From 467cae167a3066ffa2b2a5e6f16dd42642219aba Mon Sep 17 00:00:00 2001 From: TinkTheBoush Date: Tue, 1 Nov 2022 23:29:12 +0900 Subject: append_tag_shuffle --- modules/textual_inversion/dataset.py | 10 ++++++++-- modules/textual_inversion/textual_inversion.py | 4 ++-- 2 files changed, 10 insertions(+), 4 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index ad726577..e9d97cc1 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -24,7 +24,7 @@ class DatasetEntry: class PersonalizedBase(Dataset): - def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, device=None, template_file=None, include_cond=False, batch_size=1): + def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", shuffle_tags=True, model=None, device=None, template_file=None, include_cond=False, batch_size=1): 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 @@ -33,6 +33,7 @@ class PersonalizedBase(Dataset): self.width = width self.height = height self.flip = transforms.RandomHorizontalFlip(p=flip_p) + self.shuffle_tags = shuffle_tags self.dataset = [] @@ -98,7 +99,12 @@ class PersonalizedBase(Dataset): def create_text(self, filename_text): text = random.choice(self.lines) text = text.replace("[name]", self.placeholder_token) - text = text.replace("[filewords]", filename_text) + if self.tag_shuffle: + tags = filename_text.split(',') + random.shuffle(tags) + text = text.replace("[filewords]", ','.join(tags)) + else: + text = text.replace("[filewords]", filename_text) return text def __len__(self): diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index e0babb46..64700e23 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -224,7 +224,7 @@ def validate_train_inputs(model_name, learn_rate, batch_size, data_root, templat if save_model_every or create_image_every: assert log_directory, "Log directory is empty" -def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): +def train_embedding(embedding_name, learn_rate, batch_size, 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, shuffle_tags, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): save_embedding_every = save_embedding_every or 0 create_image_every = create_image_every or 0 validate_train_inputs(embedding_name, learn_rate, batch_size, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding") @@ -271,7 +271,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc # dataset loading may take a while, so input validations and early returns should be done before this shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..." with torch.autocast("cuda"): - ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file, batch_size=batch_size) + 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, shuffle_tags=shuffle_tags, model=shared.sd_model, device=devices.device, template_file=template_file, batch_size=batch_size) embedding.vec.requires_grad = True optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate) -- cgit v1.2.1 From 39541d7725bc42f456a604b07c50aba503a5a09a Mon Sep 17 00:00:00 2001 From: Fampai <> Date: Fri, 4 Nov 2022 04:50:22 -0400 Subject: Fixes race condition in training when VAE is unloaded set_current_image can attempt to use the VAE when it is unloaded to the CPU while training --- modules/textual_inversion/textual_inversion.py | 5 +++++ 1 file changed, 5 insertions(+) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 0aeb0459..55892c57 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -273,7 +273,11 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..." with torch.autocast("cuda"): ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file, batch_size=batch_size) + + old_parallel_processing_allowed = shared.parallel_processing_allowed + if unload: + shared.parallel_processing_allowed = False shared.sd_model.first_stage_model.to(devices.cpu) embedding.vec.requires_grad = True @@ -410,6 +414,7 @@ Last saved image: {html.escape(last_saved_image)}
filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt') save_embedding(embedding, checkpoint, embedding_name, filename, remove_cached_checksum=True) shared.sd_model.first_stage_model.to(devices.device) + shared.parallel_processing_allowed = old_parallel_processing_allowed return embedding, filename -- cgit v1.2.1 From 821e2b883dbb42a187bc37379175cd55b7cd7e81 Mon Sep 17 00:00:00 2001 From: TinkTheBoush Date: Fri, 4 Nov 2022 19:39:03 +0900 Subject: change option position to Training setting --- modules/textual_inversion/dataset.py | 5 ++--- modules/textual_inversion/textual_inversion.py | 4 ++-- 2 files changed, 4 insertions(+), 5 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index e9d97cc1..df278dc2 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -24,7 +24,7 @@ class DatasetEntry: class PersonalizedBase(Dataset): - def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", shuffle_tags=True, model=None, device=None, template_file=None, include_cond=False, batch_size=1): + def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, device=None, template_file=None, include_cond=False, batch_size=1): 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 @@ -33,7 +33,6 @@ class PersonalizedBase(Dataset): self.width = width self.height = height self.flip = transforms.RandomHorizontalFlip(p=flip_p) - self.shuffle_tags = shuffle_tags self.dataset = [] @@ -99,7 +98,7 @@ class PersonalizedBase(Dataset): def create_text(self, filename_text): text = random.choice(self.lines) text = text.replace("[name]", self.placeholder_token) - if self.tag_shuffle: + if shared.opts.shuffle_tags: tags = filename_text.split(',') random.shuffle(tags) text = text.replace("[filewords]", ','.join(tags)) diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 82dde931..0aeb0459 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -224,7 +224,7 @@ def validate_train_inputs(model_name, learn_rate, batch_size, data_root, templat if save_model_every or create_image_every: assert log_directory, "Log directory is empty" -def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, shuffle_tags, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): +def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): save_embedding_every = save_embedding_every or 0 create_image_every = create_image_every or 0 validate_train_inputs(embedding_name, learn_rate, batch_size, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding") @@ -272,7 +272,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc # dataset loading may take a while, so input validations and early returns should be done before this shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..." with torch.autocast("cuda"): - ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, shuffle_tags=shuffle_tags, model=shared.sd_model, device=devices.device, template_file=template_file, batch_size=batch_size) + ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file, batch_size=batch_size) if unload: shared.sd_model.first_stage_model.to(devices.cpu) -- cgit v1.2.1 From 8011be33c36eb7aa9e9498fc714614034e07f67a Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 8 Nov 2022 08:37:05 +0300 Subject: move functions out of main body for image preprocessing for easier hijacking --- modules/textual_inversion/preprocess.py | 162 ++++++++++++++++++-------------- 1 file changed, 93 insertions(+), 69 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/preprocess.py b/modules/textual_inversion/preprocess.py index e13b1894..488aa5b5 100644 --- a/modules/textual_inversion/preprocess.py +++ b/modules/textual_inversion/preprocess.py @@ -35,6 +35,84 @@ def preprocess(process_src, process_dst, process_width, process_height, preproce deepbooru.release_process() +def listfiles(dirname): + return os.listdir(dirname) + + +class PreprocessParams: + src = None + dstdir = None + subindex = 0 + flip = False + process_caption = False + process_caption_deepbooru = False + preprocess_txt_action = None + + +def save_pic_with_caption(image, index, params: PreprocessParams, existing_caption=None): + caption = "" + + if params.process_caption: + caption += shared.interrogator.generate_caption(image) + + if params.process_caption_deepbooru: + if len(caption) > 0: + caption += ", " + caption += deepbooru.get_tags_from_process(image) + + filename_part = params.src + filename_part = os.path.splitext(filename_part)[0] + filename_part = os.path.basename(filename_part) + + basename = f"{index:05}-{params.subindex}-{filename_part}" + image.save(os.path.join(params.dstdir, f"{basename}.png")) + + if params.preprocess_txt_action == 'prepend' and existing_caption: + caption = existing_caption + ' ' + caption + elif params.preprocess_txt_action == 'append' and existing_caption: + caption = caption + ' ' + existing_caption + elif params.preprocess_txt_action == 'copy' and existing_caption: + caption = existing_caption + + caption = caption.strip() + + if len(caption) > 0: + with open(os.path.join(params.dstdir, f"{basename}.txt"), "w", encoding="utf8") as file: + file.write(caption) + + params.subindex += 1 + + +def save_pic(image, index, params, existing_caption=None): + save_pic_with_caption(image, index, params, existing_caption=existing_caption) + + if params.flip: + save_pic_with_caption(ImageOps.mirror(image), index, params, existing_caption=existing_caption) + + +def split_pic(image, inverse_xy, width, height, overlap_ratio): + if inverse_xy: + from_w, from_h = image.height, image.width + to_w, to_h = height, width + else: + from_w, from_h = image.width, image.height + to_w, to_h = width, height + h = from_h * to_w // from_w + if inverse_xy: + image = image.resize((h, to_w)) + else: + image = image.resize((to_w, h)) + + split_count = math.ceil((h - to_h * overlap_ratio) / (to_h * (1.0 - overlap_ratio))) + y_step = (h - to_h) / (split_count - 1) + for i in range(split_count): + y = int(y_step * i) + if inverse_xy: + splitted = image.crop((y, 0, y + to_h, to_w)) + else: + splitted = image.crop((0, y, to_w, y + to_h)) + yield splitted + 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): width = process_width @@ -48,82 +126,28 @@ def preprocess_work(process_src, process_dst, process_width, process_height, pre os.makedirs(dst, exist_ok=True) - files = os.listdir(src) + files = listfiles(src) shared.state.textinfo = "Preprocessing..." shared.state.job_count = len(files) - def save_pic_with_caption(image, index, existing_caption=None): - caption = "" - - if process_caption: - caption += shared.interrogator.generate_caption(image) - - if process_caption_deepbooru: - if len(caption) > 0: - caption += ", " - caption += deepbooru.get_tags_from_process(image) - - filename_part = filename - filename_part = os.path.splitext(filename_part)[0] - filename_part = os.path.basename(filename_part) - - basename = f"{index:05}-{subindex[0]}-{filename_part}" - image.save(os.path.join(dst, f"{basename}.png")) - - if preprocess_txt_action == 'prepend' and existing_caption: - caption = existing_caption + ' ' + caption - elif preprocess_txt_action == 'append' and existing_caption: - caption = caption + ' ' + existing_caption - elif preprocess_txt_action == 'copy' and existing_caption: - caption = existing_caption - - caption = caption.strip() - - if len(caption) > 0: - with open(os.path.join(dst, f"{basename}.txt"), "w", encoding="utf8") as file: - file.write(caption) - - subindex[0] += 1 - - def save_pic(image, index, existing_caption=None): - save_pic_with_caption(image, index, existing_caption=existing_caption) - - if process_flip: - save_pic_with_caption(ImageOps.mirror(image), index, existing_caption=existing_caption) - - def split_pic(image, inverse_xy): - if inverse_xy: - from_w, from_h = image.height, image.width - to_w, to_h = height, width - else: - from_w, from_h = image.width, image.height - to_w, to_h = width, height - h = from_h * to_w // from_w - if inverse_xy: - image = image.resize((h, to_w)) - else: - image = image.resize((to_w, h)) - - split_count = math.ceil((h - to_h * overlap_ratio) / (to_h * (1.0 - overlap_ratio))) - y_step = (h - to_h) / (split_count - 1) - for i in range(split_count): - y = int(y_step * i) - if inverse_xy: - splitted = image.crop((y, 0, y + to_h, to_w)) - else: - splitted = image.crop((0, y, to_w, y + to_h)) - yield splitted - + params = PreprocessParams() + params.dstdir = dst + params.flip = process_flip + params.process_caption = process_caption + params.process_caption_deepbooru = process_caption_deepbooru + params.preprocess_txt_action = preprocess_txt_action for index, imagefile in enumerate(tqdm.tqdm(files)): - subindex = [0] + params.subindex = 0 filename = os.path.join(src, imagefile) try: img = Image.open(filename).convert("RGB") except Exception: continue + params.src = filename + existing_caption = None existing_caption_filename = os.path.splitext(filename)[0] + '.txt' if os.path.exists(existing_caption_filename): @@ -143,8 +167,8 @@ def preprocess_work(process_src, process_dst, process_width, process_height, pre process_default_resize = True if process_split and ratio < 1.0 and ratio <= split_threshold: - for splitted in split_pic(img, inverse_xy): - save_pic(splitted, index, existing_caption=existing_caption) + for splitted in split_pic(img, inverse_xy, width, height, overlap_ratio): + save_pic(splitted, index, params, existing_caption=existing_caption) process_default_resize = False if process_focal_crop and img.height != img.width: @@ -165,11 +189,11 @@ def preprocess_work(process_src, process_dst, process_width, process_height, pre dnn_model_path = dnn_model_path, ) for focal in autocrop.crop_image(img, autocrop_settings): - save_pic(focal, index, existing_caption=existing_caption) + save_pic(focal, index, params, existing_caption=existing_caption) process_default_resize = False if process_default_resize: img = images.resize_image(1, img, width, height) - save_pic(img, index, existing_caption=existing_caption) + save_pic(img, index, params, existing_caption=existing_caption) - shared.state.nextjob() \ No newline at end of file + shared.state.nextjob() -- cgit v1.2.1 From 13a2f1dca32980339e1fb4d1995cde428db798c5 Mon Sep 17 00:00:00 2001 From: KyuSeok Jung Date: Fri, 11 Nov 2022 10:29:55 +0900 Subject: adding tag drop out option --- modules/textual_inversion/dataset.py | 8 ++++---- 1 file 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 df278dc2..a95c7835 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -98,12 +98,12 @@ class PersonalizedBase(Dataset): def create_text(self, filename_text): text = random.choice(self.lines) text = text.replace("[name]", self.placeholder_token) + tags = filename_text.split(',') + if shared.opt.tag_drop_out != 0: + tags = [t for t in tags if random.random() > shared.opt.tag_drop_out] if shared.opts.shuffle_tags: - tags = filename_text.split(',') random.shuffle(tags) - text = text.replace("[filewords]", ','.join(tags)) - else: - text = text.replace("[filewords]", filename_text) + text = text.replace("[filewords]", ','.join(tags)) return text def __len__(self): -- cgit v1.2.1 From b19af67d29356f97fea5cccfdfa12583f605243f Mon Sep 17 00:00:00 2001 From: KyuSeok Jung Date: Fri, 11 Nov 2022 10:54:19 +0900 Subject: Update dataset.py --- modules/textual_inversion/dataset.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index a95c7835..e2cb8428 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -99,7 +99,7 @@ class PersonalizedBase(Dataset): text = random.choice(self.lines) text = text.replace("[name]", self.placeholder_token) tags = filename_text.split(',') - if shared.opt.tag_drop_out != 0: + if shared.opts.tag_drop_out != 0: tags = [t for t in tags if random.random() > shared.opt.tag_drop_out] if shared.opts.shuffle_tags: random.shuffle(tags) -- cgit v1.2.1 From a1e271207dfc3e89b1286ba41d96b459f210c4b2 Mon Sep 17 00:00:00 2001 From: KyuSeok Jung Date: Fri, 11 Nov 2022 10:56:53 +0900 Subject: Update dataset.py --- modules/textual_inversion/dataset.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index e2cb8428..eb75c376 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -100,7 +100,7 @@ class PersonalizedBase(Dataset): text = text.replace("[name]", self.placeholder_token) tags = filename_text.split(',') if shared.opts.tag_drop_out != 0: - tags = [t for t in tags if random.random() > shared.opt.tag_drop_out] + tags = [t for t in tags if random.random() > shared.opts.tag_drop_out] if shared.opts.shuffle_tags: random.shuffle(tags) text = text.replace("[filewords]", ','.join(tags)) -- cgit v1.2.1 From 9a1aff645a4bea745145c57c96950fbd3fcca27c Mon Sep 17 00:00:00 2001 From: parasi Date: Sun, 13 Nov 2022 13:44:27 -0600 Subject: resolve [name] after resolving [filewords] in training --- modules/textual_inversion/dataset.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index eb75c376..06f271f9 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -97,13 +97,13 @@ class PersonalizedBase(Dataset): def create_text(self, filename_text): text = random.choice(self.lines) - text = text.replace("[name]", self.placeholder_token) tags = filename_text.split(',') if shared.opts.tag_drop_out != 0: tags = [t for t in tags if random.random() > shared.opts.tag_drop_out] if shared.opts.shuffle_tags: random.shuffle(tags) text = text.replace("[filewords]", ','.join(tags)) + text = text.replace("[name]", self.placeholder_token) return text def __len__(self): -- cgit v1.2.1 From c8c40c8a643f2d20e3475e4d9ae7aae6d36c7e85 Mon Sep 17 00:00:00 2001 From: space-nuko <24979496+space-nuko@users.noreply.github.com> Date: Thu, 17 Nov 2022 18:03:57 -0800 Subject: Add interrupt button to preprocessing --- modules/textual_inversion/ui.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/ui.py b/modules/textual_inversion/ui.py index d679e6f4..35c4feef 100644 --- a/modules/textual_inversion/ui.py +++ b/modules/textual_inversion/ui.py @@ -18,7 +18,7 @@ def create_embedding(name, initialization_text, nvpt, overwrite_old): def preprocess(*args): modules.textual_inversion.preprocess.preprocess(*args) - return "Preprocessing finished.", "" + return f"Preprocessing {'interrupted' if shared.state.interrupted else 'finished'}.", "" def train_embedding(*args): -- cgit v1.2.1 From cdc8020d13c5eef099c609b0a911ccf3568afc0d Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sat, 19 Nov 2022 12:01:51 +0300 Subject: change StableDiffusionProcessing to internally use sampler name instead of sampler index --- modules/textual_inversion/textual_inversion.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 0aeb0459..5e4d8688 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -10,7 +10,7 @@ import csv from PIL import Image, PngImagePlugin -from modules import shared, devices, sd_hijack, processing, sd_models, images +from modules import shared, devices, sd_hijack, processing, sd_models, images, sd_samplers import modules.textual_inversion.dataset from modules.textual_inversion.learn_schedule import LearnRateScheduler @@ -345,7 +345,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc p.prompt = preview_prompt p.negative_prompt = preview_negative_prompt p.steps = preview_steps - p.sampler_index = preview_sampler_index + p.sampler_name = sd_samplers.samplers[preview_sampler_index].name p.cfg_scale = preview_cfg_scale p.seed = preview_seed p.width = preview_width -- cgit v1.2.1 From bd68e35de3b7cf7547ed97d8bdf60147402133cc Mon Sep 17 00:00:00 2001 From: flamelaw Date: Sun, 20 Nov 2022 12:35:26 +0900 Subject: Gradient accumulation, autocast fix, new latent sampling method, etc --- modules/textual_inversion/dataset.py | 134 +++++++---- modules/textual_inversion/textual_inversion.py | 320 ++++++++++++++----------- 2 files changed, 269 insertions(+), 185 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index eb75c376..d594b49d 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -3,7 +3,7 @@ import numpy as np import PIL import torch from PIL import Image -from torch.utils.data import Dataset +from torch.utils.data import Dataset, DataLoader from torchvision import transforms import random @@ -11,25 +11,28 @@ import tqdm from modules import devices, shared import re +from ldm.modules.distributions.distributions import DiagonalGaussianDistribution + re_numbers_at_start = re.compile(r"^[-\d]+\s*") class DatasetEntry: - def __init__(self, filename=None, latent=None, filename_text=None): + def __init__(self, filename=None, filename_text=None, latent_dist=None, latent_sample=None, cond=None, cond_text=None, pixel_values=None): self.filename = filename - self.latent = latent self.filename_text = filename_text - self.cond = None - self.cond_text = None + self.latent_dist = latent_dist + self.latent_sample = latent_sample + self.cond = cond + self.cond_text = cond_text + self.pixel_values = pixel_values class PersonalizedBase(Dataset): - def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, device=None, template_file=None, include_cond=False, batch_size=1): + def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, cond_model=None, device=None, template_file=None, include_cond=False, batch_size=1, gradient_step=1, shuffle_tags=False, tag_drop_out=0, latent_sampling_method='once'): re_word = re.compile(shared.opts.dataset_filename_word_regex) if len(shared.opts.dataset_filename_word_regex) > 0 else None - + self.placeholder_token = placeholder_token - self.batch_size = batch_size self.width = width self.height = height self.flip = transforms.RandomHorizontalFlip(p=flip_p) @@ -45,11 +48,16 @@ class PersonalizedBase(Dataset): assert os.path.isdir(data_root), "Dataset directory doesn't exist" assert os.listdir(data_root), "Dataset directory is empty" - cond_model = shared.sd_model.cond_stage_model - self.image_paths = [os.path.join(data_root, file_path) for file_path in os.listdir(data_root)] + + + self.shuffle_tags = shuffle_tags + self.tag_drop_out = tag_drop_out + print("Preparing dataset...") for path in tqdm.tqdm(self.image_paths): + if shared.state.interrupted: + raise Exception("inturrupted") try: image = Image.open(path).convert('RGB').resize((self.width, self.height), PIL.Image.BICUBIC) except Exception: @@ -71,37 +79,58 @@ class PersonalizedBase(Dataset): npimage = np.array(image).astype(np.uint8) npimage = (npimage / 127.5 - 1.0).astype(np.float32) - torchdata = torch.from_numpy(npimage).to(device=device, dtype=torch.float32) - torchdata = torch.moveaxis(torchdata, 2, 0) - - init_latent = model.get_first_stage_encoding(model.encode_first_stage(torchdata.unsqueeze(dim=0))).squeeze() - init_latent = init_latent.to(devices.cpu) - - entry = DatasetEntry(filename=path, filename_text=filename_text, latent=init_latent) - - if include_cond: + torchdata = torch.from_numpy(npimage).permute(2, 0, 1).to(device=device, dtype=torch.float32) + latent_sample = None + + with torch.autocast("cuda"): + latent_dist = model.encode_first_stage(torchdata.unsqueeze(dim=0)) + + if latent_sampling_method == "once" or (latent_sampling_method == "deterministic" and not isinstance(latent_dist, DiagonalGaussianDistribution)): + latent_sample = model.get_first_stage_encoding(latent_dist).squeeze().to(devices.cpu) + latent_sampling_method = "once" + entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample) + elif latent_sampling_method == "deterministic": + # Works only for DiagonalGaussianDistribution + latent_dist.std = 0 + latent_sample = model.get_first_stage_encoding(latent_dist).squeeze().to(devices.cpu) + entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample) + elif latent_sampling_method == "random": + entry = DatasetEntry(filename=path, filename_text=filename_text, latent_dist=latent_dist) + + if not (self.tag_drop_out != 0 or self.shuffle_tags): entry.cond_text = self.create_text(filename_text) - entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0) - self.dataset.append(entry) - - assert len(self.dataset) > 0, "No images have been found in the dataset." - self.length = len(self.dataset) * repeats // batch_size + if include_cond and not (self.tag_drop_out != 0 or self.shuffle_tags): + with torch.autocast("cuda"): + entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0) + # elif not include_cond: + # _, _, _, _, hijack_fixes, token_count = cond_model.process_text([entry.cond_text]) + # max_n = token_count // 75 + # index_list = [ [] for _ in range(max_n + 1) ] + # for n, (z, _) in hijack_fixes[0]: + # index_list[n].append(z) + # with torch.autocast("cuda"): + # entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0) + # entry.emb_index = index_list - self.dataset_length = len(self.dataset) - self.indexes = None - self.shuffle() + self.dataset.append(entry) + del torchdata + del latent_dist + del latent_sample - def shuffle(self): - self.indexes = np.random.permutation(self.dataset_length) + self.length = len(self.dataset) + assert self.length > 0, "No images have been found in the dataset." + self.batch_size = min(batch_size, self.length) + self.gradient_step = min(gradient_step, self.length // self.batch_size) + self.latent_sampling_method = latent_sampling_method def create_text(self, filename_text): text = random.choice(self.lines) text = text.replace("[name]", self.placeholder_token) tags = filename_text.split(',') - if shared.opts.tag_drop_out != 0: - tags = [t for t in tags if random.random() > shared.opts.tag_drop_out] - if shared.opts.shuffle_tags: + if self.tag_drop_out != 0: + tags = [t for t in tags if random.random() > self.tag_drop_out] + if self.shuffle_tags: random.shuffle(tags) text = text.replace("[filewords]", ','.join(tags)) return text @@ -110,19 +139,28 @@ class PersonalizedBase(Dataset): return self.length def __getitem__(self, i): - res = [] - - for j in range(self.batch_size): - position = i * self.batch_size + j - if position % len(self.indexes) == 0: - self.shuffle() - - index = self.indexes[position % len(self.indexes)] - entry = self.dataset[index] - - if entry.cond is None: - entry.cond_text = self.create_text(entry.filename_text) - - res.append(entry) - - return res + entry = self.dataset[i] + if self.tag_drop_out != 0 or self.shuffle_tags: + entry.cond_text = self.create_text(entry.filename_text) + if self.latent_sampling_method == "random": + entry.latent_sample = shared.sd_model.get_first_stage_encoding(entry.latent_dist) + return entry + +class PersonalizedDataLoader(DataLoader): + def __init__(self, *args, **kwargs): + super(PersonalizedDataLoader, self).__init__(shuffle=True, drop_last=True, *args, **kwargs) + self.collate_fn = collate_wrapper + + +class BatchLoader: + def __init__(self, data): + self.cond_text = [entry.cond_text for entry in data] + self.cond = [entry.cond for entry in data] + self.latent_sample = torch.stack([entry.latent_sample for entry in data]).squeeze(1) + + def pin_memory(self): + self.latent_sample = self.latent_sample.pin_memory() + return self + +def collate_wrapper(batch): + return BatchLoader(batch) \ No newline at end of file diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 5e4d8688..1d5e3a32 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -184,7 +184,7 @@ def write_loss(log_directory, filename, step, epoch_len, values): if shared.opts.training_write_csv_every == 0: return - if (step + 1) % shared.opts.training_write_csv_every != 0: + if step % shared.opts.training_write_csv_every != 0: return write_csv_header = False if os.path.exists(os.path.join(log_directory, filename)) else True @@ -194,21 +194,23 @@ def write_loss(log_directory, filename, step, epoch_len, values): if write_csv_header: csv_writer.writeheader() - epoch = step // epoch_len - epoch_step = step % epoch_len + epoch = (step - 1) // epoch_len + epoch_step = (step - 1) % epoch_len csv_writer.writerow({ - "step": step + 1, + "step": step, "epoch": epoch, - "epoch_step": epoch_step + 1, + "epoch_step": epoch_step, **values, }) -def validate_train_inputs(model_name, learn_rate, batch_size, data_root, template_file, steps, save_model_every, create_image_every, log_directory, name="embedding"): +def validate_train_inputs(model_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_model_every, create_image_every, log_directory, name="embedding"): assert model_name, f"{name} not selected" assert learn_rate, "Learning rate is empty or 0" assert isinstance(batch_size, int), "Batch size must be integer" assert batch_size > 0, "Batch size must be positive" + assert isinstance(gradient_step, int), "Gradient accumulation step must be integer" + assert gradient_step > 0, "Gradient accumulation step must be positive" assert data_root, "Dataset directory is empty" assert os.path.isdir(data_root), "Dataset directory doesn't exist" assert os.listdir(data_root), "Dataset directory is empty" @@ -224,10 +226,10 @@ def validate_train_inputs(model_name, learn_rate, batch_size, data_root, templat if save_model_every or create_image_every: assert log_directory, "Log directory is empty" -def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): +def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, steps, shuffle_tags, tag_drop_out, latent_sampling_method, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): save_embedding_every = save_embedding_every or 0 create_image_every = create_image_every or 0 - validate_train_inputs(embedding_name, learn_rate, batch_size, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding") + validate_train_inputs(embedding_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding") shared.state.textinfo = "Initializing textual inversion training..." shared.state.job_count = steps @@ -255,161 +257,205 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc else: images_embeds_dir = None - cond_model = shared.sd_model.cond_stage_model - hijack = sd_hijack.model_hijack embedding = hijack.embedding_db.word_embeddings[embedding_name] checkpoint = sd_models.select_checkpoint() - ititial_step = embedding.step or 0 - if ititial_step >= steps: + initial_step = embedding.step or 0 + if initial_step >= steps: shared.state.textinfo = f"Model has already been trained beyond specified max steps" return embedding, filename + scheduler = LearnRateScheduler(learn_rate, steps, initial_step) - scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) - - # dataset loading may take a while, so input validations and early returns should be done before this + # dataset loading may take a while, so input validations and early returns should be done before this shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..." - with torch.autocast("cuda"): - ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file, batch_size=batch_size) + + pin_memory = shared.opts.pin_memory + + ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, cond_model=shared.sd_model.cond_stage_model, device=devices.device, template_file=template_file, batch_size=batch_size, gradient_step=gradient_step, shuffle_tags=shuffle_tags, tag_drop_out=tag_drop_out, latent_sampling_method=latent_sampling_method) + + latent_sampling_method = ds.latent_sampling_method + + dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, batch_size=ds.batch_size, pin_memory=False) + if unload: shared.sd_model.first_stage_model.to(devices.cpu) embedding.vec.requires_grad = True optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate) + scaler = torch.cuda.amp.GradScaler() - losses = torch.zeros((32,)) + batch_size = ds.batch_size + gradient_step = ds.gradient_step + # n steps = batch_size * gradient_step * n image processed + steps_per_epoch = len(ds) // batch_size // gradient_step + max_steps_per_epoch = len(ds) // batch_size - (len(ds) // batch_size) % gradient_step + loss_step = 0 + _loss_step = 0 #internal + last_saved_file = "" last_saved_image = "" forced_filename = "" embedding_yet_to_be_embedded = False - - pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) - for i, entries in pbar: - embedding.step = i + ititial_step - - scheduler.apply(optimizer, embedding.step) - if scheduler.finished: - break - - if shared.state.interrupted: - break - - with torch.autocast("cuda"): - c = cond_model([entry.cond_text for entry in entries]) - x = torch.stack([entry.latent for entry in entries]).to(devices.device) - loss = shared.sd_model(x, c)[0] - del x - - losses[embedding.step % losses.shape[0]] = loss.item() - - optimizer.zero_grad() - loss.backward() - optimizer.step() - - steps_done = embedding.step + 1 - - epoch_num = embedding.step // len(ds) - epoch_step = embedding.step % len(ds) - - pbar.set_description(f"[Epoch {epoch_num}: {epoch_step+1}/{len(ds)}]loss: {losses.mean():.7f}") - - if embedding_dir is not None and steps_done % save_embedding_every == 0: - # Before saving, change name to match current checkpoint. - embedding_name_every = f'{embedding_name}-{steps_done}' - last_saved_file = os.path.join(embedding_dir, f'{embedding_name_every}.pt') - save_embedding(embedding, checkpoint, embedding_name_every, last_saved_file, remove_cached_checksum=True) - embedding_yet_to_be_embedded = True - - write_loss(log_directory, "textual_inversion_loss.csv", embedding.step, len(ds), { - "loss": f"{losses.mean():.7f}", - "learn_rate": scheduler.learn_rate - }) - - if images_dir is not None and steps_done % create_image_every == 0: - forced_filename = f'{embedding_name}-{steps_done}' - last_saved_image = os.path.join(images_dir, forced_filename) - - shared.sd_model.first_stage_model.to(devices.device) - - p = processing.StableDiffusionProcessingTxt2Img( - sd_model=shared.sd_model, - do_not_save_grid=True, - do_not_save_samples=True, - do_not_reload_embeddings=True, - ) - - if preview_from_txt2img: - p.prompt = preview_prompt - p.negative_prompt = preview_negative_prompt - p.steps = preview_steps - p.sampler_name = sd_samplers.samplers[preview_sampler_index].name - p.cfg_scale = preview_cfg_scale - p.seed = preview_seed - p.width = preview_width - p.height = preview_height - else: - p.prompt = entries[0].cond_text - p.steps = 20 - p.width = training_width - p.height = training_height - - preview_text = p.prompt - - processed = processing.process_images(p) - image = processed.images[0] - - if unload: - shared.sd_model.first_stage_model.to(devices.cpu) - - shared.state.current_image = image - - if save_image_with_stored_embedding and os.path.exists(last_saved_file) and embedding_yet_to_be_embedded: - - last_saved_image_chunks = os.path.join(images_embeds_dir, f'{embedding_name}-{steps_done}.png') - - info = PngImagePlugin.PngInfo() - data = torch.load(last_saved_file) - info.add_text("sd-ti-embedding", embedding_to_b64(data)) - - title = "<{}>".format(data.get('name', '???')) - - try: - vectorSize = list(data['string_to_param'].values())[0].shape[0] - except Exception as e: - vectorSize = '?' - - checkpoint = sd_models.select_checkpoint() - footer_left = checkpoint.model_name - footer_mid = '[{}]'.format(checkpoint.hash) - footer_right = '{}v {}s'.format(vectorSize, steps_done) - - captioned_image = caption_image_overlay(image, title, footer_left, footer_mid, footer_right) - captioned_image = insert_image_data_embed(captioned_image, data) - - captioned_image.save(last_saved_image_chunks, "PNG", pnginfo=info) - embedding_yet_to_be_embedded = False - - last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False) - last_saved_image += f", prompt: {preview_text}" - - shared.state.job_no = embedding.step - - shared.state.textinfo = f""" + + pbar = tqdm.tqdm(total=steps - initial_step) + try: + for i in range((steps-initial_step) * gradient_step): + if scheduler.finished: + break + if shared.state.interrupted: + break + for j, batch in enumerate(dl): + # works as a drop_last=True for gradient accumulation + if j == max_steps_per_epoch: + break + scheduler.apply(optimizer, embedding.step) + if scheduler.finished: + break + if shared.state.interrupted: + break + + with torch.autocast("cuda"): + # c = stack_conds(batch.cond).to(devices.device) + # mask = torch.tensor(batch.emb_index).to(devices.device, non_blocking=pin_memory) + # print(mask) + # c[:, 1:1+embedding.vec.shape[0]] = embedding.vec.to(devices.device, non_blocking=pin_memory) + x = batch.latent_sample.to(devices.device, non_blocking=pin_memory) + c = shared.sd_model.cond_stage_model(batch.cond_text) + loss = shared.sd_model(x, c)[0] / gradient_step + del x + + _loss_step += loss.item() + scaler.scale(loss).backward() + + # go back until we reach gradient accumulation steps + if (j + 1) % gradient_step != 0: + continue + #print(f"grad:{embedding.vec.grad.detach().cpu().abs().mean().item():.7f}") + #scaler.unscale_(optimizer) + #print(f"grad:{embedding.vec.grad.detach().cpu().abs().mean().item():.7f}") + #torch.nn.utils.clip_grad_norm_(embedding.vec, max_norm=1.0) + #print(f"grad:{embedding.vec.grad.detach().cpu().abs().mean().item():.7f}") + scaler.step(optimizer) + scaler.update() + embedding.step += 1 + pbar.update() + optimizer.zero_grad(set_to_none=True) + loss_step = _loss_step + _loss_step = 0 + + steps_done = embedding.step + 1 + + epoch_num = embedding.step // steps_per_epoch + epoch_step = embedding.step % steps_per_epoch + + pbar.set_description(f"[Epoch {epoch_num}: {epoch_step+1}/{steps_per_epoch}]loss: {loss_step:.7f}") + if embedding_dir is not None and steps_done % save_embedding_every == 0: + # Before saving, change name to match current checkpoint. + embedding_name_every = f'{embedding_name}-{steps_done}' + last_saved_file = os.path.join(embedding_dir, f'{embedding_name_every}.pt') + #if shared.opts.save_optimizer_state: + #embedding.optimizer_state_dict = optimizer.state_dict() + save_embedding(embedding, checkpoint, embedding_name_every, last_saved_file, remove_cached_checksum=True) + embedding_yet_to_be_embedded = True + + write_loss(log_directory, "textual_inversion_loss.csv", embedding.step, steps_per_epoch, { + "loss": f"{loss_step:.7f}", + "learn_rate": scheduler.learn_rate + }) + + if images_dir is not None and steps_done % create_image_every == 0: + forced_filename = f'{embedding_name}-{steps_done}' + last_saved_image = os.path.join(images_dir, forced_filename) + + shared.sd_model.first_stage_model.to(devices.device) + + p = processing.StableDiffusionProcessingTxt2Img( + sd_model=shared.sd_model, + do_not_save_grid=True, + do_not_save_samples=True, + do_not_reload_embeddings=True, + ) + + if preview_from_txt2img: + p.prompt = preview_prompt + p.negative_prompt = preview_negative_prompt + p.steps = preview_steps + p.sampler_name = sd_samplers.samplers[preview_sampler_index].name + p.cfg_scale = preview_cfg_scale + p.seed = preview_seed + p.width = preview_width + p.height = preview_height + else: + p.prompt = batch.cond_text[0] + p.steps = 20 + p.width = training_width + p.height = training_height + + preview_text = p.prompt + + processed = processing.process_images(p) + image = processed.images[0] if len(processed.images) > 0 else None + + if unload: + shared.sd_model.first_stage_model.to(devices.cpu) + + if image is not None: + shared.state.current_image = image + last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False) + last_saved_image += f", prompt: {preview_text}" + + if save_image_with_stored_embedding and os.path.exists(last_saved_file) and embedding_yet_to_be_embedded: + + last_saved_image_chunks = os.path.join(images_embeds_dir, f'{embedding_name}-{steps_done}.png') + + info = PngImagePlugin.PngInfo() + data = torch.load(last_saved_file) + info.add_text("sd-ti-embedding", embedding_to_b64(data)) + + title = "<{}>".format(data.get('name', '???')) + + try: + vectorSize = list(data['string_to_param'].values())[0].shape[0] + except Exception as e: + vectorSize = '?' + + checkpoint = sd_models.select_checkpoint() + footer_left = checkpoint.model_name + footer_mid = '[{}]'.format(checkpoint.hash) + footer_right = '{}v {}s'.format(vectorSize, steps_done) + + captioned_image = caption_image_overlay(image, title, footer_left, footer_mid, footer_right) + captioned_image = insert_image_data_embed(captioned_image, data) + + captioned_image.save(last_saved_image_chunks, "PNG", pnginfo=info) + embedding_yet_to_be_embedded = False + + last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False) + last_saved_image += f", prompt: {preview_text}" + + shared.state.job_no = embedding.step + + shared.state.textinfo = f"""

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

""" - - filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt') - save_embedding(embedding, checkpoint, embedding_name, filename, remove_cached_checksum=True) - shared.sd_model.first_stage_model.to(devices.device) + filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt') + save_embedding(embedding, checkpoint, embedding_name, filename, remove_cached_checksum=True) + except Exception: + print(traceback.format_exc(), file=sys.stderr) + pass + finally: + pbar.leave = False + pbar.close() + shared.sd_model.first_stage_model.to(devices.device) return embedding, filename -- cgit v1.2.1 From a4a5735d0a80218e59f8a6e8401726f7209a6a8d Mon Sep 17 00:00:00 2001 From: flamelaw Date: Sun, 20 Nov 2022 12:38:18 +0900 Subject: remove unnecessary comment --- modules/textual_inversion/dataset.py | 9 --------- 1 file changed, 9 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index d594b49d..1dd53b85 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -103,15 +103,6 @@ class PersonalizedBase(Dataset): if include_cond and not (self.tag_drop_out != 0 or self.shuffle_tags): with torch.autocast("cuda"): entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0) - # elif not include_cond: - # _, _, _, _, hijack_fixes, token_count = cond_model.process_text([entry.cond_text]) - # max_n = token_count // 75 - # index_list = [ [] for _ in range(max_n + 1) ] - # for n, (z, _) in hijack_fixes[0]: - # index_list[n].append(z) - # with torch.autocast("cuda"): - # entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0) - # entry.emb_index = index_list self.dataset.append(entry) del torchdata -- cgit v1.2.1 From 2d22d72cdaaf2b78b2986b841d478c11ac855dd2 Mon Sep 17 00:00:00 2001 From: flamelaw Date: Sun, 20 Nov 2022 16:14:27 +0900 Subject: fix random sampling with pin_memory --- modules/textual_inversion/dataset.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index 1dd53b85..110c0e09 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -134,7 +134,7 @@ class PersonalizedBase(Dataset): if self.tag_drop_out != 0 or self.shuffle_tags: entry.cond_text = self.create_text(entry.filename_text) if self.latent_sampling_method == "random": - entry.latent_sample = shared.sd_model.get_first_stage_encoding(entry.latent_dist) + entry.latent_sample = shared.sd_model.get_first_stage_encoding(entry.latent_dist).to(devices.cpu) return entry class PersonalizedDataLoader(DataLoader): -- cgit v1.2.1 From c81d440d876dfd2ab3560410f37442ef56fc6632 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sun, 20 Nov 2022 16:39:20 +0300 Subject: moved deepdanbooru to pure pytorch implementation --- modules/textual_inversion/preprocess.py | 12 ++++-------- 1 file changed, 4 insertions(+), 8 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/preprocess.py b/modules/textual_inversion/preprocess.py index 488aa5b5..56b9b2eb 100644 --- a/modules/textual_inversion/preprocess.py +++ b/modules/textual_inversion/preprocess.py @@ -6,12 +6,10 @@ import sys import tqdm import time -from modules import shared, images +from modules import shared, images, deepbooru from modules.paths import models_path from modules.shared import opts, cmd_opts from modules.textual_inversion import autocrop -if cmd_opts.deepdanbooru: - import modules.deepbooru as deepbooru 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): @@ -20,9 +18,7 @@ def preprocess(process_src, process_dst, process_width, process_height, preproce shared.interrogator.load() if process_caption_deepbooru: - db_opts = deepbooru.create_deepbooru_opts() - db_opts[deepbooru.OPT_INCLUDE_RANKS] = False - deepbooru.create_deepbooru_process(opts.interrogate_deepbooru_score_threshold, db_opts) + 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) @@ -32,7 +28,7 @@ def preprocess(process_src, process_dst, process_width, process_height, preproce shared.interrogator.send_blip_to_ram() if process_caption_deepbooru: - deepbooru.release_process() + deepbooru.model.stop() def listfiles(dirname): @@ -58,7 +54,7 @@ def save_pic_with_caption(image, index, params: PreprocessParams, existing_capti if params.process_caption_deepbooru: if len(caption) > 0: caption += ", " - caption += deepbooru.get_tags_from_process(image) + caption += deepbooru.model.tag_multi(image) filename_part = params.src filename_part = os.path.splitext(filename_part)[0] -- cgit v1.2.1 From 5b57f61ba47f8b11d19a5b46e7fb5a52458abae5 Mon Sep 17 00:00:00 2001 From: flamelaw Date: Mon, 21 Nov 2022 10:15:46 +0900 Subject: fix pin_memory with different latent sampling method --- modules/textual_inversion/dataset.py | 23 +++++++++++++++++++---- modules/textual_inversion/textual_inversion.py | 7 +------ 2 files changed, 20 insertions(+), 10 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index 110c0e09..f470324a 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -138,9 +138,12 @@ class PersonalizedBase(Dataset): return entry class PersonalizedDataLoader(DataLoader): - def __init__(self, *args, **kwargs): - super(PersonalizedDataLoader, self).__init__(shuffle=True, drop_last=True, *args, **kwargs) - self.collate_fn = collate_wrapper + def __init__(self, dataset, latent_sampling_method="once", batch_size=1, pin_memory=False): + super(PersonalizedDataLoader, self).__init__(dataset, shuffle=True, drop_last=True, batch_size=batch_size, pin_memory=pin_memory) + if latent_sampling_method == "random": + self.collate_fn = collate_wrapper_random + else: + self.collate_fn = collate_wrapper class BatchLoader: @@ -148,10 +151,22 @@ class BatchLoader: self.cond_text = [entry.cond_text for entry in data] self.cond = [entry.cond for entry in data] self.latent_sample = torch.stack([entry.latent_sample for entry in data]).squeeze(1) + #self.emb_index = [entry.emb_index for entry in data] + #print(self.latent_sample.device) def pin_memory(self): self.latent_sample = self.latent_sample.pin_memory() return self def collate_wrapper(batch): - return BatchLoader(batch) \ No newline at end of file + return BatchLoader(batch) + +class BatchLoaderRandom(BatchLoader): + def __init__(self, data): + super().__init__(data) + + def pin_memory(self): + return self + +def collate_wrapper_random(batch): + return BatchLoaderRandom(batch) \ No newline at end of file diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 1d5e3a32..3036e48a 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -277,7 +277,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ latent_sampling_method = ds.latent_sampling_method - dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, batch_size=ds.batch_size, pin_memory=False) + dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, latent_sampling_method=latent_sampling_method, batch_size=ds.batch_size, pin_memory=pin_memory) if unload: shared.sd_model.first_stage_model.to(devices.cpu) @@ -333,11 +333,6 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ # go back until we reach gradient accumulation steps if (j + 1) % gradient_step != 0: continue - #print(f"grad:{embedding.vec.grad.detach().cpu().abs().mean().item():.7f}") - #scaler.unscale_(optimizer) - #print(f"grad:{embedding.vec.grad.detach().cpu().abs().mean().item():.7f}") - #torch.nn.utils.clip_grad_norm_(embedding.vec, max_norm=1.0) - #print(f"grad:{embedding.vec.grad.detach().cpu().abs().mean().item():.7f}") scaler.step(optimizer) scaler.update() embedding.step += 1 -- cgit v1.2.1 From 89d8ecff09b426ddc89eb5b432825f8f4c218051 Mon Sep 17 00:00:00 2001 From: flamelaw Date: Wed, 23 Nov 2022 02:49:01 +0900 Subject: small fixes --- modules/textual_inversion/textual_inversion.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 3036e48a..fee08e33 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -436,7 +436,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ shared.state.textinfo = f"""

Loss: {loss_step:.7f}
-Step: {embedding.step}
+Step: {steps_done}
Last prompt: {html.escape(batch.cond_text[0])}
Last saved embedding: {html.escape(last_saved_file)}
Last saved image: {html.escape(last_saved_image)}
-- cgit v1.2.1 From ce6911158b5b2f9cf79b405a1f368f875492044d Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sat, 26 Nov 2022 16:10:46 +0300 Subject: Add support Stable Diffusion 2.0 --- modules/textual_inversion/textual_inversion.py | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 5e4d8688..a273e663 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -64,7 +64,8 @@ class EmbeddingDatabase: self.word_embeddings[embedding.name] = embedding - ids = model.cond_stage_model.tokenizer([embedding.name], add_special_tokens=False)['input_ids'][0] + # TODO changing between clip and open clip changes tokenization, which will cause embeddings to stop working + ids = model.cond_stage_model.tokenize([embedding.name])[0] first_id = ids[0] if first_id not in self.ids_lookup: @@ -155,13 +156,11 @@ class EmbeddingDatabase: def create_embedding(name, num_vectors_per_token, overwrite_old, init_text='*'): cond_model = shared.sd_model.cond_stage_model - embedding_layer = cond_model.wrapped.transformer.text_model.embeddings with devices.autocast(): cond_model([""]) # will send cond model to GPU if lowvram/medvram is active - ids = cond_model.tokenizer(init_text, max_length=num_vectors_per_token, return_tensors="pt", add_special_tokens=False)["input_ids"] - embedded = embedding_layer.token_embedding.wrapped(ids.to(devices.device)).squeeze(0) + embedded = cond_model.encode_embedding_init_text(init_text, num_vectors_per_token) vec = torch.zeros((num_vectors_per_token, embedded.shape[1]), device=devices.device) for i in range(num_vectors_per_token): -- cgit v1.2.1 From 755df94b2aa62eabd96f900e0dd7ddc83c2f692c Mon Sep 17 00:00:00 2001 From: flamelaw Date: Sun, 27 Nov 2022 00:35:44 +0900 Subject: set TI AdamW default weight decay to 0 --- modules/textual_inversion/textual_inversion.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index fee08e33..b9b1394f 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -283,7 +283,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ shared.sd_model.first_stage_model.to(devices.cpu) embedding.vec.requires_grad = True - optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate) + optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate, weight_decay=0.0) scaler = torch.cuda.amp.GradScaler() batch_size = ds.batch_size -- cgit v1.2.1 From 4d5f1691dda971ec7b461dd880426300fd54ccee Mon Sep 17 00:00:00 2001 From: brkirch Date: Mon, 28 Nov 2022 21:36:35 -0500 Subject: Use devices.autocast instead of torch.autocast --- modules/textual_inversion/dataset.py | 4 ++-- modules/textual_inversion/textual_inversion.py | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index e5725f33..2dc64c3c 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -82,7 +82,7 @@ class PersonalizedBase(Dataset): torchdata = torch.from_numpy(npimage).permute(2, 0, 1).to(device=device, dtype=torch.float32) latent_sample = None - with torch.autocast("cuda"): + with devices.autocast(): latent_dist = model.encode_first_stage(torchdata.unsqueeze(dim=0)) if latent_sampling_method == "once" or (latent_sampling_method == "deterministic" and not isinstance(latent_dist, DiagonalGaussianDistribution)): @@ -101,7 +101,7 @@ class PersonalizedBase(Dataset): entry.cond_text = self.create_text(filename_text) if include_cond and not (self.tag_drop_out != 0 or self.shuffle_tags): - with torch.autocast("cuda"): + with devices.autocast(): entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0) self.dataset.append(entry) diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 4eb75cb5..daf8d1b8 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -316,7 +316,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ if shared.state.interrupted: break - with torch.autocast("cuda"): + with devices.autocast(): # c = stack_conds(batch.cond).to(devices.device) # mask = torch.tensor(batch.emb_index).to(devices.device, non_blocking=pin_memory) # print(mask) -- cgit v1.2.1 From 119a945ef7569128eb7d6772468ffc5567c2e161 Mon Sep 17 00:00:00 2001 From: PhytoEpidemic <64293310+PhytoEpidemic@users.noreply.github.com> Date: Fri, 2 Dec 2022 12:16:29 -0600 Subject: Fix divide by 0 error Fix of the edge case 0 weight that occasionally will pop up in some specific situations. This was crashing the script. --- modules/textual_inversion/autocrop.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'modules/textual_inversion') diff --git a/modules/textual_inversion/autocrop.py b/modules/textual_inversion/autocrop.py index 9859974a..68e1103c 100644 --- a/modules/textual_inversion/autocrop.py +++ b/modules/textual_inversion/autocrop.py @@ -276,8 +276,8 @@ def poi_average(pois, settings): weight += poi.weight x += poi.x * poi.weight y += poi.y * poi.weight - avg_x = round(x / weight) - avg_y = round(y / weight) + avg_x = round(weight and x / weight) + avg_y = round(weight and y / weight) return PointOfInterest(avg_x, avg_y) @@ -338,4 +338,4 @@ class Settings: self.face_points_weight = face_points_weight self.annotate_image = annotate_image self.destop_view_image = False - self.dnn_model_path = dnn_model_path \ No newline at end of file + self.dnn_model_path = dnn_model_path -- cgit v1.2.1