diff options
Diffstat (limited to 'modules/textual_inversion/textual_inversion.py')
| -rw-r--r-- | modules/textual_inversion/textual_inversion.py | 143 |
1 files changed, 124 insertions, 19 deletions
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index cd9f3498..529ed3e2 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -6,11 +6,17 @@ import torch import tqdm
import html
import datetime
+import csv
+from PIL import Image, PngImagePlugin
from modules import shared, devices, sd_hijack, processing, sd_models
import modules.textual_inversion.dataset
+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)
class Embedding:
def __init__(self, vec, name, step=None):
@@ -80,7 +86,18 @@ class EmbeddingDatabase: def process_file(path, filename):
name = os.path.splitext(filename)[0]
- data = torch.load(path, map_location="cpu")
+ data = []
+
+ if os.path.splitext(filename.upper())[-1] in ['.PNG', '.WEBP', '.JXL', '.AVIF']:
+ embed_image = Image.open(path)
+ if hasattr(embed_image, 'text') and 'sd-ti-embedding' in embed_image.text:
+ data = embedding_from_b64(embed_image.text['sd-ti-embedding'])
+ name = data.get('name', name)
+ else:
+ data = extract_image_data_embed(embed_image)
+ name = data.get('name', name)
+ else:
+ data = torch.load(path, map_location="cpu")
# textual inversion embeddings
if 'string_to_param' in data:
@@ -120,6 +137,7 @@ class EmbeddingDatabase: continue
print(f"Loaded a total of {len(self.word_embeddings)} textual inversion embeddings.")
+ print("Embeddings:", ', '.join(self.word_embeddings.keys()))
def find_embedding_at_position(self, tokens, offset):
token = tokens[offset]
@@ -135,7 +153,7 @@ class EmbeddingDatabase: return None, None
-def create_embedding(name, num_vectors_per_token, init_text='*'):
+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
@@ -147,7 +165,8 @@ def create_embedding(name, num_vectors_per_token, init_text='*'): vec[i] = embedded[i * int(embedded.shape[0]) // num_vectors_per_token]
fn = os.path.join(shared.cmd_opts.embeddings_dir, f"{name}.pt")
- assert not os.path.exists(fn), f"file {fn} already exists"
+ if not overwrite_old:
+ assert not os.path.exists(fn), f"file {fn} already exists"
embedding = Embedding(vec, name)
embedding.step = 0
@@ -156,7 +175,33 @@ 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 write_loss(log_directory, filename, step, epoch_len, values):
+ if shared.opts.training_write_csv_every == 0:
+ return
+
+ 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
+
+ with open(os.path.join(log_directory, filename), "a+", newline='') as fout:
+ csv_writer = csv.DictWriter(fout, fieldnames=["step", "epoch", "epoch_step", *(values.keys())])
+
+ if write_csv_header:
+ csv_writer.writeheader()
+
+ epoch = step // epoch_len
+ epoch_step = step - epoch * epoch_len
+
+ csv_writer.writerow({
+ "step": step + 1,
+ "epoch": epoch + 1,
+ "epoch_step": epoch_step + 1,
+ **values,
+ })
+
+
+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):
assert embedding_name, 'embedding not selected'
shared.state.textinfo = "Initializing textual inversion training..."
@@ -178,43 +223,51 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, else:
images_dir = None
+ if create_image_every > 0 and save_image_with_stored_embedding:
+ images_embeds_dir = os.path.join(log_directory, "image_embeddings")
+ os.makedirs(images_embeds_dir, exist_ok=True)
+ else:
+ images_embeds_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, 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, 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)
hijack = sd_hijack.model_hijack
embedding = hijack.embedding_db.word_embeddings[embedding_name]
embedding.vec.requires_grad = True
- optimizer = torch.optim.AdamW([embedding.vec], lr=learn_rate)
-
losses = torch.zeros((32,))
last_saved_file = "<none>"
last_saved_image = "<none>"
+ embedding_yet_to_be_embedded = False
ititial_step = embedding.step or 0
if ititial_step > steps:
return embedding, filename
+ scheduler = LearnRateScheduler(learn_rate, steps, ititial_step)
+ optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate)
+
pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step)
- for i, (x, text) in pbar:
+ for i, entries in pbar:
embedding.step = i + ititial_step
- if embedding.step > steps:
+ scheduler.apply(optimizer, embedding.step)
+ if scheduler.finished:
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]
+ 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()
@@ -223,30 +276,83 @@ 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 = embedding.step // len(ds)
+ epoch_step = embedding.step - (epoch_num * len(ds)) + 1
+
+ pbar.set_description(f"[Epoch {epoch_num}: {epoch_step}/{len(ds)}]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')
embedding.save(last_saved_file)
+ 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 embedding.step > 0 and images_dir is not None and embedding.step % create_image_every == 0:
last_saved_image = os.path.join(images_dir, f'{embedding_name}-{embedding.step}.png')
p = processing.StableDiffusionProcessingTxt2Img(
sd_model=shared.sd_model,
- prompt=text,
- steps=20,
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_index = preview_sampler_index
+ 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]
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}-{embedding.step}.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, embedding.step)
+
+ 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
+
image.save(last_saved_image)
- last_saved_image += f", prompt: {text}"
+ last_saved_image += f", prompt: {preview_text}"
shared.state.job_no = embedding.step
@@ -254,7 +360,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, steps, <p>
Loss: {losses.mean():.7f}<br/>
Step: {embedding.step}<br/>
-Last prompt: {html.escape(text)}<br/>
+Last prompt: {html.escape(entries[0].cond_text)}<br/>
Last saved embedding: {html.escape(last_saved_file)}<br/>
Last saved image: {html.escape(last_saved_image)}<br/>
</p>
@@ -268,4 +374,3 @@ Last saved image: {html.escape(last_saved_image)}<br/> embedding.save(filename)
return embedding, filename
-
|