From 873efeed49bb5197a42da18272115b326c5d68f3 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 11 Oct 2022 15:51:22 +0300 Subject: rename hypernetwork dir to hypernetworks to prevent clash with an old filename that people who use zip instead of git clone will have --- modules/hypernetworks/hypernetwork.py | 283 ++++++++++++++++++++++++++++++++++ 1 file changed, 283 insertions(+) create mode 100644 modules/hypernetworks/hypernetwork.py (limited to 'modules/hypernetworks/hypernetwork.py') diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py new file mode 100644 index 00000000..aa701bda --- /dev/null +++ b/modules/hypernetworks/hypernetwork.py @@ -0,0 +1,283 @@ +import datetime +import glob +import html +import os +import sys +import traceback +import tqdm + +import torch + +from ldm.util import default +from modules import devices, shared, processing, sd_models +import torch +from torch import einsum +from einops import rearrange, repeat +import modules.textual_inversion.dataset + + +class HypernetworkModule(torch.nn.Module): + def __init__(self, dim, state_dict=None): + super().__init__() + + self.linear1 = torch.nn.Linear(dim, dim * 2) + self.linear2 = torch.nn.Linear(dim * 2, dim) + + if state_dict is not None: + self.load_state_dict(state_dict, strict=True) + else: + + self.linear1.weight.data.normal_(mean=0.0, std=0.01) + self.linear1.bias.data.zero_() + self.linear2.weight.data.normal_(mean=0.0, std=0.01) + self.linear2.bias.data.zero_() + + self.to(devices.device) + + def forward(self, x): + return x + (self.linear2(self.linear1(x))) + + +class Hypernetwork: + filename = None + name = None + + def __init__(self, name=None): + self.filename = None + self.name = name + self.layers = {} + self.step = 0 + self.sd_checkpoint = None + self.sd_checkpoint_name = None + + for size in [320, 640, 768, 1280]: + self.layers[size] = (HypernetworkModule(size), HypernetworkModule(size)) + + def weights(self): + res = [] + + for k, layers in self.layers.items(): + for layer in layers: + layer.train() + res += [layer.linear1.weight, layer.linear1.bias, layer.linear2.weight, layer.linear2.bias] + + return res + + def save(self, filename): + state_dict = {} + + for k, v in self.layers.items(): + state_dict[k] = (v[0].state_dict(), v[1].state_dict()) + + state_dict['step'] = self.step + state_dict['name'] = self.name + state_dict['sd_checkpoint'] = self.sd_checkpoint + state_dict['sd_checkpoint_name'] = self.sd_checkpoint_name + + torch.save(state_dict, filename) + + def load(self, filename): + self.filename = filename + if self.name is None: + self.name = os.path.splitext(os.path.basename(filename))[0] + + state_dict = torch.load(filename, map_location='cpu') + + for size, sd in state_dict.items(): + if type(size) == int: + self.layers[size] = (HypernetworkModule(size, sd[0]), HypernetworkModule(size, sd[1])) + + self.name = state_dict.get('name', self.name) + self.step = state_dict.get('step', 0) + self.sd_checkpoint = state_dict.get('sd_checkpoint', None) + self.sd_checkpoint_name = state_dict.get('sd_checkpoint_name', None) + + +def list_hypernetworks(path): + res = {} + for filename in glob.iglob(os.path.join(path, '**/*.pt'), recursive=True): + name = os.path.splitext(os.path.basename(filename))[0] + res[name] = filename + return res + + +def load_hypernetwork(filename): + path = shared.hypernetworks.get(filename, None) + if path is not None: + print(f"Loading hypernetwork {filename}") + try: + shared.loaded_hypernetwork = Hypernetwork() + shared.loaded_hypernetwork.load(path) + + except Exception: + print(f"Error loading hypernetwork {path}", file=sys.stderr) + print(traceback.format_exc(), file=sys.stderr) + else: + if shared.loaded_hypernetwork is not None: + print(f"Unloading hypernetwork") + + shared.loaded_hypernetwork = None + + +def apply_hypernetwork(hypernetwork, context, layer=None): + hypernetwork_layers = (hypernetwork.layers if hypernetwork is not None else {}).get(context.shape[2], None) + + if hypernetwork_layers is None: + return context, context + + if layer is not None: + layer.hyper_k = hypernetwork_layers[0] + layer.hyper_v = hypernetwork_layers[1] + + context_k = hypernetwork_layers[0](context) + context_v = hypernetwork_layers[1](context) + return context_k, context_v + + +def attention_CrossAttention_forward(self, x, context=None, mask=None): + h = self.heads + + q = self.to_q(x) + context = default(context, x) + + context_k, context_v = apply_hypernetwork(shared.loaded_hypernetwork, context, self) + k = self.to_k(context_k) + v = self.to_v(context_v) + + q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v)) + + sim = einsum('b i d, b j d -> b i j', q, k) * self.scale + + if mask is not None: + mask = rearrange(mask, 'b ... -> b (...)') + max_neg_value = -torch.finfo(sim.dtype).max + mask = repeat(mask, 'b j -> (b h) () j', h=h) + sim.masked_fill_(~mask, max_neg_value) + + # attention, what we cannot get enough of + attn = sim.softmax(dim=-1) + + out = einsum('b i j, b j d -> b i d', attn, v) + out = rearrange(out, '(b h) n d -> b n (h d)', h=h) + return self.to_out(out) + + +def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory, steps, create_image_every, save_hypernetwork_every, template_file, preview_image_prompt): + assert hypernetwork_name, 'embedding not selected' + + path = shared.hypernetworks.get(hypernetwork_name, None) + shared.loaded_hypernetwork = Hypernetwork() + shared.loaded_hypernetwork.load(path) + + shared.state.textinfo = "Initializing hypernetwork training..." + shared.state.job_count = steps + + filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt') + + log_directory = os.path.join(log_directory, datetime.datetime.now().strftime("%Y-%m-%d"), hypernetwork_name) + + if save_hypernetwork_every > 0: + hypernetwork_dir = os.path.join(log_directory, "hypernetworks") + os.makedirs(hypernetwork_dir, exist_ok=True) + else: + hypernetwork_dir = None + + if create_image_every > 0: + images_dir = os.path.join(log_directory, "images") + os.makedirs(images_dir, exist_ok=True) + else: + images_dir = None + + cond_model = shared.sd_model.cond_stage_model + + shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..." + with torch.autocast("cuda"): + ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=512, height=512, repeats=1, placeholder_token=hypernetwork_name, model=shared.sd_model, device=devices.device, template_file=template_file) + + hypernetwork = shared.loaded_hypernetwork + weights = hypernetwork.weights() + for weight in weights: + weight.requires_grad = True + + optimizer = torch.optim.AdamW(weights, lr=learn_rate) + + losses = torch.zeros((32,)) + + last_saved_file = "" + last_saved_image = "" + + ititial_step = hypernetwork.step or 0 + if ititial_step > steps: + return hypernetwork, filename + + pbar = tqdm.tqdm(enumerate(ds), total=steps - ititial_step) + for i, (x, text) in pbar: + hypernetwork.step = i + ititial_step + + if hypernetwork.step > steps: + break + + if shared.state.interrupted: + break + + with torch.autocast("cuda"): + c = cond_model([text]) + + x = x.to(devices.device) + loss = shared.sd_model(x.unsqueeze(0), c)[0] + del x + + losses[hypernetwork.step % losses.shape[0]] = loss.item() + + optimizer.zero_grad() + loss.backward() + optimizer.step() + + pbar.set_description(f"loss: {losses.mean():.7f}") + + if hypernetwork.step > 0 and hypernetwork_dir is not None and hypernetwork.step % save_hypernetwork_every == 0: + last_saved_file = os.path.join(hypernetwork_dir, f'{hypernetwork_name}-{hypernetwork.step}.pt') + hypernetwork.save(last_saved_file) + + if hypernetwork.step > 0 and images_dir is not None and hypernetwork.step % create_image_every == 0: + last_saved_image = os.path.join(images_dir, f'{hypernetwork_name}-{hypernetwork.step}.png') + + preview_text = text if preview_image_prompt == "" else preview_image_prompt + + p = processing.StableDiffusionProcessingTxt2Img( + sd_model=shared.sd_model, + prompt=preview_text, + steps=20, + do_not_save_grid=True, + do_not_save_samples=True, + ) + + processed = processing.process_images(p) + image = processed.images[0] + + shared.state.current_image = image + image.save(last_saved_image) + + last_saved_image += f", prompt: {preview_text}" + + shared.state.job_no = hypernetwork.step + + shared.state.textinfo = f""" +

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

+""" + + checkpoint = sd_models.select_checkpoint() + + hypernetwork.sd_checkpoint = checkpoint.hash + hypernetwork.sd_checkpoint_name = checkpoint.model_name + hypernetwork.save(filename) + + return hypernetwork, filename + + -- cgit v1.2.3 From d682444ecc99319fbd2b142a12727501e2884ba7 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 11 Oct 2022 18:04:47 +0300 Subject: add option to select hypernetwork modules when creating --- modules/hypernetworks/hypernetwork.py | 4 ++-- modules/hypernetworks/ui.py | 4 ++-- modules/ui.py | 2 ++ 3 files changed, 6 insertions(+), 4 deletions(-) (limited to 'modules/hypernetworks/hypernetwork.py') diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index aa701bda..b081f14e 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -42,7 +42,7 @@ class Hypernetwork: filename = None name = None - def __init__(self, name=None): + def __init__(self, name=None, enable_sizes=None): self.filename = None self.name = name self.layers = {} @@ -50,7 +50,7 @@ class Hypernetwork: self.sd_checkpoint = None self.sd_checkpoint_name = None - for size in [320, 640, 768, 1280]: + for size in enable_sizes or [320, 640, 768, 1280]: self.layers[size] = (HypernetworkModule(size), HypernetworkModule(size)) def weights(self): diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py index e7540f41..cdddcce1 100644 --- a/modules/hypernetworks/ui.py +++ b/modules/hypernetworks/ui.py @@ -9,11 +9,11 @@ from modules import sd_hijack, shared from modules.hypernetworks import hypernetwork -def create_hypernetwork(name): +def create_hypernetwork(name, enable_sizes): fn = os.path.join(shared.cmd_opts.hypernetwork_dir, f"{name}.pt") assert not os.path.exists(fn), f"file {fn} already exists" - hypernet = modules.hypernetworks.hypernetwork.Hypernetwork(name=name) + hypernet = modules.hypernetworks.hypernetwork.Hypernetwork(name=name, enable_sizes=[int(x) for x in enable_sizes]) hypernet.save(fn) shared.reload_hypernetworks() diff --git a/modules/ui.py b/modules/ui.py index f2d16b12..14b87b92 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -1037,6 +1037,7 @@ def create_ui(wrap_gradio_gpu_call): gr.HTML(value="

Create a new hypernetwork

") new_hypernetwork_name = gr.Textbox(label="Name") + new_hypernetwork_sizes = gr.CheckboxGroup(label="Modules", value=["768", "320", "640", "1280"], choices=["768", "320", "640", "1280"]) with gr.Row(): with gr.Column(scale=3): @@ -1114,6 +1115,7 @@ def create_ui(wrap_gradio_gpu_call): fn=modules.hypernetworks.ui.create_hypernetwork, inputs=[ new_hypernetwork_name, + new_hypernetwork_sizes, ], outputs=[ train_hypernetwork_name, -- cgit v1.2.3 From d4ea5f4d8631f778d11efcde397e4a5b8801d43b Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 11 Oct 2022 19:03:08 +0300 Subject: add an option to unload models during hypernetwork training to save VRAM --- modules/hypernetworks/hypernetwork.py | 25 +++++++++++++++------- modules/hypernetworks/ui.py | 4 +++- modules/shared.py | 4 ++++ modules/textual_inversion/dataset.py | 29 ++++++++++++++++++-------- modules/textual_inversion/textual_inversion.py | 2 +- 5 files changed, 46 insertions(+), 18 deletions(-) (limited to 'modules/hypernetworks/hypernetwork.py') diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index b081f14e..4700e1ec 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -175,6 +175,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory, filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt') log_directory = os.path.join(log_directory, datetime.datetime.now().strftime("%Y-%m-%d"), hypernetwork_name) + unload = shared.opts.unload_models_when_training if save_hypernetwork_every > 0: hypernetwork_dir = os.path.join(log_directory, "hypernetworks") @@ -188,11 +189,13 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory, else: images_dir = None - cond_model = shared.sd_model.cond_stage_model - shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..." with torch.autocast("cuda"): - ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=512, height=512, repeats=1, placeholder_token=hypernetwork_name, model=shared.sd_model, device=devices.device, template_file=template_file) + ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=512, height=512, repeats=1, placeholder_token=hypernetwork_name, model=shared.sd_model, device=devices.device, template_file=template_file, include_cond=True) + + if unload: + shared.sd_model.cond_stage_model.to(devices.cpu) + shared.sd_model.first_stage_model.to(devices.cpu) hypernetwork = shared.loaded_hypernetwork weights = hypernetwork.weights() @@ -211,7 +214,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory, return hypernetwork, filename pbar = tqdm.tqdm(enumerate(ds), total=steps - ititial_step) - for i, (x, text) in pbar: + for i, (x, text, cond) in pbar: hypernetwork.step = i + ititial_step if hypernetwork.step > steps: @@ -221,11 +224,11 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory, break with torch.autocast("cuda"): - c = cond_model([text]) - + cond = cond.to(devices.device) x = x.to(devices.device) - loss = shared.sd_model(x.unsqueeze(0), c)[0] + loss = shared.sd_model(x.unsqueeze(0), cond)[0] del x + del cond losses[hypernetwork.step % losses.shape[0]] = loss.item() @@ -244,6 +247,10 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory, preview_text = text if preview_image_prompt == "" else preview_image_prompt + optimizer.zero_grad() + shared.sd_model.cond_stage_model.to(devices.device) + shared.sd_model.first_stage_model.to(devices.device) + p = processing.StableDiffusionProcessingTxt2Img( sd_model=shared.sd_model, prompt=preview_text, @@ -255,6 +262,10 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory, processed = processing.process_images(p) image = processed.images[0] + if unload: + shared.sd_model.cond_stage_model.to(devices.cpu) + shared.sd_model.first_stage_model.to(devices.cpu) + shared.state.current_image = image image.save(last_saved_image) diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py index 3541a388..c67facbb 100644 --- a/modules/hypernetworks/ui.py +++ b/modules/hypernetworks/ui.py @@ -5,7 +5,7 @@ import gradio as gr import modules.textual_inversion.textual_inversion import modules.textual_inversion.preprocess -from modules import sd_hijack, shared +from modules import sd_hijack, shared, devices from modules.hypernetworks import hypernetwork @@ -41,5 +41,7 @@ Hypernetwork saved to {html.escape(filename)} raise finally: shared.loaded_hypernetwork = initial_hypernetwork + shared.sd_model.cond_stage_model.to(devices.device) + shared.sd_model.first_stage_model.to(devices.device) sd_hijack.apply_optimizations() diff --git a/modules/shared.py b/modules/shared.py index 20b45f23..c1092ff7 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -228,6 +228,10 @@ options_templates.update(options_section(('system', "System"), { "multiple_tqdm": OptionInfo(True, "Add a second progress bar to the console that shows progress for an entire job."), })) +options_templates.update(options_section(('training', "Training"), { + "unload_models_when_training": OptionInfo(False, "Unload VAE and CLIP form VRAM when training"), +})) + options_templates.update(options_section(('sd', "Stable Diffusion"), { "sd_model_checkpoint": OptionInfo(None, "Stable Diffusion checkpoint", gr.Dropdown, lambda: {"choices": modules.sd_models.checkpoint_tiles()}, show_on_main_page=True), "sd_hypernetwork": OptionInfo("None", "Stable Diffusion finetune hypernetwork", gr.Dropdown, lambda: {"choices": ["None"] + [x for x in hypernetworks.keys()]}), 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 diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index bb05cdc6..35f4bd9e 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -201,7 +201,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini return embedding, filename pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) - for i, (x, text) in pbar: + for i, (x, text, _) in pbar: embedding.step = i + ititial_step if embedding.step > steps: -- cgit v1.2.3 From 6a9ea5b41cf92cd9e980349bb5034439f4e7a58b Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 11 Oct 2022 19:22:30 +0300 Subject: prevent extra modules from being saved/loaded with hypernet --- modules/hypernetworks/hypernetwork.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/hypernetworks/hypernetwork.py') diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 4700e1ec..5608e799 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -50,7 +50,7 @@ class Hypernetwork: self.sd_checkpoint = None self.sd_checkpoint_name = None - for size in enable_sizes or [320, 640, 768, 1280]: + for size in enable_sizes or []: self.layers[size] = (HypernetworkModule(size), HypernetworkModule(size)) def weights(self): -- cgit v1.2.3