From 820f1dc96b1979d7e92170c161db281ee8bd988b Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sun, 2 Oct 2022 15:03:39 +0300 Subject: initial support for training textual inversion --- modules/textual_inversion/dataset.py | 76 ++++++++++++++++++++++++++++++++++++ 1 file changed, 76 insertions(+) create mode 100644 modules/textual_inversion/dataset.py (limited to 'modules/textual_inversion/dataset.py') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py new file mode 100644 index 00000000..7e134a08 --- /dev/null +++ b/modules/textual_inversion/dataset.py @@ -0,0 +1,76 @@ +import os +import numpy as np +import PIL +import torch +from PIL import Image +from torch.utils.data import Dataset +from torchvision import transforms + +import random +import tqdm + + +class PersonalizedBase(Dataset): + def __init__(self, data_root, size=None, repeats=100, flip_p=0.5, placeholder_token="*", width=512, height=512, model=None, device=None, template_file=None): + + self.placeholder_token = placeholder_token + + self.size = size + self.width = width + self.height = height + self.flip = transforms.RandomHorizontalFlip(p=flip_p) + + self.dataset = [] + + with open(template_file, "r") as file: + lines = [x.strip() for x in file.readlines()] + + self.lines = lines + + assert data_root, 'dataset directory not specified' + + self.image_paths = [os.path.join(data_root, file_path) for file_path in os.listdir(data_root)] + print("Preparing dataset...") + for path in tqdm.tqdm(self.image_paths): + image = Image.open(path) + image = image.convert('RGB') + image = image.resize((self.width, self.height), PIL.Image.BICUBIC) + + filename = os.path.basename(path) + filename_tokens = os.path.splitext(filename)[0].replace('_', '-').replace(' ', '-').split('-') + filename_tokens = [token for token in filename_tokens if token.isalpha()] + + 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() + + self.dataset.append((init_latent, filename_tokens)) + + self.length = len(self.dataset) * repeats + + self.initial_indexes = np.arange(self.length) % len(self.dataset) + self.indexes = None + self.shuffle() + + def shuffle(self): + self.indexes = self.initial_indexes[torch.randperm(self.initial_indexes.shape[0])] + + def __len__(self): + return self.length + + def __getitem__(self, i): + if i % len(self.dataset) == 0: + self.shuffle() + + index = self.indexes[i % len(self.indexes)] + x, filename_tokens = self.dataset[index] + + text = random.choice(self.lines) + text = text.replace("[name]", self.placeholder_token) + text = text.replace("[filewords]", ' '.join(filename_tokens)) + + return x, text -- cgit v1.2.1 From 6785331e22d6a488fbf5905fab56d7fec867e038 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sun, 2 Oct 2022 22:59:01 +0300 Subject: keep textual inversion dataset latents in CPU memory to save a bit of VRAM --- modules/textual_inversion/dataset.py | 2 ++ 1 file changed, 2 insertions(+) (limited to 'modules/textual_inversion/dataset.py') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index 7e134a08..e8394ff6 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -8,6 +8,7 @@ from torchvision import transforms import random import tqdm +from modules import devices class PersonalizedBase(Dataset): @@ -47,6 +48,7 @@ class PersonalizedBase(Dataset): 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) self.dataset.append((init_latent, filename_tokens)) -- cgit v1.2.1 From 5ef0baf5eaec7f21a1666af424405cbee19f3764 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 4 Oct 2022 08:52:11 +0300 Subject: add support for gelbooru tags in filenames for textual inversion --- modules/textual_inversion/dataset.py | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) (limited to 'modules/textual_inversion/dataset.py') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index e8394ff6..7c44ea5b 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -9,6 +9,9 @@ from torchvision import transforms import random import tqdm from modules import devices +import re + +re_tag = re.compile(r"[a-zA-Z][_\w\d()]+") class PersonalizedBase(Dataset): @@ -38,8 +41,8 @@ class PersonalizedBase(Dataset): image = image.resize((self.width, self.height), PIL.Image.BICUBIC) filename = os.path.basename(path) - filename_tokens = os.path.splitext(filename)[0].replace('_', '-').replace(' ', '-').split('-') - filename_tokens = [token for token in filename_tokens if token.isalpha()] + filename_tokens = os.path.splitext(filename)[0] + filename_tokens = re_tag.findall(filename_tokens) npimage = np.array(image).astype(np.uint8) npimage = (npimage / 127.5 - 1.0).astype(np.float32) -- cgit v1.2.1 From 3110f895b2718a3a25aae419fdf5c87c177ec9f4 Mon Sep 17 00:00:00 2001 From: alg-wiki Date: Mon, 10 Oct 2022 17:07:46 +0900 Subject: Textual Inversion: Added custom training image size and number of repeats per input image in a single epoch --- modules/textual_inversion/dataset.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'modules/textual_inversion/dataset.py') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index 7c44ea5b..acc4ce59 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -15,13 +15,13 @@ re_tag = re.compile(r"[a-zA-Z][_\w\d()]+") class PersonalizedBase(Dataset): - def __init__(self, data_root, size=None, repeats=100, flip_p=0.5, placeholder_token="*", width=512, height=512, model=None, device=None, template_file=None): + def __init__(self, data_root, size, repeats, flip_p=0.5, placeholder_token="*", model=None, device=None, template_file=None): self.placeholder_token = placeholder_token self.size = size - self.width = width - self.height = height + self.width = size + self.height = size self.flip = transforms.RandomHorizontalFlip(p=flip_p) self.dataset = [] -- cgit v1.2.1 From 04c745ea4f81518999927fee5f78500560c25e29 Mon Sep 17 00:00:00 2001 From: alg-wiki Date: Mon, 10 Oct 2022 22:35:35 +0900 Subject: Custom Width and Height --- modules/textual_inversion/dataset.py | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) (limited to 'modules/textual_inversion/dataset.py') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index acc4ce59..bcf772d2 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -15,13 +15,12 @@ re_tag = re.compile(r"[a-zA-Z][_\w\d()]+") class PersonalizedBase(Dataset): - def __init__(self, data_root, size, repeats, flip_p=0.5, placeholder_token="*", model=None, device=None, template_file=None): + def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, device=None, template_file=None): self.placeholder_token = placeholder_token - self.size = size - self.width = size - self.height = size + self.width = width + self.height = height self.flip = transforms.RandomHorizontalFlip(p=flip_p) self.dataset = [] -- cgit v1.2.1 From bc3e183b739913e7be91213a256f038b10eb71e9 Mon Sep 17 00:00:00 2001 From: alg-wiki Date: Tue, 11 Oct 2022 04:30:13 +0900 Subject: Textual Inversion: Preprocess and Training will only pick-up image files --- modules/textual_inversion/dataset.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (limited to 'modules/textual_inversion/dataset.py') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index bcf772d2..d4baf066 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -22,6 +22,7 @@ class PersonalizedBase(Dataset): self.width = width self.height = height self.flip = transforms.RandomHorizontalFlip(p=flip_p) + self.extns = [".jpg",".jpeg",".png"] self.dataset = [] @@ -32,7 +33,7 @@ class PersonalizedBase(Dataset): assert data_root, 'dataset directory not specified' - self.image_paths = [os.path.join(data_root, file_path) for file_path in os.listdir(data_root)] + self.image_paths = [os.path.join(data_root, file_path) for file_path in os.listdir(data_root) if os.path.splitext(file_path.casefold())[1] in self.extns] print("Preparing dataset...") for path in tqdm.tqdm(self.image_paths): image = Image.open(path) -- cgit v1.2.1 From 907a88b2d0be320575c2129d8d6a1d4f3a68f9eb Mon Sep 17 00:00:00 2001 From: alg-wiki Date: Tue, 11 Oct 2022 06:33:08 +0900 Subject: Added .webp .bmp --- modules/textual_inversion/dataset.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/textual_inversion/dataset.py') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index d4baf066..0dc54fb7 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -22,7 +22,7 @@ class PersonalizedBase(Dataset): self.width = width self.height = height self.flip = transforms.RandomHorizontalFlip(p=flip_p) - self.extns = [".jpg",".jpeg",".png"] + self.extns = [".jpg",".jpeg",".png",".webp",".bmp"] self.dataset = [] -- cgit v1.2.1 From b2368a3bce663f19a7209d9cb38617e635ca6e3c Mon Sep 17 00:00:00 2001 From: alg-wiki Date: Tue, 11 Oct 2022 17:32:46 +0900 Subject: Switched to exception handling --- modules/textual_inversion/dataset.py | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) (limited to 'modules/textual_inversion/dataset.py') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index 0dc54fb7..4d006366 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -22,7 +22,6 @@ class PersonalizedBase(Dataset): self.width = width self.height = height self.flip = transforms.RandomHorizontalFlip(p=flip_p) - self.extns = [".jpg",".jpeg",".png",".webp",".bmp"] self.dataset = [] @@ -33,12 +32,13 @@ class PersonalizedBase(Dataset): assert data_root, 'dataset directory not specified' - self.image_paths = [os.path.join(data_root, file_path) for file_path in os.listdir(data_root) if os.path.splitext(file_path.casefold())[1] in self.extns] + self.image_paths = [os.path.join(data_root, file_path) for file_path in os.listdir(data_root)] print("Preparing dataset...") for path in tqdm.tqdm(self.image_paths): - image = Image.open(path) - image = image.convert('RGB') - image = image.resize((self.width, self.height), PIL.Image.BICUBIC) + try: + image = Image.open(path).convert('RGB').resize((self.width, self.height), PIL.Image.BICUBIC) + except Exception: + continue filename = os.path.basename(path) filename_tokens = os.path.splitext(filename)[0] -- cgit v1.2.1 From d4ea5f4d8631f778d11efcde397e4a5b8801d43b Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 11 Oct 2022 19:03:08 +0300 Subject: add an option to unload models during hypernetwork training to save VRAM --- modules/textual_inversion/dataset.py | 29 ++++++++++++++++++++--------- 1 file changed, 20 insertions(+), 9 deletions(-) (limited to 'modules/textual_inversion/dataset.py') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index 4d006366..f61f40d3 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -8,14 +8,14 @@ from torchvision import transforms import random import tqdm -from modules import devices +from modules import devices, shared import re re_tag = re.compile(r"[a-zA-Z][_\w\d()]+") class PersonalizedBase(Dataset): - def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, device=None, template_file=None): + def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, device=None, template_file=None, include_cond=False): self.placeholder_token = placeholder_token @@ -32,6 +32,8 @@ class PersonalizedBase(Dataset): assert data_root, 'dataset directory not specified' + 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)] print("Preparing dataset...") for path in tqdm.tqdm(self.image_paths): @@ -53,7 +55,13 @@ class PersonalizedBase(Dataset): init_latent = model.get_first_stage_encoding(model.encode_first_stage(torchdata.unsqueeze(dim=0))).squeeze() init_latent = init_latent.to(devices.cpu) - self.dataset.append((init_latent, filename_tokens)) + if include_cond: + text = self.create_text(filename_tokens) + cond = cond_model([text]).to(devices.cpu) + else: + cond = None + + self.dataset.append((init_latent, filename_tokens, cond)) self.length = len(self.dataset) * repeats @@ -64,6 +72,12 @@ class PersonalizedBase(Dataset): def shuffle(self): self.indexes = self.initial_indexes[torch.randperm(self.initial_indexes.shape[0])] + def create_text(self, filename_tokens): + text = random.choice(self.lines) + text = text.replace("[name]", self.placeholder_token) + text = text.replace("[filewords]", ' '.join(filename_tokens)) + return text + def __len__(self): return self.length @@ -72,10 +86,7 @@ class PersonalizedBase(Dataset): self.shuffle() index = self.indexes[i % len(self.indexes)] - x, filename_tokens = self.dataset[index] - - text = random.choice(self.lines) - text = text.replace("[name]", self.placeholder_token) - text = text.replace("[filewords]", ' '.join(filename_tokens)) + x, filename_tokens, cond = self.dataset[index] - return x, text + text = self.create_text(filename_tokens) + return x, text, cond -- cgit v1.2.1 From c3c8eef9fd5a0c8b26319e32ca4a19b56204e6df Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Wed, 12 Oct 2022 20:49:47 +0300 Subject: train: change filename processing to be more simple and configurable train: make it possible to make text files with prompts train: rework scheduler so that there's less repeating code in textual inversion and hypernets train: move epochs setting to options --- modules/textual_inversion/dataset.py | 47 ++++++++++++++++++++++++++---------- 1 file changed, 34 insertions(+), 13 deletions(-) (limited to 'modules/textual_inversion/dataset.py') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index f61f40d3..67e90afe 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -11,11 +11,21 @@ import tqdm from modules import devices, shared import re -re_tag = re.compile(r"[a-zA-Z][_\w\d()]+") +re_numbers_at_start = re.compile(r"^[-\d]+\s*") + + +class DatasetEntry: + def __init__(self, filename=None, latent=None, filename_text=None): + self.filename = filename + self.latent = latent + self.filename_text = filename_text + self.cond = None + self.cond_text = None 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): + 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 @@ -42,9 +52,18 @@ class PersonalizedBase(Dataset): except Exception: continue + text_filename = os.path.splitext(path)[0] + ".txt" filename = os.path.basename(path) - filename_tokens = os.path.splitext(filename)[0] - filename_tokens = re_tag.findall(filename_tokens) + + if os.path.exists(text_filename): + with open(text_filename, "r", encoding="utf8") as file: + filename_text = file.read() + else: + filename_text = os.path.splitext(filename)[0] + filename_text = re.sub(re_numbers_at_start, '', filename_text) + if re_word: + tokens = re_word.findall(filename_text) + filename_text = (shared.opts.dataset_filename_join_string or "").join(tokens) npimage = np.array(image).astype(np.uint8) npimage = (npimage / 127.5 - 1.0).astype(np.float32) @@ -55,13 +74,13 @@ class PersonalizedBase(Dataset): 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: - text = self.create_text(filename_tokens) - cond = cond_model([text]).to(devices.cpu) - else: - cond = None + entry.cond_text = self.create_text(filename_text) + entry.cond = cond_model([entry.cond_text]).to(devices.cpu) - self.dataset.append((init_latent, filename_tokens, cond)) + self.dataset.append(entry) self.length = len(self.dataset) * repeats @@ -72,10 +91,10 @@ class PersonalizedBase(Dataset): def shuffle(self): self.indexes = self.initial_indexes[torch.randperm(self.initial_indexes.shape[0])] - def create_text(self, filename_tokens): + def create_text(self, filename_text): text = random.choice(self.lines) text = text.replace("[name]", self.placeholder_token) - text = text.replace("[filewords]", ' '.join(filename_tokens)) + text = text.replace("[filewords]", filename_text) return text def __len__(self): @@ -86,7 +105,9 @@ class PersonalizedBase(Dataset): self.shuffle() index = self.indexes[i % len(self.indexes)] - x, filename_tokens, cond = self.dataset[index] + entry = self.dataset[index] + + if entry.cond is None: + entry.cond_text = self.create_text(entry.filename_text) - text = self.create_text(filename_tokens) - return x, text, cond + return entry -- cgit v1.2.1 From 4d19f3b7d461fe0f63e7ccff936909b0ce0c6126 Mon Sep 17 00:00:00 2001 From: Melan Date: Fri, 14 Oct 2022 22:45:26 +0200 Subject: Raise an assertion error if no training images have been found. --- modules/textual_inversion/dataset.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (limited to 'modules/textual_inversion/dataset.py') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index 67e90afe..12e2f43b 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -81,7 +81,8 @@ class PersonalizedBase(Dataset): entry.cond = cond_model([entry.cond_text]).to(devices.cpu) self.dataset.append(entry) - + + assert len(self.dataset) > 1, "No images have been found in the dataset." self.length = len(self.dataset) * repeats self.initial_indexes = np.arange(self.length) % len(self.dataset) -- cgit v1.2.1 From c7a86f7fe9c0b8967a87e8d709f507d2f44400d8 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sat, 15 Oct 2022 09:24:59 +0300 Subject: add option to use batch size for training --- modules/textual_inversion/dataset.py | 31 +++++++++++++++++++------------ 1 file changed, 19 insertions(+), 12 deletions(-) (limited to 'modules/textual_inversion/dataset.py') diff --git a/modules/textual_inversion/dataset.py b/modules/textual_inversion/dataset.py index 67e90afe..bd99c0cb 100644 --- a/modules/textual_inversion/dataset.py +++ b/modules/textual_inversion/dataset.py @@ -24,11 +24,12 @@ 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): - re_word = re.compile(shared.opts.dataset_filename_word_regex) if len(shared.opts.dataset_filename_word_regex)>0 else None + 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 + self.batch_size = batch_size self.width = width self.height = height self.flip = transforms.RandomHorizontalFlip(p=flip_p) @@ -78,13 +79,13 @@ class PersonalizedBase(Dataset): if include_cond: entry.cond_text = self.create_text(filename_text) - entry.cond = cond_model([entry.cond_text]).to(devices.cpu) + entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0) self.dataset.append(entry) - self.length = len(self.dataset) * repeats + self.length = len(self.dataset) * repeats // batch_size - self.initial_indexes = np.arange(self.length) % len(self.dataset) + self.initial_indexes = np.arange(len(self.dataset)) self.indexes = None self.shuffle() @@ -101,13 +102,19 @@ class PersonalizedBase(Dataset): return self.length def __getitem__(self, i): - if i % len(self.dataset) == 0: - self.shuffle() + res = [] - index = self.indexes[i % len(self.indexes)] - entry = self.dataset[index] + for j in range(self.batch_size): + position = i * self.batch_size + j + if position % len(self.indexes) == 0: + self.shuffle() - if entry.cond is None: - entry.cond_text = self.create_text(entry.filename_text) + index = self.indexes[position % len(self.indexes)] + entry = self.dataset[index] - return entry + if entry.cond is None: + entry.cond_text = self.create_text(entry.filename_text) + + res.append(entry) + + return res -- cgit v1.2.1