # 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 -- cgit v1.2.3 From 81133d4168ae0bae9bf8bf1a1d4983319a589112 Mon Sep 17 00:00:00 2001 From: Faber Date: Fri, 6 Jan 2023 03:38:37 +0700 Subject: allow loading embeddings from subdirectories --- modules/textual_inversion/textual_inversion.py | 23 ++++++++++++----------- 1 file changed, 12 insertions(+), 11 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 24b43045..0a059044 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -149,19 +149,20 @@ class EmbeddingDatabase: else: self.skipped_embeddings[name] = embedding - for fn in os.listdir(self.embeddings_dir): - try: - fullfn = os.path.join(self.embeddings_dir, fn) - - if os.stat(fullfn).st_size == 0: + for root, dirs, fns in os.walk(self.embeddings_dir): + for fn in fns: + try: + fullfn = os.path.join(root, fn) + + if os.stat(fullfn).st_size == 0: + continue + + process_file(fullfn, fn) + except Exception: + print(f"Error loading embedding {fn}:", file=sys.stderr) + print(traceback.format_exc(), file=sys.stderr) continue - process_file(fullfn, fn) - except Exception: - print(f"Error loading embedding {fn}:", file=sys.stderr) - print(traceback.format_exc(), file=sys.stderr) - continue - 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.keys())}") -- cgit v1.2.3 From 683287d87f6401083a8d63eedc00ca7410214ca1 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 6 Jan 2023 08:52:06 +0300 Subject: rework saving training params to file #6372 --- modules/textual_inversion/logging.py | 24 ++++++++++++++++++++++++ modules/textual_inversion/textual_inversion.py | 23 +++-------------------- 2 files changed, 27 insertions(+), 20 deletions(-) create mode 100644 modules/textual_inversion/logging.py (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/logging.py b/modules/textual_inversion/logging.py new file mode 100644 index 00000000..8b1981d5 --- /dev/null +++ b/modules/textual_inversion/logging.py @@ -0,0 +1,24 @@ +import datetime +import json +import os + +saved_params_shared = {"model_name", "model_hash", "initial_step", "num_of_dataset_images", "learn_rate", "batch_size", "data_root", "log_directory", "training_width", "training_height", "steps", "create_image_every", "template_file"} +saved_params_ti = {"embedding_name", "num_vectors_per_token", "save_embedding_every", "save_image_with_stored_embedding"} +saved_params_hypernet = {"hypernetwork_name", "layer_structure", "activation_func", "weight_init", "add_layer_norm", "use_dropout", "save_hypernetwork_every"} +saved_params_all = saved_params_shared | saved_params_ti | saved_params_hypernet +saved_params_previews = {"preview_prompt", "preview_negative_prompt", "preview_steps", "preview_sampler_index", "preview_cfg_scale", "preview_seed", "preview_width", "preview_height"} + + +def save_settings_to_file(log_directory, all_params): + now = datetime.datetime.now() + params = {"datetime": now.strftime("%Y-%m-%d %H:%M:%S")} + + keys = saved_params_all + if all_params.get('preview_from_txt2img'): + keys = keys | saved_params_previews + + params.update({k: v for k, v in all_params.items() if k in keys}) + + filename = f'settings-{now.strftime("%Y-%m-%d-%H-%M-%S")}.json' + with open(os.path.join(log_directory, filename), "w") as file: + json.dump(params, file, indent=4) diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index e9cf432f..f9f5e8cd 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -18,6 +18,8 @@ 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, caption_image_overlay) +from modules.textual_inversion.logging import save_settings_to_file + class Embedding: def __init__(self, vec, name, step=None): @@ -231,25 +233,6 @@ def write_loss(log_directory, filename, step, epoch_len, values): **values, }) -# Note: hypernetwork.py has a nearly identical function of the same name. -def save_settings_to_file(model_name, model_hash, initial_step, num_of_dataset_images, embedding_name, vectors_per_token, 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): - # Starting index of preview-related arguments. - border_index = 18 - # Get a list of the argument names. - arg_names = inspect.getfullargspec(save_settings_to_file).args - # Create a list of the argument names to include in the settings string. - names = arg_names[:border_index] # Include all arguments up until the preview-related ones. - if preview_from_txt2img: - names.extend(arg_names[border_index:]) # Include preview-related arguments if applicable. - # Build the settings string. - settings_str = "datetime : " + datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S") + "\n" - for name in names: - if name != 'log_directory': # It's useless and redundant to save log_directory. - value = locals()[name] - settings_str += f"{name}: {value}\n" - # Create or append to the file. - with open(os.path.join(log_directory, 'settings.txt'), "a+") as fout: - fout.write(settings_str + "\n\n") 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" @@ -330,7 +313,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ 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) if shared.opts.save_training_settings_to_txt: - save_settings_to_file(checkpoint.model_name, '[{}]'.format(checkpoint.hash), initial_step, len(ds), embedding_name, len(embedding.vec), 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_settings_to_file(log_directory, {**dict(model_name=checkpoint.model_name, model_hash=checkpoint.hash, num_of_dataset_images=len(ds), num_vectors_per_token=len(embedding.vec)), **locals()}) latent_sampling_method = ds.latent_sampling_method -- cgit v1.2.3 From 79e39fae6110c20a3ee6255e2841c877f65e8cbd Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sat, 7 Jan 2023 01:45:28 +0300 Subject: CLIP hijack rework --- modules/textual_inversion/textual_inversion.py | 1 - 1 file changed, 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index f9f5e8cd..45882ed6 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -79,7 +79,6 @@ class EmbeddingDatabase: self.word_embeddings[embedding.name] = embedding - # 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] -- cgit v1.2.3 From 448b9cedab66e05b5b2800513ca334a769b42aa7 Mon Sep 17 00:00:00 2001 From: dan Date: Sat, 7 Jan 2023 21:07:27 +0800 Subject: Allow variable img size --- modules/textual_inversion/dataset.py | 18 +++++++++++------- modules/textual_inversion/textual_inversion.py | 4 ++-- 2 files changed, 13 insertions(+), 9 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index 88d68c76..375178ed 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -17,7 +17,7 @@ re_numbers_at_start = re.compile(r"^[-\d]+\s*") class DatasetEntry: - def __init__(self, filename=None, filename_text=None, latent_dist=None, latent_sample=None, cond=None, cond_text=None, pixel_values=None): + def __init__(self, filename=None, filename_text=None, latent_dist=None, latent_sample=None, cond=None, cond_text=None, pixel_values=None, img_shape=None): self.filename = filename self.filename_text = filename_text self.latent_dist = latent_dist @@ -25,6 +25,7 @@ class DatasetEntry: self.cond = cond self.cond_text = cond_text self.pixel_values = pixel_values + self.img_shape = img_shape class PersonalizedBase(Dataset): @@ -33,8 +34,6 @@ class PersonalizedBase(Dataset): self.placeholder_token = placeholder_token - self.width = width - self.height = height self.flip = transforms.RandomHorizontalFlip(p=flip_p) self.dataset = [] @@ -59,7 +58,11 @@ class PersonalizedBase(Dataset): if shared.state.interrupted: raise Exception("interrupted") try: - image = Image.open(path).convert('RGB').resize((self.width, self.height), PIL.Image.BICUBIC) + image = Image.open(path).convert('RGB') + if width < 2000: + image = image.resize((width, height), PIL.Image.BICUBIC) + else: + assert batch_size == 1, 'variable img size must have batch size 1' except Exception: continue @@ -88,14 +91,14 @@ class PersonalizedBase(Dataset): if latent_sampling_method == "once" or (latent_sampling_method == "deterministic" and not isinstance(latent_dist, DiagonalGaussianDistribution)): latent_sample = model.get_first_stage_encoding(latent_dist).squeeze().to(devices.cpu) latent_sampling_method = "once" - entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample) + entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample, img_shape=image.size) elif latent_sampling_method == "deterministic": # Works only for DiagonalGaussianDistribution latent_dist.std = 0 latent_sample = model.get_first_stage_encoding(latent_dist).squeeze().to(devices.cpu) - entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample) + entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample, img_shape=image.size) elif latent_sampling_method == "random": - entry = DatasetEntry(filename=path, filename_text=filename_text, latent_dist=latent_dist) + entry = DatasetEntry(filename=path, filename_text=filename_text, latent_dist=latent_dist, img_shape=image.size) if not (self.tag_drop_out != 0 or self.shuffle_tags): entry.cond_text = self.create_text(filename_text) @@ -151,6 +154,7 @@ class BatchLoader: self.cond_text = [entry.cond_text for entry in data] self.cond = [entry.cond for entry in data] self.latent_sample = torch.stack([entry.latent_sample for entry in data]).squeeze(1) + self.img_shape = [entry.img_shape for entry in data] #self.emb_index = [entry.emb_index for entry in data] #print(self.latent_sample.device) diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 45882ed6..9f96d0fd 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -451,8 +451,8 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ else: p.prompt = batch.cond_text[0] p.steps = 20 - p.width = training_width - p.height = training_height + p.width = batch.img_shape[0][0] + p.height = batch.img_shape[0][1] preview_text = p.prompt -- cgit v1.2.3 From 669fb18d5222f53ae48abe0f30393d846c50ad91 Mon Sep 17 00:00:00 2001 From: dan Date: Sun, 8 Jan 2023 01:34:52 +0800 Subject: Add checkbox for variable training dims --- modules/textual_inversion/dataset.py | 4 ++-- modules/textual_inversion/textual_inversion.py | 4 ++-- 2 files changed, 4 insertions(+), 4 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index 375178ed..7f8a314f 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -29,7 +29,7 @@ class DatasetEntry: class PersonalizedBase(Dataset): - def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, cond_model=None, device=None, template_file=None, include_cond=False, batch_size=1, gradient_step=1, shuffle_tags=False, tag_drop_out=0, latent_sampling_method='once'): + def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, cond_model=None, device=None, template_file=None, include_cond=False, batch_size=1, gradient_step=1, shuffle_tags=False, tag_drop_out=0, latent_sampling_method='once', varsize=False): re_word = re.compile(shared.opts.dataset_filename_word_regex) if len(shared.opts.dataset_filename_word_regex) > 0 else None self.placeholder_token = placeholder_token @@ -59,7 +59,7 @@ class PersonalizedBase(Dataset): raise Exception("interrupted") try: image = Image.open(path).convert('RGB') - if width < 2000: + if not varsize: image = image.resize((width, height), PIL.Image.BICUBIC) else: assert batch_size == 1, 'variable img size must have batch size 1' diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 9f96d0fd..110efd19 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -255,7 +255,7 @@ 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 train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, steps, clip_grad_mode, clip_grad_value, 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): +def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, varsize, steps, clip_grad_mode, clip_grad_value, 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, gradient_step, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding") @@ -309,7 +309,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ 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) + 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, varsize=varsize) if shared.opts.save_training_settings_to_txt: save_settings_to_file(log_directory, {**dict(model_name=checkpoint.model_name, model_hash=checkpoint.hash, num_of_dataset_images=len(ds), num_vectors_per_token=len(embedding.vec)), **locals()}) -- cgit v1.2.3 From 72497895b9b1948f86d9309fe897cbb70c20ba7e Mon Sep 17 00:00:00 2001 From: dan Date: Sun, 8 Jan 2023 01:36:00 +0800 Subject: Move batchsize check --- modules/textual_inversion/dataset.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index 7f8a314f..bcad6848 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -46,6 +46,8 @@ class PersonalizedBase(Dataset): assert data_root, 'dataset directory not specified' assert os.path.isdir(data_root), "Dataset directory doesn't exist" assert os.listdir(data_root), "Dataset directory is empty" + if varsize: + assert batch_size == 1, 'variable img size must have batch size 1' self.image_paths = [os.path.join(data_root, file_path) for file_path in os.listdir(data_root)] @@ -61,8 +63,6 @@ class PersonalizedBase(Dataset): image = Image.open(path).convert('RGB') if not varsize: image = image.resize((width, height), PIL.Image.BICUBIC) - else: - assert batch_size == 1, 'variable img size must have batch size 1' except Exception: continue -- cgit v1.2.3 From a0c87f1fdf2b76b2ae4ef6c4b01ddaede3afab06 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sun, 8 Jan 2023 08:52:26 +0300 Subject: skip images in embeddings dir if they have a second .preview extension --- modules/textual_inversion/textual_inversion.py | 4 ++++ 1 file changed, 4 insertions(+) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 45882ed6..e85dd549 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -109,6 +109,10 @@ class EmbeddingDatabase: ext = ext.upper() if ext in ['.PNG', '.WEBP', '.JXL', '.AVIF']: + _, second_ext = os.path.splitext(name) + if second_ext.upper() == '.PREVIEW': + return + 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']) -- cgit v1.2.3 From 085427de0efc9e9e7a6e9a5aebc6b5a69f0365e7 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sun, 8 Jan 2023 09:37:33 +0300 Subject: make it possible for extensions/scripts to add their own embedding directories --- modules/textual_inversion/textual_inversion.py | 170 +++++++++++++++---------- 1 file changed, 104 insertions(+), 66 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index e85dd549..217fe9eb 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -66,17 +66,41 @@ class Embedding: return self.cached_checksum +class DirWithTextualInversionEmbeddings: + def __init__(self, path): + self.path = path + self.mtime = None + + def has_changed(self): + if not os.path.isdir(self.path): + return False + + mt = os.path.getmtime(self.path) + if self.mtime is None or mt > self.mtime: + return True + + def update(self): + if not os.path.isdir(self.path): + return + + self.mtime = os.path.getmtime(self.path) + + class EmbeddingDatabase: - def __init__(self, embeddings_dir): + def __init__(self): self.ids_lookup = {} self.word_embeddings = {} self.skipped_embeddings = {} - self.dir_mtime = None - self.embeddings_dir = embeddings_dir self.expected_shape = -1 + self.embedding_dirs = {} - def register_embedding(self, embedding, model): + def add_embedding_dir(self, path): + self.embedding_dirs[path] = DirWithTextualInversionEmbeddings(path) + + def clear_embedding_dirs(self): + self.embedding_dirs.clear() + def register_embedding(self, embedding, model): self.word_embeddings[embedding.name] = embedding ids = model.cond_stage_model.tokenize([embedding.name])[0] @@ -93,69 +117,62 @@ class EmbeddingDatabase: 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) - if not force_reload and self.dir_mtime is not None and mt <= self.dir_mtime: - return + def load_from_file(self, path, filename): + name, ext = os.path.splitext(filename) + ext = ext.upper() - self.dir_mtime = mt - self.ids_lookup.clear() - self.word_embeddings.clear() - self.skipped_embeddings.clear() - self.expected_shape = self.get_expected_shape() - - def process_file(path, filename): - name, ext = os.path.splitext(filename) - ext = ext.upper() - - if ext in ['.PNG', '.WEBP', '.JXL', '.AVIF']: - _, second_ext = os.path.splitext(name) - if second_ext.upper() == '.PREVIEW': - return - - 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']) - name = data.get('name', name) - else: - data = extract_image_data_embed(embed_image) - name = data.get('name', name) - elif ext in ['.BIN', '.PT']: - data = torch.load(path, map_location="cpu") - else: + if ext in ['.PNG', '.WEBP', '.JXL', '.AVIF']: + _, second_ext = os.path.splitext(name) + if second_ext.upper() == '.PREVIEW': return - # 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) - embedding.sd_checkpoint = data.get('sd_checkpoint', None) - embedding.sd_checkpoint_name = data.get('sd_checkpoint_name', None) - 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) + 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']) + name = data.get('name', name) else: - self.skipped_embeddings[name] = embedding + data = extract_image_data_embed(embed_image) + name = data.get('name', name) + elif ext in ['.BIN', '.PT']: + data = torch.load(path, map_location="cpu") + else: + return + + # 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) + embedding.sd_checkpoint = data.get('sd_checkpoint', None) + embedding.sd_checkpoint_name = data.get('sd_checkpoint_name', None) + 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[name] = embedding - for root, dirs, fns in os.walk(self.embeddings_dir): + def load_from_dir(self, embdir): + if not os.path.isdir(embdir.path): + return + + for root, dirs, fns in os.walk(embdir.path): for fn in fns: try: fullfn = os.path.join(root, fn) @@ -163,12 +180,32 @@ class EmbeddingDatabase: if os.stat(fullfn).st_size == 0: continue - process_file(fullfn, fn) + self.load_from_file(fullfn, fn) except Exception: print(f"Error loading embedding {fn}:", file=sys.stderr) print(traceback.format_exc(), file=sys.stderr) continue + def load_textual_inversion_embeddings(self, force_reload=False): + if not force_reload: + need_reload = False + for path, embdir in self.embedding_dirs.items(): + if embdir.has_changed(): + need_reload = True + break + + if not need_reload: + return + + self.ids_lookup.clear() + self.word_embeddings.clear() + self.skipped_embeddings.clear() + self.expected_shape = self.get_expected_shape() + + for path, embdir in self.embedding_dirs.items(): + self.load_from_dir(embdir) + embdir.update() + 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.keys())}") @@ -251,14 +288,15 @@ def validate_train_inputs(model_name, learn_rate, batch_size, gradient_step, dat 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 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 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" + 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, gradient_step, data_root, log_directory, training_width, training_height, steps, clip_grad_mode, clip_grad_value, 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 -- cgit v1.2.3 From 43bb5190fc9e7ae479a5dc6640be202c9a71e464 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Mon, 9 Jan 2023 22:52:23 +0300 Subject: remove/simplify some changes from #6481 --- modules/textual_inversion/dataset.py | 14 +++++--------- modules/textual_inversion/textual_inversion.py | 4 ++-- 2 files changed, 7 insertions(+), 11 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index bcad6848..fa48708e 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -17,7 +17,7 @@ re_numbers_at_start = re.compile(r"^[-\d]+\s*") class DatasetEntry: - def __init__(self, filename=None, filename_text=None, latent_dist=None, latent_sample=None, cond=None, cond_text=None, pixel_values=None, img_shape=None): + def __init__(self, filename=None, filename_text=None, latent_dist=None, latent_sample=None, cond=None, cond_text=None, pixel_values=None): self.filename = filename self.filename_text = filename_text self.latent_dist = latent_dist @@ -25,7 +25,6 @@ class DatasetEntry: self.cond = cond self.cond_text = cond_text self.pixel_values = pixel_values - self.img_shape = img_shape class PersonalizedBase(Dataset): @@ -46,12 +45,10 @@ class PersonalizedBase(Dataset): assert data_root, 'dataset directory not specified' assert os.path.isdir(data_root), "Dataset directory doesn't exist" assert os.listdir(data_root), "Dataset directory is empty" - if varsize: - assert batch_size == 1, 'variable img size must have batch size 1' + assert batch_size == 1 or not varsize, 'variable img size must have batch size 1' self.image_paths = [os.path.join(data_root, file_path) for file_path in os.listdir(data_root)] - self.shuffle_tags = shuffle_tags self.tag_drop_out = tag_drop_out @@ -91,14 +88,14 @@ class PersonalizedBase(Dataset): if latent_sampling_method == "once" or (latent_sampling_method == "deterministic" and not isinstance(latent_dist, DiagonalGaussianDistribution)): latent_sample = model.get_first_stage_encoding(latent_dist).squeeze().to(devices.cpu) latent_sampling_method = "once" - entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample, img_shape=image.size) + entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample) elif latent_sampling_method == "deterministic": # Works only for DiagonalGaussianDistribution latent_dist.std = 0 latent_sample = model.get_first_stage_encoding(latent_dist).squeeze().to(devices.cpu) - entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample, img_shape=image.size) + entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample) elif latent_sampling_method == "random": - entry = DatasetEntry(filename=path, filename_text=filename_text, latent_dist=latent_dist, img_shape=image.size) + entry = DatasetEntry(filename=path, filename_text=filename_text, latent_dist=latent_dist) if not (self.tag_drop_out != 0 or self.shuffle_tags): entry.cond_text = self.create_text(filename_text) @@ -154,7 +151,6 @@ class BatchLoader: self.cond_text = [entry.cond_text for entry in data] self.cond = [entry.cond for entry in data] self.latent_sample = torch.stack([entry.latent_sample for entry in data]).squeeze(1) - self.img_shape = [entry.img_shape for entry in data] #self.emb_index = [entry.emb_index for entry in data] #print(self.latent_sample.device) diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index ad76297e..14be2c96 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -492,8 +492,8 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ else: p.prompt = batch.cond_text[0] p.steps = 20 - p.width = batch.img_shape[0][0] - p.height = batch.img_shape[0][1] + p.width = training_width + p.height = training_height preview_text = p.prompt -- cgit v1.2.3 From 1fbb6f9ebe48326a3b12ecf611105dbc4a46891e Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Mon, 9 Jan 2023 23:35:40 +0300 Subject: make a dropdown for prompt template selection --- modules/textual_inversion/textual_inversion.py | 35 ++++++++++++++++++++------ 1 file changed, 27 insertions(+), 8 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 14be2c96..5420903f 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -2,6 +2,7 @@ import os import sys import traceback import inspect +from collections import namedtuple import torch import tqdm @@ -15,12 +16,26 @@ from modules import shared, devices, sd_hijack, processing, sd_models, images, s import modules.textual_inversion.dataset 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, - 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 from modules.textual_inversion.logging import save_settings_to_file +TextualInversionTemplate = namedtuple("TextualInversionTemplate", ["name", "path"]) +textual_inversion_templates = {} + + +def list_textual_inversion_templates(): + textual_inversion_templates.clear() + + for root, dirs, fns in os.walk(shared.cmd_opts.textual_inversion_templates_dir): + for fn in fns: + path = os.path.join(root, fn) + + textual_inversion_templates[fn] = TextualInversionTemplate(fn, path) + + return textual_inversion_templates + + class Embedding: def __init__(self, vec, name, step=None): self.vec = vec @@ -274,7 +289,7 @@ def write_loss(log_directory, filename, step, epoch_len, values): }) -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"): +def validate_train_inputs(model_name, learn_rate, batch_size, gradient_step, data_root, template_file, template_filename, 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" @@ -284,8 +299,9 @@ def validate_train_inputs(model_name, learn_rate, batch_size, gradient_step, dat 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 template_filename, "Prompt template file not selected" + assert template_file, f"Prompt template file {template_filename} not found" + assert os.path.isfile(template_file.path), f"Prompt template file {template_filename} 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" @@ -296,10 +312,13 @@ 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 train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, varsize, steps, clip_grad_mode, clip_grad_value, 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): + +def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, varsize, steps, clip_grad_mode, clip_grad_value, shuffle_tags, tag_drop_out, latent_sampling_method, create_image_every, save_embedding_every, template_filename, 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, gradient_step, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding") + template_file = textual_inversion_templates.get(template_filename, None) + validate_train_inputs(embedding_name, learn_rate, batch_size, gradient_step, data_root, template_file, template_filename, steps, save_embedding_every, create_image_every, log_directory, name="embedding") + template_file = template_file.path shared.state.job = "train-embedding" shared.state.textinfo = "Initializing textual inversion training..." -- cgit v1.2.3 From 6be644fa04ce1542f3a01804310cbbc0a4a91620 Mon Sep 17 00:00:00 2001 From: dan Date: Wed, 11 Jan 2023 05:31:58 +0800 Subject: Enable batch_size>1 for mixed-sized training --- modules/textual_inversion/dataset.py | 36 ++++++++++++++++++++++++++++++++---- 1 file changed, 32 insertions(+), 4 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index fa48708e..b47414f3 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -3,8 +3,10 @@ import numpy as np import PIL import torch from PIL import Image -from torch.utils.data import Dataset, DataLoader +from torch.utils.data import Dataset, DataLoader, Sampler from torchvision import transforms +from collections import defaultdict +from random import shuffle, choices import random import tqdm @@ -45,12 +47,12 @@ class PersonalizedBase(Dataset): assert data_root, 'dataset directory not specified' assert os.path.isdir(data_root), "Dataset directory doesn't exist" assert os.listdir(data_root), "Dataset directory is empty" - assert batch_size == 1 or not varsize, 'variable img size must have batch size 1' self.image_paths = [os.path.join(data_root, file_path) for file_path in os.listdir(data_root)] self.shuffle_tags = shuffle_tags self.tag_drop_out = tag_drop_out + groups = defaultdict(list) print("Preparing dataset...") for path in tqdm.tqdm(self.image_paths): @@ -103,13 +105,14 @@ class PersonalizedBase(Dataset): if include_cond and not (self.tag_drop_out != 0 or self.shuffle_tags): with devices.autocast(): entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0) - + groups[image.size].append(len(self.dataset)) self.dataset.append(entry) del torchdata del latent_dist del latent_sample self.length = len(self.dataset) + self.groups = list(groups.values()) assert self.length > 0, "No images have been found in the dataset." self.batch_size = min(batch_size, self.length) self.gradient_step = min(gradient_step, self.length // self.batch_size) @@ -137,9 +140,34 @@ class PersonalizedBase(Dataset): entry.latent_sample = shared.sd_model.get_first_stage_encoding(entry.latent_dist).to(devices.cpu) return entry +class GroupedBatchSampler(Sampler): + def __init__(self, data_source: PersonalizedBase, batch_size: int): + n = len(data_source) + self.groups = data_source.groups + self.len = n_batch = n // batch_size + expected = [len(g) / n * n_batch * batch_size for g in data_source.groups] + self.base = [int(e) // batch_size for e in expected] + self.n_rand_batches = nrb = n_batch - sum(self.base) + self.probs = [e%batch_size/nrb/batch_size if nrb>0 else 0 for e in expected] + self.batch_size = batch_size + def __len__(self): + return self.len + def __iter__(self): + b = self.batch_size + for g in self.groups: + shuffle(g) + batches = [] + for g in self.groups: + batches.extend(g[i*b:(i+1)*b] for i in range(len(g) // b)) + for _ in range(self.n_rand_batches): + rand_group = choices(self.groups, self.probs)[0] + batches.append(choices(rand_group, k=b)) + shuffle(batches) + yield from batches + class PersonalizedDataLoader(DataLoader): def __init__(self, dataset, latent_sampling_method="once", batch_size=1, pin_memory=False): - super(PersonalizedDataLoader, self).__init__(dataset, shuffle=True, drop_last=True, batch_size=batch_size, pin_memory=pin_memory) + super(PersonalizedDataLoader, self).__init__(dataset, batch_sampler=GroupedBatchSampler(dataset, batch_size), pin_memory=pin_memory) if latent_sampling_method == "random": self.collate_fn = collate_wrapper_random else: -- cgit v1.2.3 From f9706acf431f77e0ce9e4270e5be7299922ee963 Mon Sep 17 00:00:00 2001 From: Lee Bousfield Date: Tue, 10 Jan 2023 18:40:34 -0700 Subject: Support loading textual inversion embeddings from safetensors files --- modules/textual_inversion/textual_inversion.py | 3 +++ 1 file changed, 3 insertions(+) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 5420903f..3866c154 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -9,6 +9,7 @@ import tqdm import html import datetime import csv +import safetensors.torch from PIL import Image, PngImagePlugin @@ -150,6 +151,8 @@ class EmbeddingDatabase: name = data.get('name', name) elif ext in ['.BIN', '.PT']: data = torch.load(path, map_location="cpu") + elif ext in ['.SAFETENSORS']: + data = safetensors.torch.load_file(path, device="cpu") else: return -- cgit v1.2.3 From 3f43d8a966ba8462ba019a5ad573f94508cd45f8 Mon Sep 17 00:00:00 2001 From: Vladimir Mandic Date: Wed, 11 Jan 2023 10:28:55 -0500 Subject: set descriptions --- modules/textual_inversion/preprocess.py | 7 ++++++- modules/textual_inversion/textual_inversion.py | 4 +++- 2 files changed, 9 insertions(+), 2 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/preprocess.py b/modules/textual_inversion/preprocess.py index feb876c6..3c1042ad 100644 --- a/modules/textual_inversion/preprocess.py +++ b/modules/textual_inversion/preprocess.py @@ -135,7 +135,8 @@ def preprocess_work(process_src, process_dst, process_width, process_height, pre params.process_caption_deepbooru = process_caption_deepbooru params.preprocess_txt_action = preprocess_txt_action - for index, imagefile in enumerate(tqdm.tqdm(files)): + pbar = tqdm.tqdm(files) + for index, imagefile in enumerate(pbar): params.subindex = 0 filename = os.path.join(src, imagefile) try: @@ -143,6 +144,10 @@ def preprocess_work(process_src, process_dst, process_width, process_height, pre except Exception: continue + description = f"Preprocessing [Image {index}/{len(files)}]" + pbar.set_description(description) + shared.state.textinfo = description + params.src = filename existing_caption = None diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 3866c154..b915b091 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -476,7 +476,9 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ 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}") + description = f"Training textual inversion [Epoch {epoch_num}: {epoch_step+1}/{steps_per_epoch}]loss: {loss_step:.7f}" + pbar.set_description(description) + shared.state.textinfo = description 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}' -- cgit v1.2.3 From d52a80f7f7da160c73afd067c8f1bf491391f994 Mon Sep 17 00:00:00 2001 From: Shondoit Date: Thu, 12 Jan 2023 09:22:29 +0100 Subject: Allow creation of zero vectors for TI --- modules/textual_inversion/textual_inversion.py | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index b915b091..853246a6 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -248,11 +248,14 @@ def create_embedding(name, num_vectors_per_token, overwrite_old, init_text='*'): with devices.autocast(): cond_model([""]) # will send cond model to GPU if lowvram/medvram is active - embedded = cond_model.encode_embedding_init_text(init_text, num_vectors_per_token) + #cond_model expects at least some text, so we provide '*' as backup. + embedded = cond_model.encode_embedding_init_text(init_text or '*', 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): - vec[i] = embedded[i * int(embedded.shape[0]) // num_vectors_per_token] + #Only copy if we provided an init_text, otherwise keep vectors as zeros + if 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 "._- ")) -- cgit v1.2.3 From 0b262802b86a55c4f71faf377f2cb1aee2960b63 Mon Sep 17 00:00:00 2001 From: Josh R Date: Thu, 12 Jan 2023 17:31:05 -0800 Subject: add gradient settings to training settings log files --- modules/textual_inversion/logging.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/logging.py b/modules/textual_inversion/logging.py index 8b1981d5..31e50b64 100644 --- a/modules/textual_inversion/logging.py +++ b/modules/textual_inversion/logging.py @@ -2,7 +2,7 @@ import datetime import json import os -saved_params_shared = {"model_name", "model_hash", "initial_step", "num_of_dataset_images", "learn_rate", "batch_size", "data_root", "log_directory", "training_width", "training_height", "steps", "create_image_every", "template_file"} +saved_params_shared = {"model_name", "model_hash", "initial_step", "num_of_dataset_images", "learn_rate", "batch_size", "clip_grad_mode", "clip_grad_value", "gradient_step", "data_root", "log_directory", "training_width", "training_height", "steps", "create_image_every", "template_file"} saved_params_ti = {"embedding_name", "num_vectors_per_token", "save_embedding_every", "save_image_with_stored_embedding"} saved_params_hypernet = {"hypernetwork_name", "layer_structure", "activation_func", "weight_init", "add_layer_norm", "use_dropout", "save_hypernetwork_every"} saved_params_all = saved_params_shared | saved_params_ti | saved_params_hypernet -- cgit v1.2.3 From a176d89487d92f5a5b152401e5c424b34ff43b96 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 13 Jan 2023 14:32:15 +0300 Subject: print bucket sizes for training without resizing images #6620 fix an error when generating a picture with embedding in it --- modules/textual_inversion/dataset.py | 16 ++++++++++++++++ modules/textual_inversion/image_embedding.py | 4 ++-- modules/textual_inversion/textual_inversion.py | 2 +- 3 files changed, 19 insertions(+), 3 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index b47414f3..d31963d4 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -118,6 +118,12 @@ class PersonalizedBase(Dataset): self.gradient_step = min(gradient_step, self.length // self.batch_size) self.latent_sampling_method = latent_sampling_method + if len(groups) > 1: + print("Buckets:") + for (w, h), ids in sorted(groups.items(), key=lambda x: x[0]): + print(f" {w}x{h}: {len(ids)}") + print() + def create_text(self, filename_text): text = random.choice(self.lines) tags = filename_text.split(',') @@ -140,8 +146,11 @@ class PersonalizedBase(Dataset): entry.latent_sample = shared.sd_model.get_first_stage_encoding(entry.latent_dist).to(devices.cpu) return entry + class GroupedBatchSampler(Sampler): def __init__(self, data_source: PersonalizedBase, batch_size: int): + super().__init__(data_source) + n = len(data_source) self.groups = data_source.groups self.len = n_batch = n // batch_size @@ -150,21 +159,28 @@ class GroupedBatchSampler(Sampler): self.n_rand_batches = nrb = n_batch - sum(self.base) self.probs = [e%batch_size/nrb/batch_size if nrb>0 else 0 for e in expected] self.batch_size = batch_size + def __len__(self): return self.len + def __iter__(self): b = self.batch_size + for g in self.groups: shuffle(g) + batches = [] for g in self.groups: batches.extend(g[i*b:(i+1)*b] for i in range(len(g) // b)) for _ in range(self.n_rand_batches): rand_group = choices(self.groups, self.probs)[0] batches.append(choices(rand_group, k=b)) + shuffle(batches) + yield from batches + class PersonalizedDataLoader(DataLoader): def __init__(self, dataset, latent_sampling_method="once", batch_size=1, pin_memory=False): super(PersonalizedDataLoader, self).__init__(dataset, batch_sampler=GroupedBatchSampler(dataset, batch_size), pin_memory=pin_memory) diff --git a/modules/textual_inversion/image_embedding.py b/modules/textual_inversion/image_embedding.py index ea653806..5593f88c 100644 --- a/modules/textual_inversion/image_embedding.py +++ b/modules/textual_inversion/image_embedding.py @@ -76,10 +76,10 @@ def insert_image_data_embed(image, data): next_size = data_np_low.shape[0] + (h-(data_np_low.shape[0] % h)) next_size = next_size + ((h*d)-(next_size % (h*d))) - data_np_low.resize(next_size) + data_np_low = np.resize(data_np_low, next_size) data_np_low = data_np_low.reshape((h, -1, d)) - data_np_high.resize(next_size) + data_np_high = np.resize(data_np_high, next_size) data_np_high = data_np_high.reshape((h, -1, d)) edge_style = list(data['string_to_param'].values())[0].cpu().detach().numpy().tolist()[0][:1024] diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 853246a6..e23906ca 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -479,7 +479,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ epoch_num = embedding.step // steps_per_epoch epoch_step = embedding.step % steps_per_epoch - description = f"Training textual inversion [Epoch {epoch_num}: {epoch_step+1}/{steps_per_epoch}]loss: {loss_step:.7f}" + description = f"Training textual inversion [Epoch {epoch_num}: {epoch_step+1}/{steps_per_epoch}] loss: {loss_step:.7f}" pbar.set_description(description) shared.state.textinfo = description if embedding_dir is not None and steps_done % save_embedding_every == 0: -- cgit v1.2.3 From 82725f0ac439f7e3b67858d55900e95330bbd326 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 13 Jan 2023 15:04:37 +0300 Subject: fix a bug caused by merge --- modules/textual_inversion/textual_inversion.py | 1 + 1 file changed, 1 insertion(+) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 85210b0e..6939efcc 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -11,6 +11,7 @@ import datetime import csv import safetensors.torch +import numpy as np from PIL import Image, PngImagePlugin from torch.utils.tensorboard import SummaryWriter -- cgit v1.2.3 From a95f1353089bdeaccd7c266b40cdd79efedfe632 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sat, 14 Jan 2023 09:56:59 +0300 Subject: change hash to sha256 --- modules/textual_inversion/textual_inversion.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 6939efcc..63935878 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -407,7 +407,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ 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, varsize=varsize) if shared.opts.save_training_settings_to_txt: - save_settings_to_file(log_directory, {**dict(model_name=checkpoint.model_name, model_hash=checkpoint.hash, num_of_dataset_images=len(ds), num_vectors_per_token=len(embedding.vec)), **locals()}) + save_settings_to_file(log_directory, {**dict(model_name=checkpoint.model_name, model_hash=checkpoint.shorthash, num_of_dataset_images=len(ds), num_vectors_per_token=len(embedding.vec)), **locals()}) latent_sampling_method = ds.latent_sampling_method @@ -584,7 +584,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ checkpoint = sd_models.select_checkpoint() footer_left = checkpoint.model_name - footer_mid = '[{}]'.format(checkpoint.hash) + footer_mid = '[{}]'.format(checkpoint.shorthash) footer_right = '{}v {}s'.format(vectorSize, steps_done) captioned_image = caption_image_overlay(image, title, footer_left, footer_mid, footer_right) @@ -626,7 +626,7 @@ def save_embedding(embedding, optimizer, checkpoint, embedding_name, filename, r 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 = checkpoint.shorthash embedding.sd_checkpoint_name = checkpoint.model_name if remove_cached_checksum: embedding.cached_checksum = None -- cgit v1.2.3 From d8b90ac121cbf0c18b1dc9d56a5e1d14ca51e74e Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sun, 15 Jan 2023 18:50:56 +0300 Subject: big rework of progressbar/preview system to allow multiple users to prompts at the same time and do not get previews of each other --- modules/textual_inversion/preprocess.py | 2 +- modules/textual_inversion/textual_inversion.py | 6 +++--- 2 files changed, 4 insertions(+), 4 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/preprocess.py b/modules/textual_inversion/preprocess.py index 3c1042ad..64abff4d 100644 --- a/modules/textual_inversion/preprocess.py +++ b/modules/textual_inversion/preprocess.py @@ -12,7 +12,7 @@ from modules.shared import opts, cmd_opts from modules.textual_inversion import autocrop -def preprocess(process_src, process_dst, process_width, process_height, preprocess_txt_action, process_flip, process_split, process_caption, process_caption_deepbooru=False, split_threshold=0.5, overlap_ratio=0.2, process_focal_crop=False, process_focal_crop_face_weight=0.9, process_focal_crop_entropy_weight=0.3, process_focal_crop_edges_weight=0.5, process_focal_crop_debug=False): +def preprocess(id_task, process_src, process_dst, process_width, process_height, preprocess_txt_action, process_flip, process_split, process_caption, process_caption_deepbooru=False, split_threshold=0.5, overlap_ratio=0.2, process_focal_crop=False, process_focal_crop_face_weight=0.9, process_focal_crop_entropy_weight=0.3, process_focal_crop_edges_weight=0.5, process_focal_crop_debug=False): try: if process_caption: shared.interrogator.load() diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 63935878..7e4a6d24 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -345,7 +345,7 @@ def validate_train_inputs(model_name, learn_rate, batch_size, gradient_step, dat assert log_directory, "Log directory is empty" -def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, varsize, steps, clip_grad_mode, clip_grad_value, shuffle_tags, tag_drop_out, latent_sampling_method, create_image_every, save_embedding_every, template_filename, 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(id_task, embedding_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, varsize, steps, clip_grad_mode, clip_grad_value, shuffle_tags, tag_drop_out, latent_sampling_method, create_image_every, save_embedding_every, template_filename, 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 template_file = textual_inversion_templates.get(template_filename, None) @@ -510,7 +510,6 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ description = f"Training textual inversion [Epoch {epoch_num}: {epoch_step+1}/{steps_per_epoch}] loss: {loss_step:.7f}" pbar.set_description(description) - shared.state.textinfo = description 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}' @@ -560,7 +559,8 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ shared.sd_model.first_stage_model.to(devices.cpu) if image is not None: - shared.state.current_image = image + shared.state.assign_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}" -- cgit v1.2.3 From 110d1a2d598bcfacffe3d524df1a3422b4cbd8ec Mon Sep 17 00:00:00 2001 From: Vladimir Mandic Date: Sun, 15 Jan 2023 12:41:00 -0500 Subject: add fields to settings file --- modules/textual_inversion/logging.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/logging.py b/modules/textual_inversion/logging.py index 31e50b64..734a4b6f 100644 --- a/modules/textual_inversion/logging.py +++ b/modules/textual_inversion/logging.py @@ -2,7 +2,7 @@ import datetime import json import os -saved_params_shared = {"model_name", "model_hash", "initial_step", "num_of_dataset_images", "learn_rate", "batch_size", "clip_grad_mode", "clip_grad_value", "gradient_step", "data_root", "log_directory", "training_width", "training_height", "steps", "create_image_every", "template_file"} +saved_params_shared = {"model_name", "model_hash", "initial_step", "num_of_dataset_images", "learn_rate", "batch_size", "clip_grad_mode", "clip_grad_value", "gradient_step", "data_root", "log_directory", "training_width", "training_height", "steps", "create_image_every", "template_file", "gradient_step", "latent_sampling_method"} saved_params_ti = {"embedding_name", "num_vectors_per_token", "save_embedding_every", "save_image_with_stored_embedding"} saved_params_hypernet = {"hypernetwork_name", "layer_structure", "activation_func", "weight_init", "add_layer_norm", "use_dropout", "save_hypernetwork_every"} saved_params_all = saved_params_shared | saved_params_ti | saved_params_hypernet -- cgit v1.2.3 From 4688bfff55dd6607e6608524fb219f97dc6fe8bb Mon Sep 17 00:00:00 2001 From: dan Date: Tue, 17 Jan 2023 17:16:43 +0800 Subject: Add auto-sized cropping UI --- modules/textual_inversion/preprocess.py | 38 ++++++++++++++++++++++++++++++--- 1 file changed, 35 insertions(+), 3 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/preprocess.py b/modules/textual_inversion/preprocess.py index 64abff4d..86c1cd33 100644 --- a/modules/textual_inversion/preprocess.py +++ b/modules/textual_inversion/preprocess.py @@ -12,7 +12,7 @@ from modules.shared import opts, cmd_opts from modules.textual_inversion import autocrop -def preprocess(id_task, process_src, process_dst, process_width, process_height, preprocess_txt_action, process_flip, process_split, process_caption, process_caption_deepbooru=False, split_threshold=0.5, overlap_ratio=0.2, process_focal_crop=False, process_focal_crop_face_weight=0.9, process_focal_crop_entropy_weight=0.3, process_focal_crop_edges_weight=0.5, process_focal_crop_debug=False): +def preprocess(id_task, process_src, process_dst, process_width, process_height, preprocess_txt_action, process_flip, process_split, process_caption, process_caption_deepbooru=False, split_threshold=0.5, overlap_ratio=0.2, process_focal_crop=False, process_focal_crop_face_weight=0.9, process_focal_crop_entropy_weight=0.3, process_focal_crop_edges_weight=0.5, process_focal_crop_debug=False, process_multicrop=None, process_multicrop_mindim=None, process_multicrop_maxdim=None, process_multicrop_minarea=None, process_multicrop_maxarea=None, process_multicrop_objective=None, process_multicrop_threshold=None): try: if process_caption: shared.interrogator.load() @@ -20,7 +20,7 @@ def preprocess(id_task, process_src, process_dst, process_width, process_height, if process_caption_deepbooru: deepbooru.model.start() - preprocess_work(process_src, process_dst, process_width, process_height, preprocess_txt_action, process_flip, process_split, process_caption, process_caption_deepbooru, split_threshold, overlap_ratio, process_focal_crop, process_focal_crop_face_weight, process_focal_crop_entropy_weight, process_focal_crop_edges_weight, process_focal_crop_debug) + preprocess_work(process_src, process_dst, process_width, process_height, preprocess_txt_action, process_flip, process_split, process_caption, process_caption_deepbooru, split_threshold, overlap_ratio, process_focal_crop, process_focal_crop_face_weight, process_focal_crop_entropy_weight, process_focal_crop_edges_weight, process_focal_crop_debug, process_multicrop, process_multicrop_mindim, process_multicrop_maxdim, process_multicrop_minarea, process_multicrop_maxarea, process_multicrop_objective, process_multicrop_threshold) finally: @@ -109,8 +109,32 @@ def split_pic(image, inverse_xy, width, height, overlap_ratio): splitted = image.crop((0, y, to_w, y + to_h)) yield splitted +# not using torchvision.transforms.CenterCrop because it doesn't allow float regions +def center_crop(image: Image, w: int, h: int): + iw, ih = image.size + if ih / h < iw / w: + sw = w * ih / h + box = (iw - sw) / 2, 0, iw - (iw - sw) / 2, ih + else: + sh = h * iw / w + box = 0, (ih - sh) / 2, iw, ih - (ih - sh) / 2 + return image.resize((w, h), Image.Resampling.LANCZOS, box) + -def preprocess_work(process_src, process_dst, process_width, process_height, preprocess_txt_action, process_flip, process_split, process_caption, process_caption_deepbooru=False, split_threshold=0.5, overlap_ratio=0.2, process_focal_crop=False, process_focal_crop_face_weight=0.9, process_focal_crop_entropy_weight=0.3, process_focal_crop_edges_weight=0.5, process_focal_crop_debug=False): +def multicrop_pic(image: Image, mindim, maxdim, minarea, maxarea, objective, threshold): + iw, ih = image.size + err = lambda w, h: 1-(lambda x: x if x < 1 else 1/x)(iw/ih/(w/h)) + try: + w, h = max(((w, h) for w in range(mindim, maxdim+1, 64) for h in range(mindim, maxdim+1, 64) + if minarea <= w * h <= maxarea and err(w, h) <= threshold), + key= lambda wh: ((objective=='Maximize area')*wh[0]*wh[1], -err(*wh)) + ) + except ValueError: + return + return center_crop(image, w, h) + + +def preprocess_work(process_src, process_dst, process_width, process_height, preprocess_txt_action, process_flip, process_split, process_caption, process_caption_deepbooru=False, split_threshold=0.5, overlap_ratio=0.2, process_focal_crop=False, process_focal_crop_face_weight=0.9, process_focal_crop_entropy_weight=0.3, process_focal_crop_edges_weight=0.5, process_focal_crop_debug=False, process_multicrop=None, process_multicrop_mindim=None, process_multicrop_maxdim=None, process_multicrop_minarea=None, process_multicrop_maxarea=None, process_multicrop_objective=None, process_multicrop_threshold=None): width = process_width height = process_height src = os.path.abspath(process_src) @@ -194,6 +218,14 @@ def preprocess_work(process_src, process_dst, process_width, process_height, pre save_pic(focal, index, params, existing_caption=existing_caption) process_default_resize = False + if process_multicrop: + cropped = multicrop_pic(img, process_multicrop_mindim, process_multicrop_maxdim, process_multicrop_minarea, process_multicrop_maxarea, process_multicrop_objective, process_multicrop_threshold) + if cropped is not None: + save_pic(cropped, index, params, existing_caption=existing_caption) + else: + print(f"skipped {img.width}x{img.height} image {filename} (can't find suitable size within error threshold)") + process_default_resize = False + if process_default_resize: img = images.resize_image(1, img, width, height) save_pic(img, index, params, existing_caption=existing_caption) -- cgit v1.2.3 From 924e222004ab54273806c5f2ca7a0e7cfa76ad83 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Wed, 18 Jan 2023 23:04:24 +0300 Subject: add option to show/hide warnings removed hiding warnings from LDSR fixed/reworked few places that produced warnings --- modules/textual_inversion/textual_inversion.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 7e4a6d24..5a7be422 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -15,7 +15,7 @@ import numpy as np from PIL import Image, PngImagePlugin from torch.utils.tensorboard import SummaryWriter -from modules import shared, devices, sd_hijack, processing, sd_models, images, sd_samplers +from modules import shared, devices, sd_hijack, processing, sd_models, images, sd_samplers, sd_hijack_checkpoint import modules.textual_inversion.dataset from modules.textual_inversion.learn_schedule import LearnRateScheduler @@ -452,6 +452,8 @@ def train_embedding(id_task, embedding_name, learn_rate, batch_size, gradient_st pbar = tqdm.tqdm(total=steps - initial_step) try: + sd_hijack_checkpoint.add() + for i in range((steps-initial_step) * gradient_step): if scheduler.finished: break @@ -617,9 +619,11 @@ Last saved image: {html.escape(last_saved_image)}
pbar.close() shared.sd_model.first_stage_model.to(devices.device) shared.parallel_processing_allowed = old_parallel_processing_allowed + sd_hijack_checkpoint.remove() return embedding, filename + 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 -- cgit v1.2.3 From 18a09c7e0032e2e655269e8e2b4f1ca6ed0cc7d3 Mon Sep 17 00:00:00 2001 From: dan Date: Thu, 19 Jan 2023 17:36:23 +0800 Subject: Simplification and bugfix --- modules/textual_inversion/preprocess.py | 12 +++++------- 1 file changed, 5 insertions(+), 7 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/preprocess.py b/modules/textual_inversion/preprocess.py index 86c1cd33..454dcc36 100644 --- a/modules/textual_inversion/preprocess.py +++ b/modules/textual_inversion/preprocess.py @@ -124,13 +124,11 @@ def center_crop(image: Image, w: int, h: int): def multicrop_pic(image: Image, mindim, maxdim, minarea, maxarea, objective, threshold): iw, ih = image.size err = lambda w, h: 1-(lambda x: x if x < 1 else 1/x)(iw/ih/(w/h)) - try: - w, h = max(((w, h) for w in range(mindim, maxdim+1, 64) for h in range(mindim, maxdim+1, 64) - if minarea <= w * h <= maxarea and err(w, h) <= threshold), - key= lambda wh: ((objective=='Maximize area')*wh[0]*wh[1], -err(*wh)) - ) - except ValueError: - return + w, h = max(((w, h) for w in range(mindim, maxdim+1, 64) for h in range(mindim, maxdim+1, 64) + if minarea <= w * h <= maxarea and err(w, h) <= threshold), + key= lambda wh: (wh[0]*wh[1], -err(*wh))[::1 if objective=='Maximize area' else -1], + default=None + ) return center_crop(image, w, h) -- cgit v1.2.3 From 2985b317d719f0f0580d2ff93f3008ccabb9c251 Mon Sep 17 00:00:00 2001 From: dan Date: Thu, 19 Jan 2023 17:39:30 +0800 Subject: Fix of fix --- modules/textual_inversion/preprocess.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/preprocess.py b/modules/textual_inversion/preprocess.py index 454dcc36..c0ac11d3 100644 --- a/modules/textual_inversion/preprocess.py +++ b/modules/textual_inversion/preprocess.py @@ -124,12 +124,12 @@ def center_crop(image: Image, w: int, h: int): def multicrop_pic(image: Image, mindim, maxdim, minarea, maxarea, objective, threshold): iw, ih = image.size err = lambda w, h: 1-(lambda x: x if x < 1 else 1/x)(iw/ih/(w/h)) - w, h = max(((w, h) for w in range(mindim, maxdim+1, 64) for h in range(mindim, maxdim+1, 64) + wh = max(((w, h) for w in range(mindim, maxdim+1, 64) for h in range(mindim, maxdim+1, 64) if minarea <= w * h <= maxarea and err(w, h) <= threshold), key= lambda wh: (wh[0]*wh[1], -err(*wh))[::1 if objective=='Maximize area' else -1], default=None ) - return center_crop(image, w, h) + return wh and center_crop(image, *wh) def preprocess_work(process_src, process_dst, process_width, process_height, preprocess_txt_action, process_flip, process_split, process_caption, process_caption_deepbooru=False, split_threshold=0.5, overlap_ratio=0.2, process_focal_crop=False, process_focal_crop_face_weight=0.9, process_focal_crop_entropy_weight=0.3, process_focal_crop_edges_weight=0.5, process_focal_crop_debug=False, process_multicrop=None, process_multicrop_mindim=None, process_multicrop_maxdim=None, process_multicrop_minarea=None, process_multicrop_maxarea=None, process_multicrop_objective=None, process_multicrop_threshold=None): -- cgit v1.2.3 From 40ff6db5325fc34ad4fa35e80cb1e7768d9f7e75 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sat, 21 Jan 2023 08:36:07 +0300 Subject: extra networks UI rework of hypernets: rather than via settings, hypernets are added directly to prompt as --- modules/textual_inversion/textual_inversion.py | 2 ++ 1 file changed, 2 insertions(+) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 5a7be422..4e90f690 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -50,6 +50,7 @@ class Embedding: self.sd_checkpoint = None self.sd_checkpoint_name = None self.optimizer_state_dict = None + self.filename = None def save(self, filename): embedding_data = { @@ -182,6 +183,7 @@ class EmbeddingDatabase: embedding.sd_checkpoint_name = data.get('sd_checkpoint_name', None) embedding.vectors = vec.shape[0] embedding.shape = vec.shape[-1] + embedding.filename = path if self.expected_shape == -1 or self.expected_shape == embedding.shape: self.register_embedding(embedding, shared.sd_model) -- cgit v1.2.3 From e179b6098ac1b1ce9645fef5bd9fd0bc9b918f30 Mon Sep 17 00:00:00 2001 From: "Alex \"mcmonkey\" Goodwin" Date: Wed, 25 Jan 2023 08:48:40 -0800 Subject: allow symlinks in the textual inversion embeddings folder --- modules/textual_inversion/textual_inversion.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 4e90f690..6cf00e65 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -194,7 +194,7 @@ class EmbeddingDatabase: if not os.path.isdir(embdir.path): return - for root, dirs, fns in os.walk(embdir.path): + for root, dirs, fns in os.walk(embdir.path, followlinks=True): for fn in fns: try: fullfn = os.path.join(root, fn) -- cgit v1.2.3 From 5eee2ac39863f9e44591b50d0710dd2615416a13 Mon Sep 17 00:00:00 2001 From: Max Audron Date: Wed, 25 Jan 2023 17:15:42 +0100 Subject: add data-dir flag and set all user data directories based on it --- modules/textual_inversion/preprocess.py | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/preprocess.py b/modules/textual_inversion/preprocess.py index c0ac11d3..2239cb84 100644 --- a/modules/textual_inversion/preprocess.py +++ b/modules/textual_inversion/preprocess.py @@ -6,8 +6,7 @@ import sys import tqdm import time -from modules import shared, images, deepbooru -from modules.paths import models_path +from modules import paths, shared, images, deepbooru from modules.shared import opts, cmd_opts from modules.textual_inversion import autocrop @@ -199,7 +198,7 @@ def preprocess_work(process_src, process_dst, process_width, process_height, pre dnn_model_path = None try: - dnn_model_path = autocrop.download_and_cache_models(os.path.join(models_path, "opencv")) + dnn_model_path = autocrop.download_and_cache_models(os.path.join(paths.models_path, "opencv")) except Exception as e: print("Unable to load face detection model for auto crop selection. Falling back to lower quality haar method.", e) -- cgit v1.2.3