From 683287d87f6401083a8d63eedc00ca7410214ca1 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 6 Jan 2023 08:52:06 +0300 Subject: [PATCH] rework saving training params to file #6372 --- modules/hypernetworks/hypernetwork.py | 28 +++++-------------- modules/shared.py | 2 +- modules/textual_inversion/logging.py | 24 ++++++++++++++++ .../textual_inversion/textual_inversion.py | 23 ++------------- 4 files changed, 35 insertions(+), 42 deletions(-) create mode 100644 modules/textual_inversion/logging.py diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 3237c37a..b0cfbe71 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -13,7 +13,7 @@ import tqdm from einops import rearrange, repeat from ldm.util import default from modules import devices, processing, sd_models, shared, sd_samplers -from modules.textual_inversion import textual_inversion +from modules.textual_inversion import textual_inversion, logging from modules.textual_inversion.learn_schedule import LearnRateScheduler from torch import einsum from torch.nn.init import normal_, xavier_normal_, xavier_uniform_, kaiming_normal_, kaiming_uniform_, zeros_ @@ -401,25 +401,7 @@ def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None, hypernet.save(fn) shared.reload_hypernetworks() -# Note: textual_inversion.py has a nearly identical function of the same name. -def save_settings_to_file(model_name, model_hash, initial_step, num_of_dataset_images, hypernetwork_name, layer_structure, activation_func, weight_init, add_layer_norm, use_dropout, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): - # Starting index of preview-related arguments. - border_index = 21 - # Get a list of the argument names. - arg_names = inspect.getfullargspec(save_settings_to_file).args - # Create a list of the argument names to include in the settings string. - names = arg_names[:border_index] # Include all arguments up until the preview-related ones. - if preview_from_txt2img: - names.extend(arg_names[border_index:]) # Include preview-related arguments if applicable. - # Build the settings string. - settings_str = "datetime : " + datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S") + "\n" - for name in names: - if name != 'log_directory': # It's useless and redundant to save log_directory. - value = locals()[name] - settings_str += f"{name}: {value}\n" - # Create or append to the file. - with open(os.path.join(log_directory, 'settings.txt'), "a+") as fout: - fout.write(settings_str + "\n\n") + def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, steps, clip_grad_mode, clip_grad_value, shuffle_tags, tag_drop_out, latent_sampling_method, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): # images allows training previews to have infotext. Importing it at the top causes a circular import problem. @@ -477,7 +459,11 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step, 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=hypernetwork_name, model=shared.sd_model, cond_model=shared.sd_model.cond_stage_model, device=devices.device, template_file=template_file, include_cond=True, batch_size=batch_size, gradient_step=gradient_step, shuffle_tags=shuffle_tags, tag_drop_out=tag_drop_out, latent_sampling_method=latent_sampling_method) if shared.opts.save_training_settings_to_txt: - save_settings_to_file(checkpoint.model_name, '[{}]'.format(checkpoint.hash), initial_step, len(ds), hypernetwork_name, hypernetwork.layer_structure, hypernetwork.activation_func, hypernetwork.weight_init, hypernetwork.add_layer_norm, hypernetwork.use_dropout, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height) + saved_params = dict( + model_name=checkpoint.model_name, model_hash=checkpoint.hash, num_of_dataset_images=len(ds), + **{field: getattr(hypernetwork, field) for field in ['layer_structure', 'activation_func', 'weight_init', 'add_layer_norm', 'use_dropout', ]} + ) + logging.save_settings_to_file(log_directory, {**saved_params, **locals()}) latent_sampling_method = ds.latent_sampling_method diff --git a/modules/shared.py b/modules/shared.py index f0e10b35..57e489d0 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -362,7 +362,7 @@ options_templates.update(options_section(('training', "Training"), { "unload_models_when_training": OptionInfo(False, "Move VAE and CLIP to RAM when training if possible. Saves VRAM."), "pin_memory": OptionInfo(False, "Turn on pin_memory for DataLoader. Makes training slightly faster but can increase memory usage."), "save_optimizer_state": OptionInfo(False, "Saves Optimizer state as separate *.optim file. Training of embedding or HN can be resumed with the matching optim file."), - "save_training_settings_to_txt": OptionInfo(False, "Save textual inversion and hypernet settings to a text file whenever training starts."), + "save_training_settings_to_txt": OptionInfo(True, "Save textual inversion and hypernet settings to a text file whenever training starts."), "dataset_filename_word_regex": OptionInfo("", "Filename word regex"), "dataset_filename_join_string": OptionInfo(" ", "Filename join string"), "training_image_repeats_per_epoch": OptionInfo(1, "Number of repeats for a single input image per epoch; used only for displaying epoch number", gr.Number, {"precision": 0}), diff --git a/modules/textual_inversion/logging.py b/modules/textual_inversion/logging.py new file mode 100644 index 00000000..8b1981d5 --- /dev/null +++ b/modules/textual_inversion/logging.py @@ -0,0 +1,24 @@ +import datetime +import json +import os + +saved_params_shared = {"model_name", "model_hash", "initial_step", "num_of_dataset_images", "learn_rate", "batch_size", "data_root", "log_directory", "training_width", "training_height", "steps", "create_image_every", "template_file"} +saved_params_ti = {"embedding_name", "num_vectors_per_token", "save_embedding_every", "save_image_with_stored_embedding"} +saved_params_hypernet = {"hypernetwork_name", "layer_structure", "activation_func", "weight_init", "add_layer_norm", "use_dropout", "save_hypernetwork_every"} +saved_params_all = saved_params_shared | saved_params_ti | saved_params_hypernet +saved_params_previews = {"preview_prompt", "preview_negative_prompt", "preview_steps", "preview_sampler_index", "preview_cfg_scale", "preview_seed", "preview_width", "preview_height"} + + +def save_settings_to_file(log_directory, all_params): + now = datetime.datetime.now() + params = {"datetime": now.strftime("%Y-%m-%d %H:%M:%S")} + + keys = saved_params_all + if all_params.get('preview_from_txt2img'): + keys = keys | saved_params_previews + + params.update({k: v for k, v in all_params.items() if k in keys}) + + filename = f'settings-{now.strftime("%Y-%m-%d-%H-%M-%S")}.json' + with open(os.path.join(log_directory, filename), "w") as file: + json.dump(params, file, indent=4) diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index e9cf432f..f9f5e8cd 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -18,6 +18,8 @@ from modules.textual_inversion.learn_schedule import LearnRateScheduler from modules.textual_inversion.image_embedding import (embedding_to_b64, embedding_from_b64, insert_image_data_embed, extract_image_data_embed, caption_image_overlay) +from modules.textual_inversion.logging import save_settings_to_file + class Embedding: def __init__(self, vec, name, step=None): @@ -231,25 +233,6 @@ def write_loss(log_directory, filename, step, epoch_len, values): **values, }) -# Note: hypernetwork.py has a nearly identical function of the same name. -def save_settings_to_file(model_name, model_hash, initial_step, num_of_dataset_images, embedding_name, vectors_per_token, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): - # Starting index of preview-related arguments. - border_index = 18 - # Get a list of the argument names. - arg_names = inspect.getfullargspec(save_settings_to_file).args - # Create a list of the argument names to include in the settings string. - names = arg_names[:border_index] # Include all arguments up until the preview-related ones. - if preview_from_txt2img: - names.extend(arg_names[border_index:]) # Include preview-related arguments if applicable. - # Build the settings string. - settings_str = "datetime : " + datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S") + "\n" - for name in names: - if name != 'log_directory': # It's useless and redundant to save log_directory. - value = locals()[name] - settings_str += f"{name}: {value}\n" - # Create or append to the file. - with open(os.path.join(log_directory, 'settings.txt'), "a+") as fout: - fout.write(settings_str + "\n\n") def validate_train_inputs(model_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_model_every, create_image_every, log_directory, name="embedding"): assert model_name, f"{name} not selected" @@ -330,7 +313,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, cond_model=shared.sd_model.cond_stage_model, device=devices.device, template_file=template_file, batch_size=batch_size, gradient_step=gradient_step, shuffle_tags=shuffle_tags, tag_drop_out=tag_drop_out, latent_sampling_method=latent_sampling_method) if shared.opts.save_training_settings_to_txt: - save_settings_to_file(checkpoint.model_name, '[{}]'.format(checkpoint.hash), initial_step, len(ds), embedding_name, len(embedding.vec), learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height) + save_settings_to_file(log_directory, {**dict(model_name=checkpoint.model_name, model_hash=checkpoint.hash, num_of_dataset_images=len(ds), num_vectors_per_token=len(embedding.vec)), **locals()}) latent_sampling_method = ds.latent_sampling_method