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 +++--- modules/textual_inversion/preprocess.py | 4 ++-- modules/textual_inversion/textual_inversion.py | 15 ++++++++++++--- 3 files changed, 17 insertions(+), 8 deletions(-) (limited to 'modules/textual_inversion') 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 = [] diff --git a/modules/textual_inversion/preprocess.py b/modules/textual_inversion/preprocess.py index f1c002a2..b3de6fd7 100644 --- a/modules/textual_inversion/preprocess.py +++ b/modules/textual_inversion/preprocess.py @@ -7,8 +7,8 @@ import tqdm from modules import shared, images -def preprocess(process_src, process_dst, process_flip, process_split, process_caption): - size = 512 +def preprocess(process_src, process_dst, process_size, process_flip, process_split, process_caption): + size = process_size src = os.path.abspath(process_src) dst = os.path.abspath(process_dst) diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index cd9f3498..e34dc2e8 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -6,6 +6,7 @@ import torch import tqdm import html import datetime +import math from modules import shared, devices, sd_hijack, processing, sd_models @@ -156,7 +157,7 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): return fn -def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, create_image_every, save_embedding_every, template_file): +def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_size, steps, num_repeats, create_image_every, save_embedding_every, template_file): assert embedding_name, 'embedding not selected' shared.state.textinfo = "Initializing textual inversion training..." @@ -182,7 +183,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, 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, size=512, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file) + ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, size=training_size, repeats=num_repeats, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file) hijack = sd_hijack.model_hijack @@ -200,6 +201,9 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, if ititial_step > steps: return embedding, filename + tr_img_len = len([os.path.join(data_root, file_path) for file_path in os.listdir(data_root)]) + epoch_len = (tr_img_len * num_repeats) + tr_img_len + pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step) for i, (x, text) in pbar: embedding.step = i + ititial_step @@ -223,7 +227,10 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, loss.backward() optimizer.step() - pbar.set_description(f"loss: {losses.mean():.7f}") + epoch_num = math.floor(embedding.step / epoch_len) + epoch_step = embedding.step - (epoch_num * epoch_len) + + pbar.set_description(f"[Epoch {epoch_num}: {epoch_step}/{epoch_len}]loss: {losses.mean():.7f}") if embedding.step > 0 and embedding_dir is not None and embedding.step % save_embedding_every == 0: last_saved_file = os.path.join(embedding_dir, f'{embedding_name}-{embedding.step}.pt') @@ -236,6 +243,8 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, sd_model=shared.sd_model, prompt=text, steps=20, + height=training_size, + width=training_size, do_not_save_grid=True, do_not_save_samples=True, ) -- cgit v1.2.1