Upload folder using huggingface_hub
Browse files
febb161d-54ad-40d7-af5a-e0bfbfbdbe1c/metadata.jsonl
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{"file_name": "\uc548\ub1552001.bmp", "prompt": "TOK"}
|
| 2 |
+
{"file_name": "\uc548\ub1553001.bmp", "prompt": "TOK"}
|
| 3 |
+
{"file_name": "\uc548\ub155001.bmp", "prompt": "TOK"}
|
febb161d-54ad-40d7-af5a-e0bfbfbdbe1c/μλ
001.bmp
ADDED
|
Git LFS Details
|
febb161d-54ad-40d7-af5a-e0bfbfbdbe1c/μλ
2001.bmp
ADDED
|
Git LFS Details
|
febb161d-54ad-40d7-af5a-e0bfbfbdbe1c/μλ
3001.bmp
ADDED
|
Git LFS Details
|
requirements.txt
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
peft==0.7.1
|
| 2 |
+
-huggingface_hub
|
| 3 |
+
torch
|
| 4 |
+
git+https://github.com/huggingface/diffusers@ba28006f8b2a0f7ec3b6784695790422b4f80a97
|
| 5 |
+
transformers==4.36.2
|
| 6 |
+
accelerate==0.25.0
|
| 7 |
+
safetensors==0.4.1
|
| 8 |
+
prodigyopt==1.0
|
| 9 |
+
hf-transfer==0.1.4
|
| 10 |
+
huggingface_hub==0.20.3
|
| 11 |
+
git+https://github.com/huggingface/datasets.git@3f149204a2a5948287adcade5e90707aa5207a92
|
script.py
ADDED
|
@@ -0,0 +1,129 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import sys
|
| 2 |
+
import subprocess
|
| 3 |
+
from safetensors.torch import load_file
|
| 4 |
+
from diffusers import AutoPipelineForText2Image
|
| 5 |
+
from datasets import load_dataset
|
| 6 |
+
from huggingface_hub.repocard import RepoCard
|
| 7 |
+
from huggingface_hub import HfApi
|
| 8 |
+
import torch
|
| 9 |
+
import re
|
| 10 |
+
import argparse
|
| 11 |
+
import os
|
| 12 |
+
import zipfile
|
| 13 |
+
|
| 14 |
+
def do_preprocess(class_data_dir):
|
| 15 |
+
print("Unzipping dataset")
|
| 16 |
+
zip_file_path = f"{class_data_dir}/class_images.zip"
|
| 17 |
+
with zipfile.ZipFile(zip_file_path, 'r') as zip_ref:
|
| 18 |
+
zip_ref.extractall(class_data_dir)
|
| 19 |
+
os.remove(zip_file_path)
|
| 20 |
+
|
| 21 |
+
def do_train(script_args):
|
| 22 |
+
# Pass all arguments to trainer.py
|
| 23 |
+
print("Starting training...")
|
| 24 |
+
result = subprocess.run(['python', 'trainer.py'] + script_args)
|
| 25 |
+
if result.returncode != 0:
|
| 26 |
+
raise Exception("Training failed.")
|
| 27 |
+
|
| 28 |
+
def replace_output_dir(text, output_dir, replacement):
|
| 29 |
+
# Define a pattern that matches the output_dir followed by whitespace, '/', new line, or "'"
|
| 30 |
+
# Add system name from HF only in the correct spots
|
| 31 |
+
pattern = rf"{output_dir}(?=[\s/'\n])"
|
| 32 |
+
return re.sub(pattern, replacement, text)
|
| 33 |
+
|
| 34 |
+
def do_inference(dataset_name, output_dir, num_tokens):
|
| 35 |
+
widget_content = []
|
| 36 |
+
try:
|
| 37 |
+
print("Starting inference to generate example images...")
|
| 38 |
+
dataset = load_dataset(dataset_name)
|
| 39 |
+
pipe = AutoPipelineForText2Image.from_pretrained(
|
| 40 |
+
"stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
|
| 41 |
+
)
|
| 42 |
+
pipe = pipe.to("cuda")
|
| 43 |
+
pipe.load_lora_weights(f'{output_dir}/pytorch_lora_weights.safetensors')
|
| 44 |
+
|
| 45 |
+
prompts = dataset["train"]["prompt"]
|
| 46 |
+
if(num_tokens > 0):
|
| 47 |
+
tokens_sequence = ''.join(f'<s{i}>' for i in range(num_tokens))
|
| 48 |
+
tokens_list = [f'<s{i}>' for i in range(num_tokens)]
|
| 49 |
+
|
| 50 |
+
state_dict = load_file(f"{output_dir}/{output_dir}_emb.safetensors")
|
| 51 |
+
pipe.load_textual_inversion(state_dict["clip_l"], token=tokens_list, text_encoder=pipe.text_encoder, tokenizer=pipe.tokenizer)
|
| 52 |
+
pipe.load_textual_inversion(state_dict["clip_g"], token=tokens_list, text_encoder=pipe.text_encoder_2, tokenizer=pipe.tokenizer_2)
|
| 53 |
+
|
| 54 |
+
prompts = [prompt.replace("TOK", tokens_sequence) for prompt in prompts]
|
| 55 |
+
|
| 56 |
+
for i, prompt in enumerate(prompts):
|
| 57 |
+
image = pipe(prompt, num_inference_steps=25, guidance_scale=7.5).images[0]
|
| 58 |
+
filename = f"image-{i}.png"
|
| 59 |
+
image.save(f"{output_dir}/{filename}")
|
| 60 |
+
card_dict = {
|
| 61 |
+
"text": prompt,
|
| 62 |
+
"output": {
|
| 63 |
+
"url": filename
|
| 64 |
+
}
|
| 65 |
+
}
|
| 66 |
+
widget_content.append(card_dict)
|
| 67 |
+
except Exception as e:
|
| 68 |
+
print("Something went wrong with generating images, specifically: ", e)
|
| 69 |
+
|
| 70 |
+
try:
|
| 71 |
+
api = HfApi()
|
| 72 |
+
username = api.whoami()["name"]
|
| 73 |
+
repo_id = api.create_repo(f"{username}/{output_dir}", exist_ok=True, private=True).repo_id
|
| 74 |
+
|
| 75 |
+
with open(f'{output_dir}/README.md', 'r') as file:
|
| 76 |
+
readme_content = file.read()
|
| 77 |
+
|
| 78 |
+
|
| 79 |
+
readme_content = replace_output_dir(readme_content, output_dir, f"{username}/{output_dir}")
|
| 80 |
+
|
| 81 |
+
card = RepoCard(readme_content)
|
| 82 |
+
if widget_content:
|
| 83 |
+
card.data["widget"] = widget_content
|
| 84 |
+
card.save(f'{output_dir}/README.md')
|
| 85 |
+
|
| 86 |
+
print("Starting upload...")
|
| 87 |
+
api.upload_folder(
|
| 88 |
+
folder_path=output_dir,
|
| 89 |
+
repo_id=f"{username}/{output_dir}",
|
| 90 |
+
repo_type="model",
|
| 91 |
+
)
|
| 92 |
+
except Exception as e:
|
| 93 |
+
print("Something went wrong with uploading your model, specificaly: ", e)
|
| 94 |
+
else:
|
| 95 |
+
print("Upload finished!")
|
| 96 |
+
|
| 97 |
+
import sys
|
| 98 |
+
import argparse
|
| 99 |
+
|
| 100 |
+
def main():
|
| 101 |
+
# Capture all arguments except the script name
|
| 102 |
+
script_args = sys.argv[1:]
|
| 103 |
+
|
| 104 |
+
# Create the argument parser
|
| 105 |
+
parser = argparse.ArgumentParser()
|
| 106 |
+
parser.add_argument('--dataset_name', required=True)
|
| 107 |
+
parser.add_argument('--output_dir', required=True)
|
| 108 |
+
parser.add_argument('--num_new_tokens_per_abstraction', type=int, default=0)
|
| 109 |
+
parser.add_argument('--train_text_encoder_ti', action='store_true')
|
| 110 |
+
parser.add_argument('--class_data_dir', help="Name of the class images dataset")
|
| 111 |
+
|
| 112 |
+
# Parse known arguments
|
| 113 |
+
args, _ = parser.parse_known_args(script_args)
|
| 114 |
+
|
| 115 |
+
# Set num_tokens to 0 if '--train_text_encoder_ti' is not present
|
| 116 |
+
if not args.train_text_encoder_ti:
|
| 117 |
+
args.num_new_tokens_per_abstraction = 0
|
| 118 |
+
|
| 119 |
+
# Proceed with training and inference
|
| 120 |
+
if args.class_data_dir:
|
| 121 |
+
do_preprocess(args.class_data_dir)
|
| 122 |
+
print("Pre-processing finished!")
|
| 123 |
+
do_train(script_args)
|
| 124 |
+
print("Training finished!")
|
| 125 |
+
do_inference(args.dataset_name, args.output_dir, args.num_new_tokens_per_abstraction)
|
| 126 |
+
print("All finished!")
|
| 127 |
+
|
| 128 |
+
if __name__ == "__main__":
|
| 129 |
+
main()
|
trainer.py
ADDED
|
@@ -0,0 +1,2136 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python
|
| 2 |
+
# coding=utf-8
|
| 3 |
+
# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
|
| 4 |
+
#
|
| 5 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 6 |
+
# you may not use this file except in compliance with the License.
|
| 7 |
+
# You may obtain a copy of the License at
|
| 8 |
+
#
|
| 9 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 10 |
+
#
|
| 11 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 12 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 13 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 14 |
+
# See the License for the specific language governing permissions and
|
| 15 |
+
|
| 16 |
+
import argparse
|
| 17 |
+
import gc
|
| 18 |
+
import hashlib
|
| 19 |
+
import itertools
|
| 20 |
+
import logging
|
| 21 |
+
import math
|
| 22 |
+
import os
|
| 23 |
+
import re
|
| 24 |
+
import shutil
|
| 25 |
+
import warnings
|
| 26 |
+
from pathlib import Path
|
| 27 |
+
from typing import List, Optional
|
| 28 |
+
|
| 29 |
+
import numpy as np
|
| 30 |
+
import torch
|
| 31 |
+
import torch.nn.functional as F
|
| 32 |
+
|
| 33 |
+
# imports of the TokenEmbeddingsHandler class
|
| 34 |
+
import torch.utils.checkpoint
|
| 35 |
+
import transformers
|
| 36 |
+
from accelerate import Accelerator
|
| 37 |
+
from accelerate.logging import get_logger
|
| 38 |
+
from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration, set_seed
|
| 39 |
+
from huggingface_hub import create_repo, upload_folder
|
| 40 |
+
from packaging import version
|
| 41 |
+
from peft import LoraConfig, set_peft_model_state_dict
|
| 42 |
+
from peft.utils import get_peft_model_state_dict
|
| 43 |
+
from PIL import Image
|
| 44 |
+
from PIL.ImageOps import exif_transpose
|
| 45 |
+
from safetensors.torch import load_file, save_file
|
| 46 |
+
from torch.utils.data import Dataset
|
| 47 |
+
from torchvision import transforms
|
| 48 |
+
from tqdm.auto import tqdm
|
| 49 |
+
from transformers import AutoTokenizer, PretrainedConfig
|
| 50 |
+
|
| 51 |
+
import diffusers
|
| 52 |
+
from diffusers import (
|
| 53 |
+
AutoencoderKL,
|
| 54 |
+
DDPMScheduler,
|
| 55 |
+
DPMSolverMultistepScheduler,
|
| 56 |
+
StableDiffusionXLPipeline,
|
| 57 |
+
UNet2DConditionModel,
|
| 58 |
+
)
|
| 59 |
+
from diffusers.loaders import LoraLoaderMixin
|
| 60 |
+
from diffusers.optimization import get_scheduler
|
| 61 |
+
from diffusers.training_utils import _set_state_dict_into_text_encoder, cast_training_params, compute_snr
|
| 62 |
+
from diffusers.utils import (
|
| 63 |
+
check_min_version,
|
| 64 |
+
convert_all_state_dict_to_peft,
|
| 65 |
+
convert_state_dict_to_diffusers,
|
| 66 |
+
convert_state_dict_to_kohya,
|
| 67 |
+
convert_unet_state_dict_to_peft,
|
| 68 |
+
is_wandb_available,
|
| 69 |
+
)
|
| 70 |
+
from diffusers.utils.import_utils import is_xformers_available
|
| 71 |
+
from diffusers.utils.torch_utils import is_compiled_module
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
|
| 75 |
+
check_min_version("0.26.0.dev0")
|
| 76 |
+
|
| 77 |
+
logger = get_logger(__name__)
|
| 78 |
+
|
| 79 |
+
|
| 80 |
+
def save_model_card(
|
| 81 |
+
repo_id: str,
|
| 82 |
+
images=None,
|
| 83 |
+
base_model=str,
|
| 84 |
+
train_text_encoder=False,
|
| 85 |
+
train_text_encoder_ti=False,
|
| 86 |
+
token_abstraction_dict=None,
|
| 87 |
+
instance_prompt=str,
|
| 88 |
+
validation_prompt=str,
|
| 89 |
+
repo_folder=None,
|
| 90 |
+
vae_path=None,
|
| 91 |
+
):
|
| 92 |
+
img_str = "widget:\n"
|
| 93 |
+
for i, image in enumerate(images):
|
| 94 |
+
image.save(os.path.join(repo_folder, f"image_{i}.png"))
|
| 95 |
+
img_str += f"""
|
| 96 |
+
- text: '{validation_prompt if validation_prompt else ' ' }'
|
| 97 |
+
output:
|
| 98 |
+
url:
|
| 99 |
+
"image_{i}.png"
|
| 100 |
+
"""
|
| 101 |
+
if not images:
|
| 102 |
+
img_str += f"""
|
| 103 |
+
- text: '{instance_prompt}'
|
| 104 |
+
"""
|
| 105 |
+
embeddings_filename = f"{repo_folder}_emb"
|
| 106 |
+
instance_prompt_webui = re.sub(r"<s\d+>", "", re.sub(r"<s\d+>", embeddings_filename, instance_prompt, count=1))
|
| 107 |
+
ti_keys = ", ".join(f'"{match}"' for match in re.findall(r"<s\d+>", instance_prompt))
|
| 108 |
+
if instance_prompt_webui != embeddings_filename:
|
| 109 |
+
instance_prompt_sentence = f"For example, `{instance_prompt_webui}`"
|
| 110 |
+
else:
|
| 111 |
+
instance_prompt_sentence = ""
|
| 112 |
+
trigger_str = f"You should use {instance_prompt} to trigger the image generation."
|
| 113 |
+
diffusers_imports_pivotal = ""
|
| 114 |
+
diffusers_example_pivotal = ""
|
| 115 |
+
webui_example_pivotal = ""
|
| 116 |
+
if train_text_encoder_ti:
|
| 117 |
+
trigger_str = (
|
| 118 |
+
"To trigger image generation of trained concept(or concepts) replace each concept identifier "
|
| 119 |
+
"in you prompt with the new inserted tokens:\n"
|
| 120 |
+
)
|
| 121 |
+
diffusers_imports_pivotal = """from huggingface_hub import hf_hub_download
|
| 122 |
+
from safetensors.torch import load_file
|
| 123 |
+
"""
|
| 124 |
+
diffusers_example_pivotal = f"""embedding_path = hf_hub_download(repo_id='{repo_id}', filename='{embeddings_filename}.safetensors' repo_type="model")
|
| 125 |
+
state_dict = load_file(embedding_path)
|
| 126 |
+
pipeline.load_textual_inversion(state_dict["clip_l"], token=[{ti_keys}], text_encoder=pipeline.text_encoder, tokenizer=pipeline.tokenizer)
|
| 127 |
+
pipeline.load_textual_inversion(state_dict["clip_g"], token=[{ti_keys}], text_encoder=pipeline.text_encoder_2, tokenizer=pipeline.tokenizer_2)
|
| 128 |
+
"""
|
| 129 |
+
webui_example_pivotal = f"""- *Embeddings*: download **[`{embeddings_filename}.safetensors` here πΎ](/{repo_id}/blob/main/{embeddings_filename}.safetensors)**.
|
| 130 |
+
- Place it on it on your `embeddings` folder
|
| 131 |
+
- Use it by adding `{embeddings_filename}` to your prompt. {instance_prompt_sentence}
|
| 132 |
+
(you need both the LoRA and the embeddings as they were trained together for this LoRA)
|
| 133 |
+
"""
|
| 134 |
+
if token_abstraction_dict:
|
| 135 |
+
for key, value in token_abstraction_dict.items():
|
| 136 |
+
tokens = "".join(value)
|
| 137 |
+
trigger_str += f"""
|
| 138 |
+
to trigger concept `{key}` β use `{tokens}` in your prompt \n
|
| 139 |
+
"""
|
| 140 |
+
|
| 141 |
+
yaml = f"""---
|
| 142 |
+
tags:
|
| 143 |
+
- stable-diffusion-xl
|
| 144 |
+
- stable-diffusion-xl-diffusers
|
| 145 |
+
- text-to-image
|
| 146 |
+
- diffusers
|
| 147 |
+
- lora
|
| 148 |
+
- template:sd-lora
|
| 149 |
+
{img_str}
|
| 150 |
+
base_model: {base_model}
|
| 151 |
+
instance_prompt: {instance_prompt}
|
| 152 |
+
license: openrail++
|
| 153 |
+
---
|
| 154 |
+
"""
|
| 155 |
+
|
| 156 |
+
model_card = f"""
|
| 157 |
+
# SDXL LoRA DreamBooth - {repo_id}
|
| 158 |
+
|
| 159 |
+
<Gallery />
|
| 160 |
+
|
| 161 |
+
## Model description
|
| 162 |
+
|
| 163 |
+
### These are {repo_id} LoRA adaption weights for {base_model}.
|
| 164 |
+
|
| 165 |
+
## Download model
|
| 166 |
+
|
| 167 |
+
### Use it with UIs such as AUTOMATIC1111, Comfy UI, SD.Next, Invoke
|
| 168 |
+
|
| 169 |
+
- **LoRA**: download **[`{repo_folder}.safetensors` here πΎ](/{repo_id}/blob/main/{repo_folder}.safetensors)**.
|
| 170 |
+
- Place it on your `models/Lora` folder.
|
| 171 |
+
- On AUTOMATIC1111, load the LoRA by adding `<lora:{repo_folder}:1>` to your prompt. On ComfyUI just [load it as a regular LoRA](https://comfyanonymous.github.io/ComfyUI_examples/lora/).
|
| 172 |
+
{webui_example_pivotal}
|
| 173 |
+
|
| 174 |
+
## Use it with the [𧨠diffusers library](https://github.com/huggingface/diffusers)
|
| 175 |
+
|
| 176 |
+
```py
|
| 177 |
+
from diffusers import AutoPipelineForText2Image
|
| 178 |
+
import torch
|
| 179 |
+
{diffusers_imports_pivotal}
|
| 180 |
+
pipeline = AutoPipelineForText2Image.from_pretrained('stabilityai/stable-diffusion-xl-base-1.0', torch_dtype=torch.float16).to('cuda')
|
| 181 |
+
pipeline.load_lora_weights('{repo_id}', weight_name='pytorch_lora_weights.safetensors')
|
| 182 |
+
{diffusers_example_pivotal}
|
| 183 |
+
image = pipeline('{validation_prompt if validation_prompt else instance_prompt}').images[0]
|
| 184 |
+
```
|
| 185 |
+
|
| 186 |
+
For more details, including weighting, merging and fusing LoRAs, check the [documentation on loading LoRAs in diffusers](https://huggingface.co/docs/diffusers/main/en/using-diffusers/loading_adapters)
|
| 187 |
+
|
| 188 |
+
## Trigger words
|
| 189 |
+
|
| 190 |
+
{trigger_str}
|
| 191 |
+
|
| 192 |
+
## Details
|
| 193 |
+
All [Files & versions](/{repo_id}/tree/main).
|
| 194 |
+
|
| 195 |
+
The weights were trained using [𧨠diffusers Advanced Dreambooth Training Script](https://github.com/huggingface/diffusers/blob/main/examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py).
|
| 196 |
+
|
| 197 |
+
LoRA for the text encoder was enabled. {train_text_encoder}.
|
| 198 |
+
|
| 199 |
+
Pivotal tuning was enabled: {train_text_encoder_ti}.
|
| 200 |
+
|
| 201 |
+
Special VAE used for training: {vae_path}.
|
| 202 |
+
|
| 203 |
+
"""
|
| 204 |
+
with open(os.path.join(repo_folder, "README.md"), "w") as f:
|
| 205 |
+
f.write(yaml + model_card)
|
| 206 |
+
|
| 207 |
+
|
| 208 |
+
def import_model_class_from_model_name_or_path(
|
| 209 |
+
pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
|
| 210 |
+
):
|
| 211 |
+
text_encoder_config = PretrainedConfig.from_pretrained(
|
| 212 |
+
pretrained_model_name_or_path, subfolder=subfolder, revision=revision
|
| 213 |
+
)
|
| 214 |
+
model_class = text_encoder_config.architectures[0]
|
| 215 |
+
|
| 216 |
+
if model_class == "CLIPTextModel":
|
| 217 |
+
from transformers import CLIPTextModel
|
| 218 |
+
|
| 219 |
+
return CLIPTextModel
|
| 220 |
+
elif model_class == "CLIPTextModelWithProjection":
|
| 221 |
+
from transformers import CLIPTextModelWithProjection
|
| 222 |
+
|
| 223 |
+
return CLIPTextModelWithProjection
|
| 224 |
+
else:
|
| 225 |
+
raise ValueError(f"{model_class} is not supported.")
|
| 226 |
+
|
| 227 |
+
|
| 228 |
+
def parse_args(input_args=None):
|
| 229 |
+
parser = argparse.ArgumentParser(description="Simple example of a training script.")
|
| 230 |
+
parser.add_argument(
|
| 231 |
+
"--pretrained_model_name_or_path",
|
| 232 |
+
type=str,
|
| 233 |
+
default=None,
|
| 234 |
+
required=True,
|
| 235 |
+
help="Path to pretrained model or model identifier from huggingface.co/models.",
|
| 236 |
+
)
|
| 237 |
+
parser.add_argument(
|
| 238 |
+
"--pretrained_vae_model_name_or_path",
|
| 239 |
+
type=str,
|
| 240 |
+
default=None,
|
| 241 |
+
help="Path to pretrained VAE model with better numerical stability. More details: https://github.com/huggingface/diffusers/pull/4038.",
|
| 242 |
+
)
|
| 243 |
+
parser.add_argument(
|
| 244 |
+
"--revision",
|
| 245 |
+
type=str,
|
| 246 |
+
default=None,
|
| 247 |
+
required=False,
|
| 248 |
+
help="Revision of pretrained model identifier from huggingface.co/models.",
|
| 249 |
+
)
|
| 250 |
+
parser.add_argument(
|
| 251 |
+
"--variant",
|
| 252 |
+
type=str,
|
| 253 |
+
default=None,
|
| 254 |
+
help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
|
| 255 |
+
)
|
| 256 |
+
parser.add_argument(
|
| 257 |
+
"--dataset_name",
|
| 258 |
+
type=str,
|
| 259 |
+
default=None,
|
| 260 |
+
help=(
|
| 261 |
+
"The name of the Dataset (from the HuggingFace hub) containing the training data of instance images (could be your own, possibly private,"
|
| 262 |
+
" dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
|
| 263 |
+
" or to a folder containing files that π€ Datasets can understand.To load the custom captions, the training set directory needs to follow the structure of a "
|
| 264 |
+
"datasets ImageFolder, containing both the images and the corresponding caption for each image. see: "
|
| 265 |
+
"https://huggingface.co/docs/datasets/image_dataset for more information"
|
| 266 |
+
),
|
| 267 |
+
)
|
| 268 |
+
parser.add_argument(
|
| 269 |
+
"--dataset_config_name",
|
| 270 |
+
type=str,
|
| 271 |
+
default=None,
|
| 272 |
+
help="The config of the Dataset. In some cases, a dataset may have more than one configuration (for example "
|
| 273 |
+
"if it contains different subsets of data within, and you only wish to load a specific subset - in that case specify the desired configuration using --dataset_config_name. Leave as "
|
| 274 |
+
"None if there's only one config.",
|
| 275 |
+
)
|
| 276 |
+
parser.add_argument(
|
| 277 |
+
"--instance_data_dir",
|
| 278 |
+
type=str,
|
| 279 |
+
default=None,
|
| 280 |
+
help="A path to local folder containing the training data of instance images. Specify this arg instead of "
|
| 281 |
+
"--dataset_name if you wish to train using a local folder without custom captions. If you wish to train with custom captions please specify "
|
| 282 |
+
"--dataset_name instead.",
|
| 283 |
+
)
|
| 284 |
+
|
| 285 |
+
parser.add_argument(
|
| 286 |
+
"--cache_dir",
|
| 287 |
+
type=str,
|
| 288 |
+
default=None,
|
| 289 |
+
help="The directory where the downloaded models and datasets will be stored.",
|
| 290 |
+
)
|
| 291 |
+
|
| 292 |
+
parser.add_argument(
|
| 293 |
+
"--image_column",
|
| 294 |
+
type=str,
|
| 295 |
+
default="image",
|
| 296 |
+
help="The column of the dataset containing the target image. By "
|
| 297 |
+
"default, the standard Image Dataset maps out 'file_name' "
|
| 298 |
+
"to 'image'.",
|
| 299 |
+
)
|
| 300 |
+
parser.add_argument(
|
| 301 |
+
"--caption_column",
|
| 302 |
+
type=str,
|
| 303 |
+
default=None,
|
| 304 |
+
help="The column of the dataset containing the instance prompt for each image",
|
| 305 |
+
)
|
| 306 |
+
|
| 307 |
+
parser.add_argument("--repeats", type=int, default=1, help="How many times to repeat the training data.")
|
| 308 |
+
|
| 309 |
+
parser.add_argument(
|
| 310 |
+
"--class_data_dir",
|
| 311 |
+
type=str,
|
| 312 |
+
default=None,
|
| 313 |
+
required=False,
|
| 314 |
+
help="A folder containing the training data of class images.",
|
| 315 |
+
)
|
| 316 |
+
parser.add_argument(
|
| 317 |
+
"--instance_prompt",
|
| 318 |
+
type=str,
|
| 319 |
+
default=None,
|
| 320 |
+
required=True,
|
| 321 |
+
help="The prompt with identifier specifying the instance, e.g. 'photo of a TOK dog', 'in the style of TOK'",
|
| 322 |
+
)
|
| 323 |
+
parser.add_argument(
|
| 324 |
+
"--token_abstraction",
|
| 325 |
+
type=str,
|
| 326 |
+
default="TOK",
|
| 327 |
+
help="identifier specifying the instance(or instances) as used in instance_prompt, validation prompt, "
|
| 328 |
+
"captions - e.g. TOK. To use multiple identifiers, please specify them in a comma seperated string - e.g. "
|
| 329 |
+
"'TOK,TOK2,TOK3' etc.",
|
| 330 |
+
)
|
| 331 |
+
|
| 332 |
+
parser.add_argument(
|
| 333 |
+
"--num_new_tokens_per_abstraction",
|
| 334 |
+
type=int,
|
| 335 |
+
default=2,
|
| 336 |
+
help="number of new tokens inserted to the tokenizers per token_abstraction identifier when "
|
| 337 |
+
"--train_text_encoder_ti = True. By default, each --token_abstraction (e.g. TOK) is mapped to 2 new "
|
| 338 |
+
"tokens - <si><si+1> ",
|
| 339 |
+
)
|
| 340 |
+
|
| 341 |
+
parser.add_argument(
|
| 342 |
+
"--class_prompt",
|
| 343 |
+
type=str,
|
| 344 |
+
default=None,
|
| 345 |
+
help="The prompt to specify images in the same class as provided instance images.",
|
| 346 |
+
)
|
| 347 |
+
parser.add_argument(
|
| 348 |
+
"--validation_prompt",
|
| 349 |
+
type=str,
|
| 350 |
+
default=None,
|
| 351 |
+
help="A prompt that is used during validation to verify that the model is learning.",
|
| 352 |
+
)
|
| 353 |
+
parser.add_argument(
|
| 354 |
+
"--num_validation_images",
|
| 355 |
+
type=int,
|
| 356 |
+
default=4,
|
| 357 |
+
help="Number of images that should be generated during validation with `validation_prompt`.",
|
| 358 |
+
)
|
| 359 |
+
parser.add_argument(
|
| 360 |
+
"--validation_epochs",
|
| 361 |
+
type=int,
|
| 362 |
+
default=50,
|
| 363 |
+
help=(
|
| 364 |
+
"Run dreambooth validation every X epochs. Dreambooth validation consists of running the prompt"
|
| 365 |
+
" `args.validation_prompt` multiple times: `args.num_validation_images`."
|
| 366 |
+
),
|
| 367 |
+
)
|
| 368 |
+
parser.add_argument(
|
| 369 |
+
"--with_prior_preservation",
|
| 370 |
+
default=False,
|
| 371 |
+
action="store_true",
|
| 372 |
+
help="Flag to add prior preservation loss.",
|
| 373 |
+
)
|
| 374 |
+
parser.add_argument("--prior_loss_weight", type=float, default=1.0, help="The weight of prior preservation loss.")
|
| 375 |
+
parser.add_argument(
|
| 376 |
+
"--num_class_images",
|
| 377 |
+
type=int,
|
| 378 |
+
default=100,
|
| 379 |
+
help=(
|
| 380 |
+
"Minimal class images for prior preservation loss. If there are not enough images already present in"
|
| 381 |
+
" class_data_dir, additional images will be sampled with class_prompt."
|
| 382 |
+
),
|
| 383 |
+
)
|
| 384 |
+
parser.add_argument(
|
| 385 |
+
"--output_dir",
|
| 386 |
+
type=str,
|
| 387 |
+
default="lora-dreambooth-model",
|
| 388 |
+
help="The output directory where the model predictions and checkpoints will be written.",
|
| 389 |
+
)
|
| 390 |
+
parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
|
| 391 |
+
parser.add_argument(
|
| 392 |
+
"--resolution",
|
| 393 |
+
type=int,
|
| 394 |
+
default=1024,
|
| 395 |
+
help=(
|
| 396 |
+
"The resolution for input images, all the images in the train/validation dataset will be resized to this"
|
| 397 |
+
" resolution"
|
| 398 |
+
),
|
| 399 |
+
)
|
| 400 |
+
parser.add_argument(
|
| 401 |
+
"--crops_coords_top_left_h",
|
| 402 |
+
type=int,
|
| 403 |
+
default=0,
|
| 404 |
+
help=("Coordinate for (the height) to be included in the crop coordinate embeddings needed by SDXL UNet."),
|
| 405 |
+
)
|
| 406 |
+
parser.add_argument(
|
| 407 |
+
"--crops_coords_top_left_w",
|
| 408 |
+
type=int,
|
| 409 |
+
default=0,
|
| 410 |
+
help=("Coordinate for (the height) to be included in the crop coordinate embeddings needed by SDXL UNet."),
|
| 411 |
+
)
|
| 412 |
+
parser.add_argument(
|
| 413 |
+
"--center_crop",
|
| 414 |
+
default=False,
|
| 415 |
+
action="store_true",
|
| 416 |
+
help=(
|
| 417 |
+
"Whether to center crop the input images to the resolution. If not set, the images will be randomly"
|
| 418 |
+
" cropped. The images will be resized to the resolution first before cropping."
|
| 419 |
+
),
|
| 420 |
+
)
|
| 421 |
+
parser.add_argument(
|
| 422 |
+
"--train_text_encoder",
|
| 423 |
+
action="store_true",
|
| 424 |
+
help="Whether to train the text encoder. If set, the text encoder should be float32 precision.",
|
| 425 |
+
)
|
| 426 |
+
parser.add_argument(
|
| 427 |
+
"--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader."
|
| 428 |
+
)
|
| 429 |
+
parser.add_argument(
|
| 430 |
+
"--sample_batch_size", type=int, default=4, help="Batch size (per device) for sampling images."
|
| 431 |
+
)
|
| 432 |
+
parser.add_argument("--num_train_epochs", type=int, default=1)
|
| 433 |
+
parser.add_argument(
|
| 434 |
+
"--max_train_steps",
|
| 435 |
+
type=int,
|
| 436 |
+
default=None,
|
| 437 |
+
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
|
| 438 |
+
)
|
| 439 |
+
parser.add_argument(
|
| 440 |
+
"--checkpointing_steps",
|
| 441 |
+
type=int,
|
| 442 |
+
default=500,
|
| 443 |
+
help=(
|
| 444 |
+
"Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
|
| 445 |
+
" checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
|
| 446 |
+
" training using `--resume_from_checkpoint`."
|
| 447 |
+
),
|
| 448 |
+
)
|
| 449 |
+
parser.add_argument(
|
| 450 |
+
"--checkpoints_total_limit",
|
| 451 |
+
type=int,
|
| 452 |
+
default=None,
|
| 453 |
+
help=("Max number of checkpoints to store."),
|
| 454 |
+
)
|
| 455 |
+
parser.add_argument(
|
| 456 |
+
"--resume_from_checkpoint",
|
| 457 |
+
type=str,
|
| 458 |
+
default=None,
|
| 459 |
+
help=(
|
| 460 |
+
"Whether training should be resumed from a previous checkpoint. Use a path saved by"
|
| 461 |
+
' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
|
| 462 |
+
),
|
| 463 |
+
)
|
| 464 |
+
parser.add_argument(
|
| 465 |
+
"--gradient_accumulation_steps",
|
| 466 |
+
type=int,
|
| 467 |
+
default=1,
|
| 468 |
+
help="Number of updates steps to accumulate before performing a backward/update pass.",
|
| 469 |
+
)
|
| 470 |
+
parser.add_argument(
|
| 471 |
+
"--gradient_checkpointing",
|
| 472 |
+
action="store_true",
|
| 473 |
+
help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
|
| 474 |
+
)
|
| 475 |
+
parser.add_argument(
|
| 476 |
+
"--learning_rate",
|
| 477 |
+
type=float,
|
| 478 |
+
default=1e-4,
|
| 479 |
+
help="Initial learning rate (after the potential warmup period) to use.",
|
| 480 |
+
)
|
| 481 |
+
|
| 482 |
+
parser.add_argument(
|
| 483 |
+
"--text_encoder_lr",
|
| 484 |
+
type=float,
|
| 485 |
+
default=5e-6,
|
| 486 |
+
help="Text encoder learning rate to use.",
|
| 487 |
+
)
|
| 488 |
+
parser.add_argument(
|
| 489 |
+
"--scale_lr",
|
| 490 |
+
action="store_true",
|
| 491 |
+
default=False,
|
| 492 |
+
help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
|
| 493 |
+
)
|
| 494 |
+
parser.add_argument(
|
| 495 |
+
"--lr_scheduler",
|
| 496 |
+
type=str,
|
| 497 |
+
default="constant",
|
| 498 |
+
help=(
|
| 499 |
+
'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
|
| 500 |
+
' "constant", "constant_with_warmup"]'
|
| 501 |
+
),
|
| 502 |
+
)
|
| 503 |
+
|
| 504 |
+
parser.add_argument(
|
| 505 |
+
"--snr_gamma",
|
| 506 |
+
type=float,
|
| 507 |
+
default=None,
|
| 508 |
+
help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
|
| 509 |
+
"More details here: https://arxiv.org/abs/2303.09556.",
|
| 510 |
+
)
|
| 511 |
+
parser.add_argument(
|
| 512 |
+
"--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
|
| 513 |
+
)
|
| 514 |
+
parser.add_argument(
|
| 515 |
+
"--lr_num_cycles",
|
| 516 |
+
type=int,
|
| 517 |
+
default=1,
|
| 518 |
+
help="Number of hard resets of the lr in cosine_with_restarts scheduler.",
|
| 519 |
+
)
|
| 520 |
+
parser.add_argument("--lr_power", type=float, default=1.0, help="Power factor of the polynomial scheduler.")
|
| 521 |
+
parser.add_argument(
|
| 522 |
+
"--dataloader_num_workers",
|
| 523 |
+
type=int,
|
| 524 |
+
default=0,
|
| 525 |
+
help=(
|
| 526 |
+
"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
|
| 527 |
+
),
|
| 528 |
+
)
|
| 529 |
+
|
| 530 |
+
parser.add_argument(
|
| 531 |
+
"--train_text_encoder_ti",
|
| 532 |
+
action="store_true",
|
| 533 |
+
help=("Whether to use textual inversion"),
|
| 534 |
+
)
|
| 535 |
+
|
| 536 |
+
parser.add_argument(
|
| 537 |
+
"--train_text_encoder_ti_frac",
|
| 538 |
+
type=float,
|
| 539 |
+
default=0.5,
|
| 540 |
+
help=("The percentage of epochs to perform textual inversion"),
|
| 541 |
+
)
|
| 542 |
+
|
| 543 |
+
parser.add_argument(
|
| 544 |
+
"--train_text_encoder_frac",
|
| 545 |
+
type=float,
|
| 546 |
+
default=1.0,
|
| 547 |
+
help=("The percentage of epochs to perform text encoder tuning"),
|
| 548 |
+
)
|
| 549 |
+
|
| 550 |
+
parser.add_argument(
|
| 551 |
+
"--optimizer",
|
| 552 |
+
type=str,
|
| 553 |
+
default="adamW",
|
| 554 |
+
help=('The optimizer type to use. Choose between ["AdamW", "prodigy"]'),
|
| 555 |
+
)
|
| 556 |
+
|
| 557 |
+
parser.add_argument(
|
| 558 |
+
"--use_8bit_adam",
|
| 559 |
+
action="store_true",
|
| 560 |
+
help="Whether or not to use 8-bit Adam from bitsandbytes. Ignored if optimizer is not set to AdamW",
|
| 561 |
+
)
|
| 562 |
+
|
| 563 |
+
parser.add_argument(
|
| 564 |
+
"--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam and Prodigy optimizers."
|
| 565 |
+
)
|
| 566 |
+
parser.add_argument(
|
| 567 |
+
"--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam and Prodigy optimizers."
|
| 568 |
+
)
|
| 569 |
+
parser.add_argument(
|
| 570 |
+
"--prodigy_beta3",
|
| 571 |
+
type=float,
|
| 572 |
+
default=None,
|
| 573 |
+
help="coefficients for computing the Prodidy stepsize using running averages. If set to None, "
|
| 574 |
+
"uses the value of square root of beta2. Ignored if optimizer is adamW",
|
| 575 |
+
)
|
| 576 |
+
parser.add_argument("--prodigy_decouple", type=bool, default=True, help="Use AdamW style decoupled weight decay")
|
| 577 |
+
parser.add_argument("--adam_weight_decay", type=float, default=1e-04, help="Weight decay to use for unet params")
|
| 578 |
+
parser.add_argument(
|
| 579 |
+
"--adam_weight_decay_text_encoder", type=float, default=None, help="Weight decay to use for text_encoder"
|
| 580 |
+
)
|
| 581 |
+
|
| 582 |
+
parser.add_argument(
|
| 583 |
+
"--adam_epsilon",
|
| 584 |
+
type=float,
|
| 585 |
+
default=1e-08,
|
| 586 |
+
help="Epsilon value for the Adam optimizer and Prodigy optimizers.",
|
| 587 |
+
)
|
| 588 |
+
|
| 589 |
+
parser.add_argument(
|
| 590 |
+
"--prodigy_use_bias_correction",
|
| 591 |
+
type=bool,
|
| 592 |
+
default=True,
|
| 593 |
+
help="Turn on Adam's bias correction. True by default. Ignored if optimizer is adamW",
|
| 594 |
+
)
|
| 595 |
+
parser.add_argument(
|
| 596 |
+
"--prodigy_safeguard_warmup",
|
| 597 |
+
type=bool,
|
| 598 |
+
default=True,
|
| 599 |
+
help="Remove lr from the denominator of D estimate to avoid issues during warm-up stage. True by default. "
|
| 600 |
+
"Ignored if optimizer is adamW",
|
| 601 |
+
)
|
| 602 |
+
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
|
| 603 |
+
parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
|
| 604 |
+
parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
|
| 605 |
+
parser.add_argument(
|
| 606 |
+
"--hub_model_id",
|
| 607 |
+
type=str,
|
| 608 |
+
default=None,
|
| 609 |
+
help="The name of the repository to keep in sync with the local `output_dir`.",
|
| 610 |
+
)
|
| 611 |
+
parser.add_argument(
|
| 612 |
+
"--logging_dir",
|
| 613 |
+
type=str,
|
| 614 |
+
default="logs",
|
| 615 |
+
help=(
|
| 616 |
+
"[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
|
| 617 |
+
" *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
|
| 618 |
+
),
|
| 619 |
+
)
|
| 620 |
+
parser.add_argument(
|
| 621 |
+
"--allow_tf32",
|
| 622 |
+
action="store_true",
|
| 623 |
+
help=(
|
| 624 |
+
"Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
|
| 625 |
+
" https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
|
| 626 |
+
),
|
| 627 |
+
)
|
| 628 |
+
parser.add_argument(
|
| 629 |
+
"--report_to",
|
| 630 |
+
type=str,
|
| 631 |
+
default="tensorboard",
|
| 632 |
+
help=(
|
| 633 |
+
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
|
| 634 |
+
' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
|
| 635 |
+
),
|
| 636 |
+
)
|
| 637 |
+
parser.add_argument(
|
| 638 |
+
"--mixed_precision",
|
| 639 |
+
type=str,
|
| 640 |
+
default=None,
|
| 641 |
+
choices=["no", "fp16", "bf16"],
|
| 642 |
+
help=(
|
| 643 |
+
"Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
|
| 644 |
+
" 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the"
|
| 645 |
+
" flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
|
| 646 |
+
),
|
| 647 |
+
)
|
| 648 |
+
parser.add_argument(
|
| 649 |
+
"--prior_generation_precision",
|
| 650 |
+
type=str,
|
| 651 |
+
default=None,
|
| 652 |
+
choices=["no", "fp32", "fp16", "bf16"],
|
| 653 |
+
help=(
|
| 654 |
+
"Choose prior generation precision between fp32, fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
|
| 655 |
+
" 1.10.and an Nvidia Ampere GPU. Default to fp16 if a GPU is available else fp32."
|
| 656 |
+
),
|
| 657 |
+
)
|
| 658 |
+
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
|
| 659 |
+
parser.add_argument(
|
| 660 |
+
"--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
|
| 661 |
+
)
|
| 662 |
+
parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
|
| 663 |
+
parser.add_argument(
|
| 664 |
+
"--rank",
|
| 665 |
+
type=int,
|
| 666 |
+
default=4,
|
| 667 |
+
help=("The dimension of the LoRA update matrices."),
|
| 668 |
+
)
|
| 669 |
+
parser.add_argument(
|
| 670 |
+
"--cache_latents",
|
| 671 |
+
action="store_true",
|
| 672 |
+
default=False,
|
| 673 |
+
help="Cache the VAE latents",
|
| 674 |
+
)
|
| 675 |
+
|
| 676 |
+
if input_args is not None:
|
| 677 |
+
args = parser.parse_args(input_args)
|
| 678 |
+
else:
|
| 679 |
+
args = parser.parse_args()
|
| 680 |
+
|
| 681 |
+
if args.dataset_name is None and args.instance_data_dir is None:
|
| 682 |
+
raise ValueError("Specify either `--dataset_name` or `--instance_data_dir`")
|
| 683 |
+
|
| 684 |
+
if args.dataset_name is not None and args.instance_data_dir is not None:
|
| 685 |
+
raise ValueError("Specify only one of `--dataset_name` or `--instance_data_dir`")
|
| 686 |
+
|
| 687 |
+
if args.train_text_encoder and args.train_text_encoder_ti:
|
| 688 |
+
raise ValueError(
|
| 689 |
+
"Specify only one of `--train_text_encoder` or `--train_text_encoder_ti. "
|
| 690 |
+
"For full LoRA text encoder training check --train_text_encoder, for textual "
|
| 691 |
+
"inversion training check `--train_text_encoder_ti`"
|
| 692 |
+
)
|
| 693 |
+
|
| 694 |
+
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
|
| 695 |
+
if env_local_rank != -1 and env_local_rank != args.local_rank:
|
| 696 |
+
args.local_rank = env_local_rank
|
| 697 |
+
|
| 698 |
+
if args.with_prior_preservation:
|
| 699 |
+
if args.class_data_dir is None:
|
| 700 |
+
raise ValueError("You must specify a data directory for class images.")
|
| 701 |
+
if args.class_prompt is None:
|
| 702 |
+
raise ValueError("You must specify prompt for class images.")
|
| 703 |
+
else:
|
| 704 |
+
# logger is not available yet
|
| 705 |
+
if args.class_data_dir is not None:
|
| 706 |
+
warnings.warn("You need not use --class_data_dir without --with_prior_preservation.")
|
| 707 |
+
if args.class_prompt is not None:
|
| 708 |
+
warnings.warn("You need not use --class_prompt without --with_prior_preservation.")
|
| 709 |
+
|
| 710 |
+
return args
|
| 711 |
+
|
| 712 |
+
|
| 713 |
+
# Taken from https://github.com/replicate/cog-sdxl/blob/main/dataset_and_utils.py
|
| 714 |
+
class TokenEmbeddingsHandler:
|
| 715 |
+
def __init__(self, text_encoders, tokenizers):
|
| 716 |
+
self.text_encoders = text_encoders
|
| 717 |
+
self.tokenizers = tokenizers
|
| 718 |
+
|
| 719 |
+
self.train_ids: Optional[torch.Tensor] = None
|
| 720 |
+
self.inserting_toks: Optional[List[str]] = None
|
| 721 |
+
self.embeddings_settings = {}
|
| 722 |
+
|
| 723 |
+
def initialize_new_tokens(self, inserting_toks: List[str]):
|
| 724 |
+
idx = 0
|
| 725 |
+
for tokenizer, text_encoder in zip(self.tokenizers, self.text_encoders):
|
| 726 |
+
assert isinstance(inserting_toks, list), "inserting_toks should be a list of strings."
|
| 727 |
+
assert all(
|
| 728 |
+
isinstance(tok, str) for tok in inserting_toks
|
| 729 |
+
), "All elements in inserting_toks should be strings."
|
| 730 |
+
|
| 731 |
+
self.inserting_toks = inserting_toks
|
| 732 |
+
special_tokens_dict = {"additional_special_tokens": self.inserting_toks}
|
| 733 |
+
tokenizer.add_special_tokens(special_tokens_dict)
|
| 734 |
+
text_encoder.resize_token_embeddings(len(tokenizer))
|
| 735 |
+
|
| 736 |
+
self.train_ids = tokenizer.convert_tokens_to_ids(self.inserting_toks)
|
| 737 |
+
|
| 738 |
+
# random initialization of new tokens
|
| 739 |
+
std_token_embedding = text_encoder.text_model.embeddings.token_embedding.weight.data.std()
|
| 740 |
+
|
| 741 |
+
print(f"{idx} text encodedr's std_token_embedding: {std_token_embedding}")
|
| 742 |
+
|
| 743 |
+
text_encoder.text_model.embeddings.token_embedding.weight.data[self.train_ids] = (
|
| 744 |
+
torch.randn(len(self.train_ids), text_encoder.text_model.config.hidden_size)
|
| 745 |
+
.to(device=self.device)
|
| 746 |
+
.to(dtype=self.dtype)
|
| 747 |
+
* std_token_embedding
|
| 748 |
+
)
|
| 749 |
+
self.embeddings_settings[
|
| 750 |
+
f"original_embeddings_{idx}"
|
| 751 |
+
] = text_encoder.text_model.embeddings.token_embedding.weight.data.clone()
|
| 752 |
+
self.embeddings_settings[f"std_token_embedding_{idx}"] = std_token_embedding
|
| 753 |
+
|
| 754 |
+
inu = torch.ones((len(tokenizer),), dtype=torch.bool)
|
| 755 |
+
inu[self.train_ids] = False
|
| 756 |
+
|
| 757 |
+
self.embeddings_settings[f"index_no_updates_{idx}"] = inu
|
| 758 |
+
|
| 759 |
+
print(self.embeddings_settings[f"index_no_updates_{idx}"].shape)
|
| 760 |
+
|
| 761 |
+
idx += 1
|
| 762 |
+
|
| 763 |
+
def save_embeddings(self, file_path: str):
|
| 764 |
+
assert self.train_ids is not None, "Initialize new tokens before saving embeddings."
|
| 765 |
+
tensors = {}
|
| 766 |
+
# text_encoder_0 - CLIP ViT-L/14, text_encoder_1 - CLIP ViT-G/14
|
| 767 |
+
idx_to_text_encoder_name = {0: "clip_l", 1: "clip_g"}
|
| 768 |
+
for idx, text_encoder in enumerate(self.text_encoders):
|
| 769 |
+
assert text_encoder.text_model.embeddings.token_embedding.weight.data.shape[0] == len(
|
| 770 |
+
self.tokenizers[0]
|
| 771 |
+
), "Tokenizers should be the same."
|
| 772 |
+
new_token_embeddings = text_encoder.text_model.embeddings.token_embedding.weight.data[self.train_ids]
|
| 773 |
+
|
| 774 |
+
# New tokens for each text encoder are saved under "clip_l" (for text_encoder 0), "clip_g" (for
|
| 775 |
+
# text_encoder 1) to keep compatible with the ecosystem.
|
| 776 |
+
# Note: When loading with diffusers, any name can work - simply specify in inference
|
| 777 |
+
tensors[idx_to_text_encoder_name[idx]] = new_token_embeddings
|
| 778 |
+
# tensors[f"text_encoders_{idx}"] = new_token_embeddings
|
| 779 |
+
|
| 780 |
+
save_file(tensors, file_path)
|
| 781 |
+
|
| 782 |
+
@property
|
| 783 |
+
def dtype(self):
|
| 784 |
+
return self.text_encoders[0].dtype
|
| 785 |
+
|
| 786 |
+
@property
|
| 787 |
+
def device(self):
|
| 788 |
+
return self.text_encoders[0].device
|
| 789 |
+
|
| 790 |
+
@torch.no_grad()
|
| 791 |
+
def retract_embeddings(self):
|
| 792 |
+
for idx, text_encoder in enumerate(self.text_encoders):
|
| 793 |
+
index_no_updates = self.embeddings_settings[f"index_no_updates_{idx}"]
|
| 794 |
+
text_encoder.text_model.embeddings.token_embedding.weight.data[index_no_updates] = (
|
| 795 |
+
self.embeddings_settings[f"original_embeddings_{idx}"][index_no_updates]
|
| 796 |
+
.to(device=text_encoder.device)
|
| 797 |
+
.to(dtype=text_encoder.dtype)
|
| 798 |
+
)
|
| 799 |
+
|
| 800 |
+
# for the parts that were updated, we need to normalize them
|
| 801 |
+
# to have the same std as before
|
| 802 |
+
std_token_embedding = self.embeddings_settings[f"std_token_embedding_{idx}"]
|
| 803 |
+
|
| 804 |
+
index_updates = ~index_no_updates
|
| 805 |
+
new_embeddings = text_encoder.text_model.embeddings.token_embedding.weight.data[index_updates]
|
| 806 |
+
off_ratio = std_token_embedding / new_embeddings.std()
|
| 807 |
+
|
| 808 |
+
new_embeddings = new_embeddings * (off_ratio**0.1)
|
| 809 |
+
text_encoder.text_model.embeddings.token_embedding.weight.data[index_updates] = new_embeddings
|
| 810 |
+
|
| 811 |
+
|
| 812 |
+
class DreamBoothDataset(Dataset):
|
| 813 |
+
"""
|
| 814 |
+
A dataset to prepare the instance and class images with the prompts for fine-tuning the model.
|
| 815 |
+
It pre-processes the images.
|
| 816 |
+
"""
|
| 817 |
+
|
| 818 |
+
def __init__(
|
| 819 |
+
self,
|
| 820 |
+
instance_data_root,
|
| 821 |
+
instance_prompt,
|
| 822 |
+
class_prompt,
|
| 823 |
+
dataset_name,
|
| 824 |
+
dataset_config_name,
|
| 825 |
+
cache_dir,
|
| 826 |
+
image_column,
|
| 827 |
+
caption_column,
|
| 828 |
+
train_text_encoder_ti,
|
| 829 |
+
class_data_root=None,
|
| 830 |
+
class_num=None,
|
| 831 |
+
token_abstraction_dict=None, # token mapping for textual inversion
|
| 832 |
+
size=1024,
|
| 833 |
+
repeats=1,
|
| 834 |
+
center_crop=False,
|
| 835 |
+
):
|
| 836 |
+
self.size = size
|
| 837 |
+
self.center_crop = center_crop
|
| 838 |
+
|
| 839 |
+
self.instance_prompt = instance_prompt
|
| 840 |
+
self.custom_instance_prompts = None
|
| 841 |
+
self.class_prompt = class_prompt
|
| 842 |
+
self.token_abstraction_dict = token_abstraction_dict
|
| 843 |
+
self.train_text_encoder_ti = train_text_encoder_ti
|
| 844 |
+
# if --dataset_name is provided or a metadata jsonl file is provided in the local --instance_data directory,
|
| 845 |
+
# we load the training data using load_dataset
|
| 846 |
+
if dataset_name is not None:
|
| 847 |
+
try:
|
| 848 |
+
from datasets import load_dataset
|
| 849 |
+
except ImportError:
|
| 850 |
+
raise ImportError(
|
| 851 |
+
"You are trying to load your data using the datasets library. If you wish to train using custom "
|
| 852 |
+
"captions please install the datasets library: `pip install datasets`. If you wish to load a "
|
| 853 |
+
"local folder containing images only, specify --instance_data_dir instead."
|
| 854 |
+
)
|
| 855 |
+
# Downloading and loading a dataset from the hub.
|
| 856 |
+
# See more about loading custom images at
|
| 857 |
+
# https://huggingface.co/docs/datasets/v2.0.0/en/dataset_script
|
| 858 |
+
dataset = load_dataset(
|
| 859 |
+
dataset_name,
|
| 860 |
+
dataset_config_name,
|
| 861 |
+
cache_dir=cache_dir,
|
| 862 |
+
)
|
| 863 |
+
# Preprocessing the datasets.
|
| 864 |
+
column_names = dataset["train"].column_names
|
| 865 |
+
|
| 866 |
+
# 6. Get the column names for input/target.
|
| 867 |
+
if image_column is None:
|
| 868 |
+
image_column = column_names[0]
|
| 869 |
+
logger.info(f"image column defaulting to {image_column}")
|
| 870 |
+
else:
|
| 871 |
+
if image_column not in column_names:
|
| 872 |
+
raise ValueError(
|
| 873 |
+
f"`--image_column` value '{image_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
|
| 874 |
+
)
|
| 875 |
+
instance_images = dataset["train"][image_column]
|
| 876 |
+
|
| 877 |
+
if caption_column is None:
|
| 878 |
+
logger.info(
|
| 879 |
+
"No caption column provided, defaulting to instance_prompt for all images. If your dataset "
|
| 880 |
+
"contains captions/prompts for the images, make sure to specify the "
|
| 881 |
+
"column as --caption_column"
|
| 882 |
+
)
|
| 883 |
+
self.custom_instance_prompts = None
|
| 884 |
+
else:
|
| 885 |
+
if caption_column not in column_names:
|
| 886 |
+
raise ValueError(
|
| 887 |
+
f"`--caption_column` value '{caption_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
|
| 888 |
+
)
|
| 889 |
+
custom_instance_prompts = dataset["train"][caption_column]
|
| 890 |
+
# create final list of captions according to --repeats
|
| 891 |
+
self.custom_instance_prompts = []
|
| 892 |
+
for caption in custom_instance_prompts:
|
| 893 |
+
self.custom_instance_prompts.extend(itertools.repeat(caption, repeats))
|
| 894 |
+
else:
|
| 895 |
+
self.instance_data_root = Path(instance_data_root)
|
| 896 |
+
if not self.instance_data_root.exists():
|
| 897 |
+
raise ValueError("Instance images root doesn't exists.")
|
| 898 |
+
|
| 899 |
+
instance_images = [Image.open(path) for path in list(Path(instance_data_root).iterdir())]
|
| 900 |
+
self.custom_instance_prompts = None
|
| 901 |
+
|
| 902 |
+
self.instance_images = []
|
| 903 |
+
for img in instance_images:
|
| 904 |
+
self.instance_images.extend(itertools.repeat(img, repeats))
|
| 905 |
+
self.num_instance_images = len(self.instance_images)
|
| 906 |
+
self._length = self.num_instance_images
|
| 907 |
+
|
| 908 |
+
if class_data_root is not None:
|
| 909 |
+
self.class_data_root = Path(class_data_root)
|
| 910 |
+
self.class_data_root.mkdir(parents=True, exist_ok=True)
|
| 911 |
+
self.class_images_path = list(self.class_data_root.iterdir())
|
| 912 |
+
if class_num is not None:
|
| 913 |
+
self.num_class_images = min(len(self.class_images_path), class_num)
|
| 914 |
+
else:
|
| 915 |
+
self.num_class_images = len(self.class_images_path)
|
| 916 |
+
self._length = max(self.num_class_images, self.num_instance_images)
|
| 917 |
+
else:
|
| 918 |
+
self.class_data_root = None
|
| 919 |
+
|
| 920 |
+
self.image_transforms = transforms.Compose(
|
| 921 |
+
[
|
| 922 |
+
transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR),
|
| 923 |
+
transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size),
|
| 924 |
+
transforms.ToTensor(),
|
| 925 |
+
transforms.Normalize([0.5], [0.5]),
|
| 926 |
+
]
|
| 927 |
+
)
|
| 928 |
+
|
| 929 |
+
def __len__(self):
|
| 930 |
+
return self._length
|
| 931 |
+
|
| 932 |
+
def __getitem__(self, index):
|
| 933 |
+
example = {}
|
| 934 |
+
instance_image = self.instance_images[index % self.num_instance_images]
|
| 935 |
+
instance_image = exif_transpose(instance_image)
|
| 936 |
+
|
| 937 |
+
if not instance_image.mode == "RGB":
|
| 938 |
+
instance_image = instance_image.convert("RGB")
|
| 939 |
+
example["instance_images"] = self.image_transforms(instance_image)
|
| 940 |
+
|
| 941 |
+
if self.custom_instance_prompts:
|
| 942 |
+
caption = self.custom_instance_prompts[index % self.num_instance_images]
|
| 943 |
+
if caption:
|
| 944 |
+
if self.train_text_encoder_ti:
|
| 945 |
+
# replace instances of --token_abstraction in caption with the new tokens: "<si><si+1>" etc.
|
| 946 |
+
for token_abs, token_replacement in self.token_abstraction_dict.items():
|
| 947 |
+
caption = caption.replace(token_abs, "".join(token_replacement))
|
| 948 |
+
example["instance_prompt"] = caption
|
| 949 |
+
else:
|
| 950 |
+
example["instance_prompt"] = self.instance_prompt
|
| 951 |
+
|
| 952 |
+
else: # costum prompts were provided, but length does not match size of image dataset
|
| 953 |
+
example["instance_prompt"] = self.instance_prompt
|
| 954 |
+
|
| 955 |
+
if self.class_data_root:
|
| 956 |
+
class_image = Image.open(self.class_images_path[index % self.num_class_images])
|
| 957 |
+
class_image = exif_transpose(class_image)
|
| 958 |
+
|
| 959 |
+
if not class_image.mode == "RGB":
|
| 960 |
+
class_image = class_image.convert("RGB")
|
| 961 |
+
example["class_images"] = self.image_transforms(class_image)
|
| 962 |
+
example["class_prompt"] = self.class_prompt
|
| 963 |
+
|
| 964 |
+
return example
|
| 965 |
+
|
| 966 |
+
|
| 967 |
+
def collate_fn(examples, with_prior_preservation=False):
|
| 968 |
+
pixel_values = [example["instance_images"] for example in examples]
|
| 969 |
+
prompts = [example["instance_prompt"] for example in examples]
|
| 970 |
+
|
| 971 |
+
# Concat class and instance examples for prior preservation.
|
| 972 |
+
# We do this to avoid doing two forward passes.
|
| 973 |
+
if with_prior_preservation:
|
| 974 |
+
pixel_values += [example["class_images"] for example in examples]
|
| 975 |
+
prompts += [example["class_prompt"] for example in examples]
|
| 976 |
+
|
| 977 |
+
pixel_values = torch.stack(pixel_values)
|
| 978 |
+
pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
|
| 979 |
+
|
| 980 |
+
batch = {"pixel_values": pixel_values, "prompts": prompts}
|
| 981 |
+
return batch
|
| 982 |
+
|
| 983 |
+
|
| 984 |
+
class PromptDataset(Dataset):
|
| 985 |
+
"A simple dataset to prepare the prompts to generate class images on multiple GPUs."
|
| 986 |
+
|
| 987 |
+
def __init__(self, prompt, num_samples):
|
| 988 |
+
self.prompt = prompt
|
| 989 |
+
self.num_samples = num_samples
|
| 990 |
+
|
| 991 |
+
def __len__(self):
|
| 992 |
+
return self.num_samples
|
| 993 |
+
|
| 994 |
+
def __getitem__(self, index):
|
| 995 |
+
example = {}
|
| 996 |
+
example["prompt"] = self.prompt
|
| 997 |
+
example["index"] = index
|
| 998 |
+
return example
|
| 999 |
+
|
| 1000 |
+
|
| 1001 |
+
def tokenize_prompt(tokenizer, prompt, add_special_tokens=False):
|
| 1002 |
+
text_inputs = tokenizer(
|
| 1003 |
+
prompt,
|
| 1004 |
+
padding="max_length",
|
| 1005 |
+
max_length=tokenizer.model_max_length,
|
| 1006 |
+
truncation=True,
|
| 1007 |
+
add_special_tokens=add_special_tokens,
|
| 1008 |
+
return_tensors="pt",
|
| 1009 |
+
)
|
| 1010 |
+
text_input_ids = text_inputs.input_ids
|
| 1011 |
+
return text_input_ids
|
| 1012 |
+
|
| 1013 |
+
|
| 1014 |
+
# Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt
|
| 1015 |
+
def encode_prompt(text_encoders, tokenizers, prompt, text_input_ids_list=None):
|
| 1016 |
+
prompt_embeds_list = []
|
| 1017 |
+
|
| 1018 |
+
for i, text_encoder in enumerate(text_encoders):
|
| 1019 |
+
if tokenizers is not None:
|
| 1020 |
+
tokenizer = tokenizers[i]
|
| 1021 |
+
text_input_ids = tokenize_prompt(tokenizer, prompt)
|
| 1022 |
+
else:
|
| 1023 |
+
assert text_input_ids_list is not None
|
| 1024 |
+
text_input_ids = text_input_ids_list[i]
|
| 1025 |
+
|
| 1026 |
+
prompt_embeds = text_encoder(
|
| 1027 |
+
text_input_ids.to(text_encoder.device),
|
| 1028 |
+
output_hidden_states=True,
|
| 1029 |
+
)
|
| 1030 |
+
|
| 1031 |
+
# We are only ALWAYS interested in the pooled output of the final text encoder
|
| 1032 |
+
pooled_prompt_embeds = prompt_embeds[0]
|
| 1033 |
+
prompt_embeds = prompt_embeds.hidden_states[-2]
|
| 1034 |
+
bs_embed, seq_len, _ = prompt_embeds.shape
|
| 1035 |
+
prompt_embeds = prompt_embeds.view(bs_embed, seq_len, -1)
|
| 1036 |
+
prompt_embeds_list.append(prompt_embeds)
|
| 1037 |
+
|
| 1038 |
+
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
|
| 1039 |
+
pooled_prompt_embeds = pooled_prompt_embeds.view(bs_embed, -1)
|
| 1040 |
+
return prompt_embeds, pooled_prompt_embeds
|
| 1041 |
+
|
| 1042 |
+
|
| 1043 |
+
def main(args):
|
| 1044 |
+
logging_dir = Path(args.output_dir, args.logging_dir)
|
| 1045 |
+
|
| 1046 |
+
accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
|
| 1047 |
+
kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
|
| 1048 |
+
accelerator = Accelerator(
|
| 1049 |
+
gradient_accumulation_steps=args.gradient_accumulation_steps,
|
| 1050 |
+
mixed_precision=args.mixed_precision,
|
| 1051 |
+
log_with=args.report_to,
|
| 1052 |
+
project_config=accelerator_project_config,
|
| 1053 |
+
kwargs_handlers=[kwargs],
|
| 1054 |
+
)
|
| 1055 |
+
|
| 1056 |
+
if args.report_to == "wandb":
|
| 1057 |
+
if not is_wandb_available():
|
| 1058 |
+
raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
|
| 1059 |
+
import wandb
|
| 1060 |
+
|
| 1061 |
+
# Make one log on every process with the configuration for debugging.
|
| 1062 |
+
logging.basicConfig(
|
| 1063 |
+
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
|
| 1064 |
+
datefmt="%m/%d/%Y %H:%M:%S",
|
| 1065 |
+
level=logging.INFO,
|
| 1066 |
+
)
|
| 1067 |
+
logger.info(accelerator.state, main_process_only=False)
|
| 1068 |
+
if accelerator.is_local_main_process:
|
| 1069 |
+
transformers.utils.logging.set_verbosity_warning()
|
| 1070 |
+
diffusers.utils.logging.set_verbosity_info()
|
| 1071 |
+
else:
|
| 1072 |
+
transformers.utils.logging.set_verbosity_error()
|
| 1073 |
+
diffusers.utils.logging.set_verbosity_error()
|
| 1074 |
+
|
| 1075 |
+
# If passed along, set the training seed now.
|
| 1076 |
+
if args.seed is not None:
|
| 1077 |
+
set_seed(args.seed)
|
| 1078 |
+
|
| 1079 |
+
# Generate class images if prior preservation is enabled.
|
| 1080 |
+
if args.with_prior_preservation:
|
| 1081 |
+
class_images_dir = Path(args.class_data_dir)
|
| 1082 |
+
if not class_images_dir.exists():
|
| 1083 |
+
class_images_dir.mkdir(parents=True)
|
| 1084 |
+
cur_class_images = len(list(class_images_dir.iterdir()))
|
| 1085 |
+
|
| 1086 |
+
if cur_class_images < args.num_class_images:
|
| 1087 |
+
torch_dtype = torch.float16 if accelerator.device.type == "cuda" else torch.float32
|
| 1088 |
+
if args.prior_generation_precision == "fp32":
|
| 1089 |
+
torch_dtype = torch.float32
|
| 1090 |
+
elif args.prior_generation_precision == "fp16":
|
| 1091 |
+
torch_dtype = torch.float16
|
| 1092 |
+
elif args.prior_generation_precision == "bf16":
|
| 1093 |
+
torch_dtype = torch.bfloat16
|
| 1094 |
+
pipeline = StableDiffusionXLPipeline.from_pretrained(
|
| 1095 |
+
args.pretrained_model_name_or_path,
|
| 1096 |
+
torch_dtype=torch_dtype,
|
| 1097 |
+
revision=args.revision,
|
| 1098 |
+
variant=args.variant,
|
| 1099 |
+
)
|
| 1100 |
+
pipeline.set_progress_bar_config(disable=True)
|
| 1101 |
+
|
| 1102 |
+
num_new_images = args.num_class_images - cur_class_images
|
| 1103 |
+
logger.info(f"Number of class images to sample: {num_new_images}.")
|
| 1104 |
+
|
| 1105 |
+
sample_dataset = PromptDataset(args.class_prompt, num_new_images)
|
| 1106 |
+
sample_dataloader = torch.utils.data.DataLoader(sample_dataset, batch_size=args.sample_batch_size)
|
| 1107 |
+
|
| 1108 |
+
sample_dataloader = accelerator.prepare(sample_dataloader)
|
| 1109 |
+
pipeline.to(accelerator.device)
|
| 1110 |
+
|
| 1111 |
+
for example in tqdm(
|
| 1112 |
+
sample_dataloader, desc="Generating class images", disable=not accelerator.is_local_main_process
|
| 1113 |
+
):
|
| 1114 |
+
images = pipeline(example["prompt"]).images
|
| 1115 |
+
|
| 1116 |
+
for i, image in enumerate(images):
|
| 1117 |
+
hash_image = hashlib.sha1(image.tobytes()).hexdigest()
|
| 1118 |
+
image_filename = class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg"
|
| 1119 |
+
image.save(image_filename)
|
| 1120 |
+
|
| 1121 |
+
del pipeline
|
| 1122 |
+
if torch.cuda.is_available():
|
| 1123 |
+
torch.cuda.empty_cache()
|
| 1124 |
+
|
| 1125 |
+
# Handle the repository creation
|
| 1126 |
+
if accelerator.is_main_process:
|
| 1127 |
+
if args.output_dir is not None:
|
| 1128 |
+
os.makedirs(args.output_dir, exist_ok=True)
|
| 1129 |
+
|
| 1130 |
+
model_id = args.hub_model_id or Path(args.output_dir).name
|
| 1131 |
+
repo_id = None
|
| 1132 |
+
if args.push_to_hub:
|
| 1133 |
+
repo_id = create_repo(repo_id=model_id, exist_ok=True, token=args.hub_token).repo_id
|
| 1134 |
+
|
| 1135 |
+
# Load the tokenizers
|
| 1136 |
+
tokenizer_one = AutoTokenizer.from_pretrained(
|
| 1137 |
+
args.pretrained_model_name_or_path,
|
| 1138 |
+
subfolder="tokenizer",
|
| 1139 |
+
revision=args.revision,
|
| 1140 |
+
variant=args.variant,
|
| 1141 |
+
use_fast=False,
|
| 1142 |
+
)
|
| 1143 |
+
tokenizer_two = AutoTokenizer.from_pretrained(
|
| 1144 |
+
args.pretrained_model_name_or_path,
|
| 1145 |
+
subfolder="tokenizer_2",
|
| 1146 |
+
revision=args.revision,
|
| 1147 |
+
variant=args.variant,
|
| 1148 |
+
use_fast=False,
|
| 1149 |
+
)
|
| 1150 |
+
|
| 1151 |
+
# import correct text encoder classes
|
| 1152 |
+
text_encoder_cls_one = import_model_class_from_model_name_or_path(
|
| 1153 |
+
args.pretrained_model_name_or_path, args.revision
|
| 1154 |
+
)
|
| 1155 |
+
text_encoder_cls_two = import_model_class_from_model_name_or_path(
|
| 1156 |
+
args.pretrained_model_name_or_path, args.revision, subfolder="text_encoder_2"
|
| 1157 |
+
)
|
| 1158 |
+
|
| 1159 |
+
# Load scheduler and models
|
| 1160 |
+
noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
|
| 1161 |
+
text_encoder_one = text_encoder_cls_one.from_pretrained(
|
| 1162 |
+
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
|
| 1163 |
+
)
|
| 1164 |
+
text_encoder_two = text_encoder_cls_two.from_pretrained(
|
| 1165 |
+
args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant
|
| 1166 |
+
)
|
| 1167 |
+
vae_path = (
|
| 1168 |
+
args.pretrained_model_name_or_path
|
| 1169 |
+
if args.pretrained_vae_model_name_or_path is None
|
| 1170 |
+
else args.pretrained_vae_model_name_or_path
|
| 1171 |
+
)
|
| 1172 |
+
vae = AutoencoderKL.from_pretrained(
|
| 1173 |
+
vae_path,
|
| 1174 |
+
subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
|
| 1175 |
+
revision=args.revision,
|
| 1176 |
+
variant=args.variant,
|
| 1177 |
+
)
|
| 1178 |
+
vae_scaling_factor = vae.config.scaling_factor
|
| 1179 |
+
unet = UNet2DConditionModel.from_pretrained(
|
| 1180 |
+
args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
|
| 1181 |
+
)
|
| 1182 |
+
|
| 1183 |
+
if args.train_text_encoder_ti:
|
| 1184 |
+
# we parse the provided token identifier (or identifiers) into a list. s.t. - "TOK" -> ["TOK"], "TOK,
|
| 1185 |
+
# TOK2" -> ["TOK", "TOK2"] etc.
|
| 1186 |
+
token_abstraction_list = "".join(args.token_abstraction.split()).split(",")
|
| 1187 |
+
logger.info(f"list of token identifiers: {token_abstraction_list}")
|
| 1188 |
+
|
| 1189 |
+
token_abstraction_dict = {}
|
| 1190 |
+
token_idx = 0
|
| 1191 |
+
for i, token in enumerate(token_abstraction_list):
|
| 1192 |
+
token_abstraction_dict[token] = [
|
| 1193 |
+
f"<s{token_idx + i + j}>" for j in range(args.num_new_tokens_per_abstraction)
|
| 1194 |
+
]
|
| 1195 |
+
token_idx += args.num_new_tokens_per_abstraction - 1
|
| 1196 |
+
|
| 1197 |
+
# replace instances of --token_abstraction in --instance_prompt with the new tokens: "<si><si+1>" etc.
|
| 1198 |
+
for token_abs, token_replacement in token_abstraction_dict.items():
|
| 1199 |
+
args.instance_prompt = args.instance_prompt.replace(token_abs, "".join(token_replacement))
|
| 1200 |
+
if args.with_prior_preservation:
|
| 1201 |
+
args.class_prompt = args.class_prompt.replace(token_abs, "".join(token_replacement))
|
| 1202 |
+
|
| 1203 |
+
# initialize the new tokens for textual inversion
|
| 1204 |
+
embedding_handler = TokenEmbeddingsHandler(
|
| 1205 |
+
[text_encoder_one, text_encoder_two], [tokenizer_one, tokenizer_two]
|
| 1206 |
+
)
|
| 1207 |
+
inserting_toks = []
|
| 1208 |
+
for new_tok in token_abstraction_dict.values():
|
| 1209 |
+
inserting_toks.extend(new_tok)
|
| 1210 |
+
embedding_handler.initialize_new_tokens(inserting_toks=inserting_toks)
|
| 1211 |
+
|
| 1212 |
+
# We only train the additional adapter LoRA layers
|
| 1213 |
+
vae.requires_grad_(False)
|
| 1214 |
+
text_encoder_one.requires_grad_(False)
|
| 1215 |
+
text_encoder_two.requires_grad_(False)
|
| 1216 |
+
unet.requires_grad_(False)
|
| 1217 |
+
|
| 1218 |
+
# For mixed precision training we cast all non-trainable weights (vae, non-lora text_encoder and non-lora unet) to half-precision
|
| 1219 |
+
# as these weights are only used for inference, keeping weights in full precision is not required.
|
| 1220 |
+
weight_dtype = torch.float32
|
| 1221 |
+
if accelerator.mixed_precision == "fp16":
|
| 1222 |
+
weight_dtype = torch.float16
|
| 1223 |
+
elif accelerator.mixed_precision == "bf16":
|
| 1224 |
+
weight_dtype = torch.bfloat16
|
| 1225 |
+
|
| 1226 |
+
# Move unet, vae and text_encoder to device and cast to weight_dtype
|
| 1227 |
+
unet.to(accelerator.device, dtype=weight_dtype)
|
| 1228 |
+
|
| 1229 |
+
# The VAE is always in float32 to avoid NaN losses.
|
| 1230 |
+
vae.to(accelerator.device, dtype=torch.float32)
|
| 1231 |
+
|
| 1232 |
+
text_encoder_one.to(accelerator.device, dtype=weight_dtype)
|
| 1233 |
+
text_encoder_two.to(accelerator.device, dtype=weight_dtype)
|
| 1234 |
+
|
| 1235 |
+
if args.enable_xformers_memory_efficient_attention:
|
| 1236 |
+
if is_xformers_available():
|
| 1237 |
+
import xformers
|
| 1238 |
+
|
| 1239 |
+
xformers_version = version.parse(xformers.__version__)
|
| 1240 |
+
if xformers_version == version.parse("0.0.16"):
|
| 1241 |
+
logger.warn(
|
| 1242 |
+
"xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, "
|
| 1243 |
+
"please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
|
| 1244 |
+
)
|
| 1245 |
+
unet.enable_xformers_memory_efficient_attention()
|
| 1246 |
+
else:
|
| 1247 |
+
raise ValueError("xformers is not available. Make sure it is installed correctly")
|
| 1248 |
+
|
| 1249 |
+
if args.gradient_checkpointing:
|
| 1250 |
+
unet.enable_gradient_checkpointing()
|
| 1251 |
+
if args.train_text_encoder:
|
| 1252 |
+
text_encoder_one.gradient_checkpointing_enable()
|
| 1253 |
+
text_encoder_two.gradient_checkpointing_enable()
|
| 1254 |
+
|
| 1255 |
+
# now we will add new LoRA weights to the attention layers
|
| 1256 |
+
unet_lora_config = LoraConfig(
|
| 1257 |
+
r=args.rank,
|
| 1258 |
+
lora_alpha=args.rank,
|
| 1259 |
+
init_lora_weights="gaussian",
|
| 1260 |
+
target_modules=["to_k", "to_q", "to_v", "to_out.0"],
|
| 1261 |
+
)
|
| 1262 |
+
unet.add_adapter(unet_lora_config)
|
| 1263 |
+
|
| 1264 |
+
# The text encoder comes from π€ transformers, so we cannot directly modify it.
|
| 1265 |
+
# So, instead, we monkey-patch the forward calls of its attention-blocks.
|
| 1266 |
+
if args.train_text_encoder:
|
| 1267 |
+
text_lora_config = LoraConfig(
|
| 1268 |
+
r=args.rank,
|
| 1269 |
+
lora_alpha=args.rank,
|
| 1270 |
+
init_lora_weights="gaussian",
|
| 1271 |
+
target_modules=["q_proj", "k_proj", "v_proj", "out_proj"],
|
| 1272 |
+
)
|
| 1273 |
+
text_encoder_one.add_adapter(text_lora_config)
|
| 1274 |
+
text_encoder_two.add_adapter(text_lora_config)
|
| 1275 |
+
|
| 1276 |
+
# if we use textual inversion, we freeze all parameters except for the token embeddings
|
| 1277 |
+
# in text encoder
|
| 1278 |
+
elif args.train_text_encoder_ti:
|
| 1279 |
+
text_lora_parameters_one = []
|
| 1280 |
+
for name, param in text_encoder_one.named_parameters():
|
| 1281 |
+
if "token_embedding" in name:
|
| 1282 |
+
# ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
|
| 1283 |
+
param.data = param.to(dtype=torch.float32)
|
| 1284 |
+
param.requires_grad = True
|
| 1285 |
+
text_lora_parameters_one.append(param)
|
| 1286 |
+
else:
|
| 1287 |
+
param.requires_grad = False
|
| 1288 |
+
text_lora_parameters_two = []
|
| 1289 |
+
for name, param in text_encoder_two.named_parameters():
|
| 1290 |
+
if "token_embedding" in name:
|
| 1291 |
+
# ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
|
| 1292 |
+
param.data = param.to(dtype=torch.float32)
|
| 1293 |
+
param.requires_grad = True
|
| 1294 |
+
text_lora_parameters_two.append(param)
|
| 1295 |
+
else:
|
| 1296 |
+
param.requires_grad = False
|
| 1297 |
+
|
| 1298 |
+
def unwrap_model(model):
|
| 1299 |
+
model = accelerator.unwrap_model(model)
|
| 1300 |
+
model = model._orig_mod if is_compiled_module(model) else model
|
| 1301 |
+
return model
|
| 1302 |
+
|
| 1303 |
+
# create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
|
| 1304 |
+
def save_model_hook(models, weights, output_dir):
|
| 1305 |
+
if accelerator.is_main_process:
|
| 1306 |
+
# there are only two options here. Either are just the unet attn processor layers
|
| 1307 |
+
# or there are the unet and text encoder atten layers
|
| 1308 |
+
unet_lora_layers_to_save = None
|
| 1309 |
+
text_encoder_one_lora_layers_to_save = None
|
| 1310 |
+
text_encoder_two_lora_layers_to_save = None
|
| 1311 |
+
|
| 1312 |
+
for model in models:
|
| 1313 |
+
if isinstance(model, type(unwrap_model(unet))):
|
| 1314 |
+
unet_lora_layers_to_save = convert_state_dict_to_diffusers(get_peft_model_state_dict(model))
|
| 1315 |
+
elif isinstance(model, type(unwrap_model(text_encoder_one))):
|
| 1316 |
+
if args.train_text_encoder:
|
| 1317 |
+
text_encoder_one_lora_layers_to_save = convert_state_dict_to_diffusers(
|
| 1318 |
+
get_peft_model_state_dict(model)
|
| 1319 |
+
)
|
| 1320 |
+
elif isinstance(model, type(unwrap_model(text_encoder_two))):
|
| 1321 |
+
if args.train_text_encoder:
|
| 1322 |
+
text_encoder_two_lora_layers_to_save = convert_state_dict_to_diffusers(
|
| 1323 |
+
get_peft_model_state_dict(model)
|
| 1324 |
+
)
|
| 1325 |
+
else:
|
| 1326 |
+
raise ValueError(f"unexpected save model: {model.__class__}")
|
| 1327 |
+
|
| 1328 |
+
# make sure to pop weight so that corresponding model is not saved again
|
| 1329 |
+
weights.pop()
|
| 1330 |
+
|
| 1331 |
+
StableDiffusionXLPipeline.save_lora_weights(
|
| 1332 |
+
output_dir,
|
| 1333 |
+
unet_lora_layers=unet_lora_layers_to_save,
|
| 1334 |
+
text_encoder_lora_layers=text_encoder_one_lora_layers_to_save,
|
| 1335 |
+
text_encoder_2_lora_layers=text_encoder_two_lora_layers_to_save,
|
| 1336 |
+
)
|
| 1337 |
+
if args.train_text_encoder_ti:
|
| 1338 |
+
embedding_handler.save_embeddings(f"{output_dir}/{args.output_dir}_emb.safetensors")
|
| 1339 |
+
|
| 1340 |
+
def load_model_hook(models, input_dir):
|
| 1341 |
+
unet_ = None
|
| 1342 |
+
text_encoder_one_ = None
|
| 1343 |
+
text_encoder_two_ = None
|
| 1344 |
+
|
| 1345 |
+
while len(models) > 0:
|
| 1346 |
+
model = models.pop()
|
| 1347 |
+
|
| 1348 |
+
if isinstance(model, type(unwrap_model(unet))):
|
| 1349 |
+
unet_ = model
|
| 1350 |
+
elif isinstance(model, type(unwrap_model(text_encoder_one))):
|
| 1351 |
+
text_encoder_one_ = model
|
| 1352 |
+
elif isinstance(model, type(unwrap_model(text_encoder_two))):
|
| 1353 |
+
text_encoder_two_ = model
|
| 1354 |
+
else:
|
| 1355 |
+
raise ValueError(f"unexpected save model: {model.__class__}")
|
| 1356 |
+
|
| 1357 |
+
lora_state_dict, network_alphas = LoraLoaderMixin.lora_state_dict(input_dir)
|
| 1358 |
+
|
| 1359 |
+
unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")}
|
| 1360 |
+
unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
|
| 1361 |
+
incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
|
| 1362 |
+
if incompatible_keys is not None:
|
| 1363 |
+
# check only for unexpected keys
|
| 1364 |
+
unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
|
| 1365 |
+
if unexpected_keys:
|
| 1366 |
+
logger.warning(
|
| 1367 |
+
f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
|
| 1368 |
+
f" {unexpected_keys}. "
|
| 1369 |
+
)
|
| 1370 |
+
|
| 1371 |
+
if args.train_text_encoder:
|
| 1372 |
+
_set_state_dict_into_text_encoder(lora_state_dict, prefix="text_encoder.", text_encoder=text_encoder_one_)
|
| 1373 |
+
|
| 1374 |
+
_set_state_dict_into_text_encoder(
|
| 1375 |
+
lora_state_dict, prefix="text_encoder_2.", text_encoder=text_encoder_two_
|
| 1376 |
+
)
|
| 1377 |
+
|
| 1378 |
+
# Make sure the trainable params are in float32. This is again needed since the base models
|
| 1379 |
+
# are in `weight_dtype`. More details:
|
| 1380 |
+
# https://github.com/huggingface/diffusers/pull/6514#discussion_r1449796804
|
| 1381 |
+
if args.mixed_precision == "fp16":
|
| 1382 |
+
models = [unet_]
|
| 1383 |
+
if args.train_text_encoder:
|
| 1384 |
+
models.extend([text_encoder_one_, text_encoder_two_])
|
| 1385 |
+
cast_training_params(models)
|
| 1386 |
+
|
| 1387 |
+
accelerator.register_save_state_pre_hook(save_model_hook)
|
| 1388 |
+
accelerator.register_load_state_pre_hook(load_model_hook)
|
| 1389 |
+
|
| 1390 |
+
# Enable TF32 for faster training on Ampere GPUs,
|
| 1391 |
+
# cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
|
| 1392 |
+
if args.allow_tf32:
|
| 1393 |
+
torch.backends.cuda.matmul.allow_tf32 = True
|
| 1394 |
+
|
| 1395 |
+
if args.scale_lr:
|
| 1396 |
+
args.learning_rate = (
|
| 1397 |
+
args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
|
| 1398 |
+
)
|
| 1399 |
+
|
| 1400 |
+
# Make sure the trainable params are in float32.
|
| 1401 |
+
if args.mixed_precision == "fp16":
|
| 1402 |
+
models = [unet]
|
| 1403 |
+
if args.train_text_encoder:
|
| 1404 |
+
models.extend([text_encoder_one, text_encoder_two])
|
| 1405 |
+
cast_training_params(models, dtype=torch.float32)
|
| 1406 |
+
|
| 1407 |
+
unet_lora_parameters = list(filter(lambda p: p.requires_grad, unet.parameters()))
|
| 1408 |
+
|
| 1409 |
+
if args.train_text_encoder:
|
| 1410 |
+
text_lora_parameters_one = list(filter(lambda p: p.requires_grad, text_encoder_one.parameters()))
|
| 1411 |
+
text_lora_parameters_two = list(filter(lambda p: p.requires_grad, text_encoder_two.parameters()))
|
| 1412 |
+
|
| 1413 |
+
# If neither --train_text_encoder nor --train_text_encoder_ti, text_encoders remain frozen during training
|
| 1414 |
+
freeze_text_encoder = not (args.train_text_encoder or args.train_text_encoder_ti)
|
| 1415 |
+
|
| 1416 |
+
# Optimization parameters
|
| 1417 |
+
unet_lora_parameters_with_lr = {"params": unet_lora_parameters, "lr": args.learning_rate}
|
| 1418 |
+
if not freeze_text_encoder:
|
| 1419 |
+
# different learning rate for text encoder and unet
|
| 1420 |
+
text_lora_parameters_one_with_lr = {
|
| 1421 |
+
"params": text_lora_parameters_one,
|
| 1422 |
+
"weight_decay": args.adam_weight_decay_text_encoder
|
| 1423 |
+
if args.adam_weight_decay_text_encoder
|
| 1424 |
+
else args.adam_weight_decay,
|
| 1425 |
+
"lr": args.text_encoder_lr if args.text_encoder_lr else args.learning_rate,
|
| 1426 |
+
}
|
| 1427 |
+
text_lora_parameters_two_with_lr = {
|
| 1428 |
+
"params": text_lora_parameters_two,
|
| 1429 |
+
"weight_decay": args.adam_weight_decay_text_encoder
|
| 1430 |
+
if args.adam_weight_decay_text_encoder
|
| 1431 |
+
else args.adam_weight_decay,
|
| 1432 |
+
"lr": args.text_encoder_lr if args.text_encoder_lr else args.learning_rate,
|
| 1433 |
+
}
|
| 1434 |
+
params_to_optimize = [
|
| 1435 |
+
unet_lora_parameters_with_lr,
|
| 1436 |
+
text_lora_parameters_one_with_lr,
|
| 1437 |
+
text_lora_parameters_two_with_lr,
|
| 1438 |
+
]
|
| 1439 |
+
else:
|
| 1440 |
+
params_to_optimize = [unet_lora_parameters_with_lr]
|
| 1441 |
+
|
| 1442 |
+
# Optimizer creation
|
| 1443 |
+
if not (args.optimizer.lower() == "prodigy" or args.optimizer.lower() == "adamw"):
|
| 1444 |
+
logger.warn(
|
| 1445 |
+
f"Unsupported choice of optimizer: {args.optimizer}.Supported optimizers include [adamW, prodigy]."
|
| 1446 |
+
"Defaulting to adamW"
|
| 1447 |
+
)
|
| 1448 |
+
args.optimizer = "adamw"
|
| 1449 |
+
|
| 1450 |
+
if args.use_8bit_adam and not args.optimizer.lower() == "adamw":
|
| 1451 |
+
logger.warn(
|
| 1452 |
+
f"use_8bit_adam is ignored when optimizer is not set to 'AdamW'. Optimizer was "
|
| 1453 |
+
f"set to {args.optimizer.lower()}"
|
| 1454 |
+
)
|
| 1455 |
+
|
| 1456 |
+
if args.optimizer.lower() == "adamw":
|
| 1457 |
+
if args.use_8bit_adam:
|
| 1458 |
+
try:
|
| 1459 |
+
import bitsandbytes as bnb
|
| 1460 |
+
except ImportError:
|
| 1461 |
+
raise ImportError(
|
| 1462 |
+
"To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
|
| 1463 |
+
)
|
| 1464 |
+
|
| 1465 |
+
optimizer_class = bnb.optim.AdamW8bit
|
| 1466 |
+
else:
|
| 1467 |
+
optimizer_class = torch.optim.AdamW
|
| 1468 |
+
|
| 1469 |
+
optimizer = optimizer_class(
|
| 1470 |
+
params_to_optimize,
|
| 1471 |
+
betas=(args.adam_beta1, args.adam_beta2),
|
| 1472 |
+
weight_decay=args.adam_weight_decay,
|
| 1473 |
+
eps=args.adam_epsilon,
|
| 1474 |
+
)
|
| 1475 |
+
|
| 1476 |
+
if args.optimizer.lower() == "prodigy":
|
| 1477 |
+
try:
|
| 1478 |
+
import prodigyopt
|
| 1479 |
+
except ImportError:
|
| 1480 |
+
raise ImportError("To use Prodigy, please install the prodigyopt library: `pip install prodigyopt`")
|
| 1481 |
+
|
| 1482 |
+
optimizer_class = prodigyopt.Prodigy
|
| 1483 |
+
|
| 1484 |
+
if args.learning_rate <= 0.1:
|
| 1485 |
+
logger.warn(
|
| 1486 |
+
"Learning rate is too low. When using prodigy, it's generally better to set learning rate around 1.0"
|
| 1487 |
+
)
|
| 1488 |
+
if args.train_text_encoder and args.text_encoder_lr:
|
| 1489 |
+
logger.warn(
|
| 1490 |
+
f"Learning rates were provided both for the unet and the text encoder- e.g. text_encoder_lr:"
|
| 1491 |
+
f" {args.text_encoder_lr} and learning_rate: {args.learning_rate}. "
|
| 1492 |
+
f"When using prodigy only learning_rate is used as the initial learning rate."
|
| 1493 |
+
)
|
| 1494 |
+
# changes the learning rate of text_encoder_parameters_one and text_encoder_parameters_two to be
|
| 1495 |
+
# --learning_rate
|
| 1496 |
+
params_to_optimize[1]["lr"] = args.learning_rate
|
| 1497 |
+
params_to_optimize[2]["lr"] = args.learning_rate
|
| 1498 |
+
|
| 1499 |
+
optimizer = optimizer_class(
|
| 1500 |
+
params_to_optimize,
|
| 1501 |
+
lr=args.learning_rate,
|
| 1502 |
+
betas=(args.adam_beta1, args.adam_beta2),
|
| 1503 |
+
beta3=args.prodigy_beta3,
|
| 1504 |
+
weight_decay=args.adam_weight_decay,
|
| 1505 |
+
eps=args.adam_epsilon,
|
| 1506 |
+
decouple=args.prodigy_decouple,
|
| 1507 |
+
use_bias_correction=args.prodigy_use_bias_correction,
|
| 1508 |
+
safeguard_warmup=args.prodigy_safeguard_warmup,
|
| 1509 |
+
)
|
| 1510 |
+
|
| 1511 |
+
# Dataset and DataLoaders creation:
|
| 1512 |
+
train_dataset = DreamBoothDataset(
|
| 1513 |
+
instance_data_root=args.instance_data_dir,
|
| 1514 |
+
instance_prompt=args.instance_prompt,
|
| 1515 |
+
class_prompt=args.class_prompt,
|
| 1516 |
+
dataset_name=args.dataset_name,
|
| 1517 |
+
dataset_config_name=args.dataset_config_name,
|
| 1518 |
+
cache_dir=args.cache_dir,
|
| 1519 |
+
image_column=args.image_column,
|
| 1520 |
+
train_text_encoder_ti=args.train_text_encoder_ti,
|
| 1521 |
+
caption_column=args.caption_column,
|
| 1522 |
+
class_data_root=args.class_data_dir if args.with_prior_preservation else None,
|
| 1523 |
+
token_abstraction_dict=token_abstraction_dict if args.train_text_encoder_ti else None,
|
| 1524 |
+
class_num=args.num_class_images,
|
| 1525 |
+
size=args.resolution,
|
| 1526 |
+
repeats=args.repeats,
|
| 1527 |
+
center_crop=args.center_crop,
|
| 1528 |
+
)
|
| 1529 |
+
|
| 1530 |
+
train_dataloader = torch.utils.data.DataLoader(
|
| 1531 |
+
train_dataset,
|
| 1532 |
+
batch_size=args.train_batch_size,
|
| 1533 |
+
shuffle=True,
|
| 1534 |
+
collate_fn=lambda examples: collate_fn(examples, args.with_prior_preservation),
|
| 1535 |
+
num_workers=args.dataloader_num_workers,
|
| 1536 |
+
)
|
| 1537 |
+
|
| 1538 |
+
# Computes additional embeddings/ids required by the SDXL UNet.
|
| 1539 |
+
# regular text embeddings (when `train_text_encoder` is not True)
|
| 1540 |
+
# pooled text embeddings
|
| 1541 |
+
# time ids
|
| 1542 |
+
|
| 1543 |
+
def compute_time_ids():
|
| 1544 |
+
# Adapted from pipeline.StableDiffusionXLPipeline._get_add_time_ids
|
| 1545 |
+
original_size = (args.resolution, args.resolution)
|
| 1546 |
+
target_size = (args.resolution, args.resolution)
|
| 1547 |
+
crops_coords_top_left = (args.crops_coords_top_left_h, args.crops_coords_top_left_w)
|
| 1548 |
+
add_time_ids = list(original_size + crops_coords_top_left + target_size)
|
| 1549 |
+
add_time_ids = torch.tensor([add_time_ids])
|
| 1550 |
+
add_time_ids = add_time_ids.to(accelerator.device, dtype=weight_dtype)
|
| 1551 |
+
return add_time_ids
|
| 1552 |
+
|
| 1553 |
+
if not args.train_text_encoder:
|
| 1554 |
+
tokenizers = [tokenizer_one, tokenizer_two]
|
| 1555 |
+
text_encoders = [text_encoder_one, text_encoder_two]
|
| 1556 |
+
|
| 1557 |
+
def compute_text_embeddings(prompt, text_encoders, tokenizers):
|
| 1558 |
+
with torch.no_grad():
|
| 1559 |
+
prompt_embeds, pooled_prompt_embeds = encode_prompt(text_encoders, tokenizers, prompt)
|
| 1560 |
+
prompt_embeds = prompt_embeds.to(accelerator.device)
|
| 1561 |
+
pooled_prompt_embeds = pooled_prompt_embeds.to(accelerator.device)
|
| 1562 |
+
return prompt_embeds, pooled_prompt_embeds
|
| 1563 |
+
|
| 1564 |
+
# Handle instance prompt.
|
| 1565 |
+
instance_time_ids = compute_time_ids()
|
| 1566 |
+
|
| 1567 |
+
# If no type of tuning is done on the text_encoder and custom instance prompts are NOT
|
| 1568 |
+
# provided (i.e. the --instance_prompt is used for all images), we encode the instance prompt once to avoid
|
| 1569 |
+
# the redundant encoding.
|
| 1570 |
+
if freeze_text_encoder and not train_dataset.custom_instance_prompts:
|
| 1571 |
+
instance_prompt_hidden_states, instance_pooled_prompt_embeds = compute_text_embeddings(
|
| 1572 |
+
args.instance_prompt, text_encoders, tokenizers
|
| 1573 |
+
)
|
| 1574 |
+
|
| 1575 |
+
# Handle class prompt for prior-preservation.
|
| 1576 |
+
if args.with_prior_preservation:
|
| 1577 |
+
class_time_ids = compute_time_ids()
|
| 1578 |
+
if freeze_text_encoder:
|
| 1579 |
+
class_prompt_hidden_states, class_pooled_prompt_embeds = compute_text_embeddings(
|
| 1580 |
+
args.class_prompt, text_encoders, tokenizers
|
| 1581 |
+
)
|
| 1582 |
+
|
| 1583 |
+
# Clear the memory here
|
| 1584 |
+
if freeze_text_encoder and not train_dataset.custom_instance_prompts:
|
| 1585 |
+
del tokenizers, text_encoders
|
| 1586 |
+
gc.collect()
|
| 1587 |
+
torch.cuda.empty_cache()
|
| 1588 |
+
|
| 1589 |
+
# If custom instance prompts are NOT provided (i.e. the instance prompt is used for all images),
|
| 1590 |
+
# pack the statically computed variables appropriately here. This is so that we don't
|
| 1591 |
+
# have to pass them to the dataloader.
|
| 1592 |
+
add_time_ids = instance_time_ids
|
| 1593 |
+
if args.with_prior_preservation:
|
| 1594 |
+
add_time_ids = torch.cat([add_time_ids, class_time_ids], dim=0)
|
| 1595 |
+
|
| 1596 |
+
# if --train_text_encoder_ti we need add_special_tokens to be True fo textual inversion
|
| 1597 |
+
add_special_tokens = True if args.train_text_encoder_ti else False
|
| 1598 |
+
|
| 1599 |
+
if not train_dataset.custom_instance_prompts:
|
| 1600 |
+
if freeze_text_encoder:
|
| 1601 |
+
prompt_embeds = instance_prompt_hidden_states
|
| 1602 |
+
unet_add_text_embeds = instance_pooled_prompt_embeds
|
| 1603 |
+
if args.with_prior_preservation:
|
| 1604 |
+
prompt_embeds = torch.cat([prompt_embeds, class_prompt_hidden_states], dim=0)
|
| 1605 |
+
unet_add_text_embeds = torch.cat([unet_add_text_embeds, class_pooled_prompt_embeds], dim=0)
|
| 1606 |
+
# if we're optmizing the text encoder (both if instance prompt is used for all images or custom prompts) we need to tokenize and encode the
|
| 1607 |
+
# batch prompts on all training steps
|
| 1608 |
+
else:
|
| 1609 |
+
tokens_one = tokenize_prompt(tokenizer_one, args.instance_prompt, add_special_tokens)
|
| 1610 |
+
tokens_two = tokenize_prompt(tokenizer_two, args.instance_prompt, add_special_tokens)
|
| 1611 |
+
if args.with_prior_preservation:
|
| 1612 |
+
class_tokens_one = tokenize_prompt(tokenizer_one, args.class_prompt, add_special_tokens)
|
| 1613 |
+
class_tokens_two = tokenize_prompt(tokenizer_two, args.class_prompt, add_special_tokens)
|
| 1614 |
+
tokens_one = torch.cat([tokens_one, class_tokens_one], dim=0)
|
| 1615 |
+
tokens_two = torch.cat([tokens_two, class_tokens_two], dim=0)
|
| 1616 |
+
|
| 1617 |
+
if args.train_text_encoder_ti and args.validation_prompt:
|
| 1618 |
+
# replace instances of --token_abstraction in validation prompt with the new tokens: "<si><si+1>" etc.
|
| 1619 |
+
for token_abs, token_replacement in train_dataset.token_abstraction_dict.items():
|
| 1620 |
+
args.validation_prompt = args.validation_prompt.replace(token_abs, "".join(token_replacement))
|
| 1621 |
+
print("validation prompt:", args.validation_prompt)
|
| 1622 |
+
|
| 1623 |
+
if args.cache_latents:
|
| 1624 |
+
latents_cache = []
|
| 1625 |
+
for batch in tqdm(train_dataloader, desc="Caching latents"):
|
| 1626 |
+
with torch.no_grad():
|
| 1627 |
+
batch["pixel_values"] = batch["pixel_values"].to(
|
| 1628 |
+
accelerator.device, non_blocking=True, dtype=torch.float32
|
| 1629 |
+
)
|
| 1630 |
+
latents_cache.append(vae.encode(batch["pixel_values"]).latent_dist)
|
| 1631 |
+
|
| 1632 |
+
if args.validation_prompt is None:
|
| 1633 |
+
del vae
|
| 1634 |
+
if torch.cuda.is_available():
|
| 1635 |
+
torch.cuda.empty_cache()
|
| 1636 |
+
|
| 1637 |
+
# Scheduler and math around the number of training steps.
|
| 1638 |
+
overrode_max_train_steps = False
|
| 1639 |
+
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
|
| 1640 |
+
if args.max_train_steps is None:
|
| 1641 |
+
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
|
| 1642 |
+
overrode_max_train_steps = True
|
| 1643 |
+
|
| 1644 |
+
lr_scheduler = get_scheduler(
|
| 1645 |
+
args.lr_scheduler,
|
| 1646 |
+
optimizer=optimizer,
|
| 1647 |
+
num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
|
| 1648 |
+
num_training_steps=args.max_train_steps * accelerator.num_processes,
|
| 1649 |
+
num_cycles=args.lr_num_cycles,
|
| 1650 |
+
power=args.lr_power,
|
| 1651 |
+
)
|
| 1652 |
+
|
| 1653 |
+
# Prepare everything with our `accelerator`.
|
| 1654 |
+
if not freeze_text_encoder:
|
| 1655 |
+
unet, text_encoder_one, text_encoder_two, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
|
| 1656 |
+
unet, text_encoder_one, text_encoder_two, optimizer, train_dataloader, lr_scheduler
|
| 1657 |
+
)
|
| 1658 |
+
else:
|
| 1659 |
+
unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
|
| 1660 |
+
unet, optimizer, train_dataloader, lr_scheduler
|
| 1661 |
+
)
|
| 1662 |
+
|
| 1663 |
+
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
|
| 1664 |
+
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
|
| 1665 |
+
if overrode_max_train_steps:
|
| 1666 |
+
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
|
| 1667 |
+
# Afterwards we recalculate our number of training epochs
|
| 1668 |
+
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
|
| 1669 |
+
|
| 1670 |
+
# We need to initialize the trackers we use, and also store our configuration.
|
| 1671 |
+
# The trackers initializes automatically on the main process.
|
| 1672 |
+
if accelerator.is_main_process:
|
| 1673 |
+
accelerator.init_trackers("dreambooth-lora-sd-xl", config=vars(args))
|
| 1674 |
+
|
| 1675 |
+
# Train!
|
| 1676 |
+
total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
|
| 1677 |
+
|
| 1678 |
+
logger.info("***** Running training *****")
|
| 1679 |
+
logger.info(f" Num examples = {len(train_dataset)}")
|
| 1680 |
+
logger.info(f" Num batches each epoch = {len(train_dataloader)}")
|
| 1681 |
+
logger.info(f" Num Epochs = {args.num_train_epochs}")
|
| 1682 |
+
logger.info(f" Instantaneous batch size per device = {args.train_batch_size}")
|
| 1683 |
+
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
|
| 1684 |
+
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
|
| 1685 |
+
logger.info(f" Total optimization steps = {args.max_train_steps}")
|
| 1686 |
+
global_step = 0
|
| 1687 |
+
first_epoch = 0
|
| 1688 |
+
|
| 1689 |
+
# Potentially load in the weights and states from a previous save
|
| 1690 |
+
if args.resume_from_checkpoint:
|
| 1691 |
+
if args.resume_from_checkpoint != "latest":
|
| 1692 |
+
path = os.path.basename(args.resume_from_checkpoint)
|
| 1693 |
+
else:
|
| 1694 |
+
# Get the mos recent checkpoint
|
| 1695 |
+
dirs = os.listdir(args.output_dir)
|
| 1696 |
+
dirs = [d for d in dirs if d.startswith("checkpoint")]
|
| 1697 |
+
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
|
| 1698 |
+
path = dirs[-1] if len(dirs) > 0 else None
|
| 1699 |
+
|
| 1700 |
+
if path is None:
|
| 1701 |
+
accelerator.print(
|
| 1702 |
+
f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
|
| 1703 |
+
)
|
| 1704 |
+
args.resume_from_checkpoint = None
|
| 1705 |
+
initial_global_step = 0
|
| 1706 |
+
else:
|
| 1707 |
+
accelerator.print(f"Resuming from checkpoint {path}")
|
| 1708 |
+
accelerator.load_state(os.path.join(args.output_dir, path))
|
| 1709 |
+
global_step = int(path.split("-")[1])
|
| 1710 |
+
|
| 1711 |
+
initial_global_step = global_step
|
| 1712 |
+
first_epoch = global_step // num_update_steps_per_epoch
|
| 1713 |
+
|
| 1714 |
+
else:
|
| 1715 |
+
initial_global_step = 0
|
| 1716 |
+
|
| 1717 |
+
progress_bar = tqdm(
|
| 1718 |
+
range(0, args.max_train_steps),
|
| 1719 |
+
initial=initial_global_step,
|
| 1720 |
+
desc="Steps",
|
| 1721 |
+
# Only show the progress bar once on each machine.
|
| 1722 |
+
disable=not accelerator.is_local_main_process,
|
| 1723 |
+
)
|
| 1724 |
+
|
| 1725 |
+
if args.train_text_encoder:
|
| 1726 |
+
num_train_epochs_text_encoder = int(args.train_text_encoder_frac * args.num_train_epochs)
|
| 1727 |
+
elif args.train_text_encoder_ti: # args.train_text_encoder_ti
|
| 1728 |
+
num_train_epochs_text_encoder = int(args.train_text_encoder_ti_frac * args.num_train_epochs)
|
| 1729 |
+
# flag used for textual inversion
|
| 1730 |
+
pivoted = False
|
| 1731 |
+
for epoch in range(first_epoch, args.num_train_epochs):
|
| 1732 |
+
# if performing any kind of optimization of text_encoder params
|
| 1733 |
+
if args.train_text_encoder or args.train_text_encoder_ti:
|
| 1734 |
+
if epoch == num_train_epochs_text_encoder:
|
| 1735 |
+
print("PIVOT HALFWAY", epoch)
|
| 1736 |
+
# stopping optimization of text_encoder params
|
| 1737 |
+
# this flag is used to reset the optimizer to optimize only on unet params
|
| 1738 |
+
pivoted = True
|
| 1739 |
+
|
| 1740 |
+
else:
|
| 1741 |
+
# still optimizing the text encoder
|
| 1742 |
+
text_encoder_one.train()
|
| 1743 |
+
text_encoder_two.train()
|
| 1744 |
+
# set top parameter requires_grad = True for gradient checkpointing works
|
| 1745 |
+
if args.train_text_encoder:
|
| 1746 |
+
text_encoder_one.text_model.embeddings.requires_grad_(True)
|
| 1747 |
+
text_encoder_two.text_model.embeddings.requires_grad_(True)
|
| 1748 |
+
|
| 1749 |
+
unet.train()
|
| 1750 |
+
for step, batch in enumerate(train_dataloader):
|
| 1751 |
+
if pivoted:
|
| 1752 |
+
# stopping optimization of text_encoder params
|
| 1753 |
+
# re setting the optimizer to optimize only on unet params
|
| 1754 |
+
optimizer.param_groups[1]["lr"] = 0.0
|
| 1755 |
+
optimizer.param_groups[2]["lr"] = 0.0
|
| 1756 |
+
|
| 1757 |
+
with accelerator.accumulate(unet):
|
| 1758 |
+
prompts = batch["prompts"]
|
| 1759 |
+
# encode batch prompts when custom prompts are provided for each image -
|
| 1760 |
+
if train_dataset.custom_instance_prompts:
|
| 1761 |
+
if freeze_text_encoder:
|
| 1762 |
+
prompt_embeds, unet_add_text_embeds = compute_text_embeddings(
|
| 1763 |
+
prompts, text_encoders, tokenizers
|
| 1764 |
+
)
|
| 1765 |
+
|
| 1766 |
+
else:
|
| 1767 |
+
tokens_one = tokenize_prompt(tokenizer_one, prompts, add_special_tokens)
|
| 1768 |
+
tokens_two = tokenize_prompt(tokenizer_two, prompts, add_special_tokens)
|
| 1769 |
+
|
| 1770 |
+
if args.cache_latents:
|
| 1771 |
+
model_input = latents_cache[step].sample()
|
| 1772 |
+
else:
|
| 1773 |
+
pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
|
| 1774 |
+
model_input = vae.encode(pixel_values).latent_dist.sample()
|
| 1775 |
+
|
| 1776 |
+
model_input = model_input * vae_scaling_factor
|
| 1777 |
+
if args.pretrained_vae_model_name_or_path is None:
|
| 1778 |
+
model_input = model_input.to(weight_dtype)
|
| 1779 |
+
|
| 1780 |
+
# Sample noise that we'll add to the latents
|
| 1781 |
+
noise = torch.randn_like(model_input)
|
| 1782 |
+
if args.noise_offset:
|
| 1783 |
+
# https://www.crosslabs.org//blog/diffusion-with-offset-noise
|
| 1784 |
+
noise += args.noise_offset * torch.randn(
|
| 1785 |
+
(model_input.shape[0], model_input.shape[1], 1, 1), device=model_input.device
|
| 1786 |
+
)
|
| 1787 |
+
|
| 1788 |
+
bsz = model_input.shape[0]
|
| 1789 |
+
# Sample a random timestep for each image
|
| 1790 |
+
timesteps = torch.randint(
|
| 1791 |
+
0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
|
| 1792 |
+
)
|
| 1793 |
+
timesteps = timesteps.long()
|
| 1794 |
+
|
| 1795 |
+
# Add noise to the model input according to the noise magnitude at each timestep
|
| 1796 |
+
# (this is the forward diffusion process)
|
| 1797 |
+
noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)
|
| 1798 |
+
|
| 1799 |
+
# Calculate the elements to repeat depending on the use of prior-preservation and custom captions.
|
| 1800 |
+
if not train_dataset.custom_instance_prompts:
|
| 1801 |
+
elems_to_repeat_text_embeds = bsz // 2 if args.with_prior_preservation else bsz
|
| 1802 |
+
elems_to_repeat_time_ids = bsz // 2 if args.with_prior_preservation else bsz
|
| 1803 |
+
|
| 1804 |
+
else:
|
| 1805 |
+
elems_to_repeat_text_embeds = 1
|
| 1806 |
+
elems_to_repeat_time_ids = bsz // 2 if args.with_prior_preservation else bsz
|
| 1807 |
+
|
| 1808 |
+
# Predict the noise residual
|
| 1809 |
+
if freeze_text_encoder:
|
| 1810 |
+
unet_added_conditions = {
|
| 1811 |
+
"time_ids": add_time_ids.repeat(elems_to_repeat_time_ids, 1),
|
| 1812 |
+
"text_embeds": unet_add_text_embeds.repeat(elems_to_repeat_text_embeds, 1),
|
| 1813 |
+
}
|
| 1814 |
+
prompt_embeds_input = prompt_embeds.repeat(elems_to_repeat_text_embeds, 1, 1)
|
| 1815 |
+
model_pred = unet(
|
| 1816 |
+
noisy_model_input,
|
| 1817 |
+
timesteps,
|
| 1818 |
+
prompt_embeds_input,
|
| 1819 |
+
added_cond_kwargs=unet_added_conditions,
|
| 1820 |
+
).sample
|
| 1821 |
+
else:
|
| 1822 |
+
unet_added_conditions = {"time_ids": add_time_ids.repeat(elems_to_repeat_time_ids, 1)}
|
| 1823 |
+
prompt_embeds, pooled_prompt_embeds = encode_prompt(
|
| 1824 |
+
text_encoders=[text_encoder_one, text_encoder_two],
|
| 1825 |
+
tokenizers=None,
|
| 1826 |
+
prompt=None,
|
| 1827 |
+
text_input_ids_list=[tokens_one, tokens_two],
|
| 1828 |
+
)
|
| 1829 |
+
unet_added_conditions.update(
|
| 1830 |
+
{"text_embeds": pooled_prompt_embeds.repeat(elems_to_repeat_text_embeds, 1)}
|
| 1831 |
+
)
|
| 1832 |
+
prompt_embeds_input = prompt_embeds.repeat(elems_to_repeat_text_embeds, 1, 1)
|
| 1833 |
+
model_pred = unet(
|
| 1834 |
+
noisy_model_input, timesteps, prompt_embeds_input, added_cond_kwargs=unet_added_conditions
|
| 1835 |
+
).sample
|
| 1836 |
+
|
| 1837 |
+
# Get the target for loss depending on the prediction type
|
| 1838 |
+
if noise_scheduler.config.prediction_type == "epsilon":
|
| 1839 |
+
target = noise
|
| 1840 |
+
elif noise_scheduler.config.prediction_type == "v_prediction":
|
| 1841 |
+
target = noise_scheduler.get_velocity(model_input, noise, timesteps)
|
| 1842 |
+
else:
|
| 1843 |
+
raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
|
| 1844 |
+
|
| 1845 |
+
if args.with_prior_preservation:
|
| 1846 |
+
# Chunk the noise and model_pred into two parts and compute the loss on each part separately.
|
| 1847 |
+
model_pred, model_pred_prior = torch.chunk(model_pred, 2, dim=0)
|
| 1848 |
+
target, target_prior = torch.chunk(target, 2, dim=0)
|
| 1849 |
+
|
| 1850 |
+
# Compute prior loss
|
| 1851 |
+
prior_loss = F.mse_loss(model_pred_prior.float(), target_prior.float(), reduction="mean")
|
| 1852 |
+
|
| 1853 |
+
if args.snr_gamma is None:
|
| 1854 |
+
loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
|
| 1855 |
+
else:
|
| 1856 |
+
# Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
|
| 1857 |
+
# Since we predict the noise instead of x_0, the original formulation is slightly changed.
|
| 1858 |
+
# This is discussed in Section 4.2 of the same paper.
|
| 1859 |
+
|
| 1860 |
+
if args.with_prior_preservation:
|
| 1861 |
+
# if we're using prior preservation, we calc snr for instance loss only -
|
| 1862 |
+
# and hence only need timesteps corresponding to instance images
|
| 1863 |
+
snr_timesteps, _ = torch.chunk(timesteps, 2, dim=0)
|
| 1864 |
+
else:
|
| 1865 |
+
snr_timesteps = timesteps
|
| 1866 |
+
|
| 1867 |
+
snr = compute_snr(noise_scheduler, snr_timesteps)
|
| 1868 |
+
base_weight = (
|
| 1869 |
+
torch.stack([snr, args.snr_gamma * torch.ones_like(snr_timesteps)], dim=1).min(dim=1)[0] / snr
|
| 1870 |
+
)
|
| 1871 |
+
|
| 1872 |
+
if noise_scheduler.config.prediction_type == "v_prediction":
|
| 1873 |
+
# Velocity objective needs to be floored to an SNR weight of one.
|
| 1874 |
+
mse_loss_weights = base_weight + 1
|
| 1875 |
+
else:
|
| 1876 |
+
# Epsilon and sample both use the same loss weights.
|
| 1877 |
+
mse_loss_weights = base_weight
|
| 1878 |
+
|
| 1879 |
+
loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
|
| 1880 |
+
loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
|
| 1881 |
+
loss = loss.mean()
|
| 1882 |
+
|
| 1883 |
+
if args.with_prior_preservation:
|
| 1884 |
+
# Add the prior loss to the instance loss.
|
| 1885 |
+
loss = loss + args.prior_loss_weight * prior_loss
|
| 1886 |
+
|
| 1887 |
+
accelerator.backward(loss)
|
| 1888 |
+
if accelerator.sync_gradients:
|
| 1889 |
+
params_to_clip = (
|
| 1890 |
+
itertools.chain(unet_lora_parameters, text_lora_parameters_one, text_lora_parameters_two)
|
| 1891 |
+
if (args.train_text_encoder or args.train_text_encoder_ti)
|
| 1892 |
+
else unet_lora_parameters
|
| 1893 |
+
)
|
| 1894 |
+
accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
|
| 1895 |
+
optimizer.step()
|
| 1896 |
+
lr_scheduler.step()
|
| 1897 |
+
optimizer.zero_grad()
|
| 1898 |
+
|
| 1899 |
+
# every step, we reset the embeddings to the original embeddings.
|
| 1900 |
+
if args.train_text_encoder_ti:
|
| 1901 |
+
embedding_handler.retract_embeddings()
|
| 1902 |
+
|
| 1903 |
+
# Checks if the accelerator has performed an optimization step behind the scenes
|
| 1904 |
+
if accelerator.sync_gradients:
|
| 1905 |
+
progress_bar.update(1)
|
| 1906 |
+
global_step += 1
|
| 1907 |
+
|
| 1908 |
+
if accelerator.is_main_process:
|
| 1909 |
+
if global_step % args.checkpointing_steps == 0:
|
| 1910 |
+
# _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
|
| 1911 |
+
if args.checkpoints_total_limit is not None:
|
| 1912 |
+
checkpoints = os.listdir(args.output_dir)
|
| 1913 |
+
checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
|
| 1914 |
+
checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
|
| 1915 |
+
|
| 1916 |
+
# before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
|
| 1917 |
+
if len(checkpoints) >= args.checkpoints_total_limit:
|
| 1918 |
+
num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
|
| 1919 |
+
removing_checkpoints = checkpoints[0:num_to_remove]
|
| 1920 |
+
|
| 1921 |
+
logger.info(
|
| 1922 |
+
f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
|
| 1923 |
+
)
|
| 1924 |
+
logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
|
| 1925 |
+
|
| 1926 |
+
for removing_checkpoint in removing_checkpoints:
|
| 1927 |
+
removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
|
| 1928 |
+
shutil.rmtree(removing_checkpoint)
|
| 1929 |
+
|
| 1930 |
+
save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
|
| 1931 |
+
accelerator.save_state(save_path)
|
| 1932 |
+
logger.info(f"Saved state to {save_path}")
|
| 1933 |
+
|
| 1934 |
+
logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
|
| 1935 |
+
progress_bar.set_postfix(**logs)
|
| 1936 |
+
accelerator.log(logs, step=global_step)
|
| 1937 |
+
|
| 1938 |
+
if global_step >= args.max_train_steps:
|
| 1939 |
+
break
|
| 1940 |
+
|
| 1941 |
+
if accelerator.is_main_process:
|
| 1942 |
+
if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
|
| 1943 |
+
logger.info(
|
| 1944 |
+
f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
|
| 1945 |
+
f" {args.validation_prompt}."
|
| 1946 |
+
)
|
| 1947 |
+
# create pipeline
|
| 1948 |
+
if freeze_text_encoder:
|
| 1949 |
+
text_encoder_one = text_encoder_cls_one.from_pretrained(
|
| 1950 |
+
args.pretrained_model_name_or_path,
|
| 1951 |
+
subfolder="text_encoder",
|
| 1952 |
+
revision=args.revision,
|
| 1953 |
+
variant=args.variant,
|
| 1954 |
+
)
|
| 1955 |
+
text_encoder_two = text_encoder_cls_two.from_pretrained(
|
| 1956 |
+
args.pretrained_model_name_or_path,
|
| 1957 |
+
subfolder="text_encoder_2",
|
| 1958 |
+
revision=args.revision,
|
| 1959 |
+
variant=args.variant,
|
| 1960 |
+
)
|
| 1961 |
+
pipeline = StableDiffusionXLPipeline.from_pretrained(
|
| 1962 |
+
args.pretrained_model_name_or_path,
|
| 1963 |
+
vae=vae,
|
| 1964 |
+
text_encoder=accelerator.unwrap_model(text_encoder_one),
|
| 1965 |
+
text_encoder_2=accelerator.unwrap_model(text_encoder_two),
|
| 1966 |
+
unet=accelerator.unwrap_model(unet),
|
| 1967 |
+
revision=args.revision,
|
| 1968 |
+
variant=args.variant,
|
| 1969 |
+
torch_dtype=weight_dtype,
|
| 1970 |
+
)
|
| 1971 |
+
|
| 1972 |
+
# We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
|
| 1973 |
+
scheduler_args = {}
|
| 1974 |
+
|
| 1975 |
+
if "variance_type" in pipeline.scheduler.config:
|
| 1976 |
+
variance_type = pipeline.scheduler.config.variance_type
|
| 1977 |
+
|
| 1978 |
+
if variance_type in ["learned", "learned_range"]:
|
| 1979 |
+
variance_type = "fixed_small"
|
| 1980 |
+
|
| 1981 |
+
scheduler_args["variance_type"] = variance_type
|
| 1982 |
+
|
| 1983 |
+
pipeline.scheduler = DPMSolverMultistepScheduler.from_config(
|
| 1984 |
+
pipeline.scheduler.config, **scheduler_args
|
| 1985 |
+
)
|
| 1986 |
+
|
| 1987 |
+
pipeline = pipeline.to(accelerator.device)
|
| 1988 |
+
pipeline.set_progress_bar_config(disable=True)
|
| 1989 |
+
|
| 1990 |
+
# run inference
|
| 1991 |
+
generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
|
| 1992 |
+
pipeline_args = {"prompt": args.validation_prompt}
|
| 1993 |
+
|
| 1994 |
+
with torch.cuda.amp.autocast():
|
| 1995 |
+
images = [
|
| 1996 |
+
pipeline(**pipeline_args, generator=generator).images[0]
|
| 1997 |
+
for _ in range(args.num_validation_images)
|
| 1998 |
+
]
|
| 1999 |
+
|
| 2000 |
+
for tracker in accelerator.trackers:
|
| 2001 |
+
if tracker.name == "tensorboard":
|
| 2002 |
+
np_images = np.stack([np.asarray(img) for img in images])
|
| 2003 |
+
tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
|
| 2004 |
+
if tracker.name == "wandb":
|
| 2005 |
+
tracker.log(
|
| 2006 |
+
{
|
| 2007 |
+
"validation": [
|
| 2008 |
+
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
|
| 2009 |
+
for i, image in enumerate(images)
|
| 2010 |
+
]
|
| 2011 |
+
}
|
| 2012 |
+
)
|
| 2013 |
+
|
| 2014 |
+
del pipeline
|
| 2015 |
+
torch.cuda.empty_cache()
|
| 2016 |
+
|
| 2017 |
+
# Save the lora layers
|
| 2018 |
+
accelerator.wait_for_everyone()
|
| 2019 |
+
if accelerator.is_main_process:
|
| 2020 |
+
unet = accelerator.unwrap_model(unet)
|
| 2021 |
+
unet = unet.to(torch.float32)
|
| 2022 |
+
unet_lora_layers = convert_state_dict_to_diffusers(get_peft_model_state_dict(unet))
|
| 2023 |
+
|
| 2024 |
+
if args.train_text_encoder:
|
| 2025 |
+
text_encoder_one = accelerator.unwrap_model(text_encoder_one)
|
| 2026 |
+
text_encoder_lora_layers = convert_state_dict_to_diffusers(
|
| 2027 |
+
get_peft_model_state_dict(text_encoder_one.to(torch.float32))
|
| 2028 |
+
)
|
| 2029 |
+
text_encoder_two = accelerator.unwrap_model(text_encoder_two)
|
| 2030 |
+
text_encoder_2_lora_layers = convert_state_dict_to_diffusers(
|
| 2031 |
+
get_peft_model_state_dict(text_encoder_two.to(torch.float32))
|
| 2032 |
+
)
|
| 2033 |
+
else:
|
| 2034 |
+
text_encoder_lora_layers = None
|
| 2035 |
+
text_encoder_2_lora_layers = None
|
| 2036 |
+
|
| 2037 |
+
StableDiffusionXLPipeline.save_lora_weights(
|
| 2038 |
+
save_directory=args.output_dir,
|
| 2039 |
+
unet_lora_layers=unet_lora_layers,
|
| 2040 |
+
text_encoder_lora_layers=text_encoder_lora_layers,
|
| 2041 |
+
text_encoder_2_lora_layers=text_encoder_2_lora_layers,
|
| 2042 |
+
)
|
| 2043 |
+
images = []
|
| 2044 |
+
if args.validation_prompt and args.num_validation_images > 0:
|
| 2045 |
+
# Final inference
|
| 2046 |
+
# Load previous pipeline
|
| 2047 |
+
vae = AutoencoderKL.from_pretrained(
|
| 2048 |
+
vae_path,
|
| 2049 |
+
subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
|
| 2050 |
+
revision=args.revision,
|
| 2051 |
+
variant=args.variant,
|
| 2052 |
+
torch_dtype=weight_dtype,
|
| 2053 |
+
)
|
| 2054 |
+
pipeline = StableDiffusionXLPipeline.from_pretrained(
|
| 2055 |
+
args.pretrained_model_name_or_path,
|
| 2056 |
+
vae=vae,
|
| 2057 |
+
revision=args.revision,
|
| 2058 |
+
variant=args.variant,
|
| 2059 |
+
torch_dtype=weight_dtype,
|
| 2060 |
+
)
|
| 2061 |
+
|
| 2062 |
+
# We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
|
| 2063 |
+
scheduler_args = {}
|
| 2064 |
+
|
| 2065 |
+
if "variance_type" in pipeline.scheduler.config:
|
| 2066 |
+
variance_type = pipeline.scheduler.config.variance_type
|
| 2067 |
+
|
| 2068 |
+
if variance_type in ["learned", "learned_range"]:
|
| 2069 |
+
variance_type = "fixed_small"
|
| 2070 |
+
|
| 2071 |
+
scheduler_args["variance_type"] = variance_type
|
| 2072 |
+
|
| 2073 |
+
pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, **scheduler_args)
|
| 2074 |
+
|
| 2075 |
+
# load attention processors
|
| 2076 |
+
pipeline.load_lora_weights(args.output_dir)
|
| 2077 |
+
|
| 2078 |
+
# run inference
|
| 2079 |
+
pipeline = pipeline.to(accelerator.device)
|
| 2080 |
+
generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
|
| 2081 |
+
images = [
|
| 2082 |
+
pipeline(args.validation_prompt, num_inference_steps=25, generator=generator).images[0]
|
| 2083 |
+
for _ in range(args.num_validation_images)
|
| 2084 |
+
]
|
| 2085 |
+
|
| 2086 |
+
for tracker in accelerator.trackers:
|
| 2087 |
+
if tracker.name == "tensorboard":
|
| 2088 |
+
np_images = np.stack([np.asarray(img) for img in images])
|
| 2089 |
+
tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
|
| 2090 |
+
if tracker.name == "wandb":
|
| 2091 |
+
tracker.log(
|
| 2092 |
+
{
|
| 2093 |
+
"test": [
|
| 2094 |
+
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
|
| 2095 |
+
for i, image in enumerate(images)
|
| 2096 |
+
]
|
| 2097 |
+
}
|
| 2098 |
+
)
|
| 2099 |
+
|
| 2100 |
+
if args.train_text_encoder_ti:
|
| 2101 |
+
embedding_handler.save_embeddings(
|
| 2102 |
+
f"{args.output_dir}/{args.output_dir}_emb.safetensors",
|
| 2103 |
+
)
|
| 2104 |
+
|
| 2105 |
+
# Conver to WebUI format
|
| 2106 |
+
lora_state_dict = load_file(f"{args.output_dir}/pytorch_lora_weights.safetensors")
|
| 2107 |
+
peft_state_dict = convert_all_state_dict_to_peft(lora_state_dict)
|
| 2108 |
+
kohya_state_dict = convert_state_dict_to_kohya(peft_state_dict)
|
| 2109 |
+
save_file(kohya_state_dict, f"{args.output_dir}/{args.output_dir}.safetensors")
|
| 2110 |
+
|
| 2111 |
+
save_model_card(
|
| 2112 |
+
model_id if not args.push_to_hub else repo_id,
|
| 2113 |
+
images=images,
|
| 2114 |
+
base_model=args.pretrained_model_name_or_path,
|
| 2115 |
+
train_text_encoder=args.train_text_encoder,
|
| 2116 |
+
train_text_encoder_ti=args.train_text_encoder_ti,
|
| 2117 |
+
token_abstraction_dict=train_dataset.token_abstraction_dict,
|
| 2118 |
+
instance_prompt=args.instance_prompt,
|
| 2119 |
+
validation_prompt=args.validation_prompt,
|
| 2120 |
+
repo_folder=args.output_dir,
|
| 2121 |
+
vae_path=args.pretrained_vae_model_name_or_path,
|
| 2122 |
+
)
|
| 2123 |
+
if args.push_to_hub:
|
| 2124 |
+
upload_folder(
|
| 2125 |
+
repo_id=repo_id,
|
| 2126 |
+
folder_path=args.output_dir,
|
| 2127 |
+
commit_message="End of training",
|
| 2128 |
+
ignore_patterns=["step_*", "epoch_*"],
|
| 2129 |
+
)
|
| 2130 |
+
|
| 2131 |
+
accelerator.end_training()
|
| 2132 |
+
|
| 2133 |
+
|
| 2134 |
+
if __name__ == "__main__":
|
| 2135 |
+
args = parse_args()
|
| 2136 |
+
main(args)
|