Spaces:
Sleeping
Sleeping
File size: 6,339 Bytes
3916ae1 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 |
from __future__ import annotations
import datetime
import os
import pathlib
import shlex
import shutil
import subprocess
import gradio as gr
import PIL.Image
import slugify
import torch
from huggingface_hub import HfApi
from accelerate.utils import write_basic_config
from app_upload import ModelUploader
from utils import save_model_card
URL_TO_JOIN_LIBRARY_ORG = 'https://huggingface.co/organizations/svdiff-library/share/PZBRRkosXikenXUdjMcvcoFmpWjcWnZjKL'
def pad_image(image: PIL.Image.Image) -> PIL.Image.Image:
w, h = image.size
if w == h:
return image
elif w > h:
new_image = PIL.Image.new(image.mode, (w, w), (0, 0, 0))
new_image.paste(image, (0, (w - h) // 2))
return new_image
else:
new_image = PIL.Image.new(image.mode, (h, h), (0, 0, 0))
new_image.paste(image, ((h - w) // 2, 0))
return new_image
class Trainer:
def __init__(self, hf_token: str | None = None):
self.hf_token = hf_token
self.api = HfApi(token=hf_token)
self.model_uploader = ModelUploader(hf_token)
def prepare_dataset(self, instance_images: list, resolution: int,
instance_data_dir: pathlib.Path) -> None:
shutil.rmtree(instance_data_dir, ignore_errors=True)
instance_data_dir.mkdir(parents=True)
for i, temp_path in enumerate(instance_images):
image = PIL.Image.open(temp_path.name)
image = pad_image(image)
image = image.resize((resolution, resolution))
image = image.convert('RGB')
out_path = instance_data_dir / f'{i:03d}.jpg'
image.save(out_path, format='JPEG', quality=100)
def join_library_org(self) -> None:
subprocess.run(
shlex.split(
f'curl -X POST -H "Authorization: Bearer {self.hf_token}" -H "Content-Type: application/json" {URL_TO_JOIN_LIBRARY_ORG}'
))
def run(
self,
instance_images: list | None,
instance_prompt: str,
output_model_name: str,
overwrite_existing_model: bool,
validation_prompt: str,
base_model: str,
resolution_s: str,
n_steps: int,
learning_rate: float,
gradient_accumulation: int,
seed: int,
fp16: bool,
use_8bit_adam: bool,
gradient_checkpointing: bool,
# enable_xformers_memory_efficient_attention: bool,
checkpointing_steps: int,
use_wandb: bool,
validation_epochs: int,
upload_to_hub: bool,
use_private_repo: bool,
delete_existing_repo: bool,
upload_to: str,
remove_gpu_after_training: bool,
) -> str:
if not torch.cuda.is_available():
raise gr.Error('CUDA is not available.')
if instance_images is None:
raise gr.Error('You need to upload images.')
if not instance_prompt:
raise gr.Error('The instance prompt is missing.')
if not validation_prompt:
raise gr.Error('The validation prompt is missing.')
resolution = int(resolution_s)
if not output_model_name:
timestamp = datetime.datetime.now().strftime('%Y-%m-%d-%H-%M-%S')
output_model_name = f'svdiff-pytorch-{timestamp}'
output_model_name = slugify.slugify(output_model_name)
repo_dir = pathlib.Path(__file__).parent
output_dir = repo_dir / 'experiments' / output_model_name
if overwrite_existing_model or upload_to_hub:
shutil.rmtree(output_dir, ignore_errors=True)
output_dir.mkdir(parents=True)
instance_data_dir = repo_dir / 'training_data' / output_model_name
self.prepare_dataset(instance_images, resolution, instance_data_dir)
if upload_to_hub:
self.join_library_org()
# accelerate config
write_basic_config()
command = f'''
accelerate launch train_svdiff.py \
--pretrained_model_name_or_path={base_model} \
--instance_data_dir={instance_data_dir} \
--output_dir={output_dir} \
--instance_prompt="{instance_prompt}" \
--resolution={resolution} \
--train_batch_size=1 \
--gradient_accumulation_steps={gradient_accumulation} \
--learning_rate={learning_rate} \
--learning_rate_1d=1e-6 \
--train_text_encoder \
--lr_scheduler=constant \
--lr_warmup_steps=0 \
--max_train_steps={n_steps} \
--checkpointing_steps={checkpointing_steps} \
--validation_prompt="{validation_prompt}" \
--validation_epochs={validation_epochs} \
--seed={seed}
'''
if fp16:
command += ' --mixed_precision="fp16"'
if use_8bit_adam:
command += ' --use_8bit_adam'
if gradient_checkpointing:
command += ' --gradient_checkpointing'
# if enable_xformers_memory_efficient_attention:
# command += ' --enable_xformers_memory_efficient_attention'
if use_wandb:
command += ' --report_to wandb'
with open(output_dir / 'train.sh', 'w') as f:
command_s = ' '.join(command.split())
f.write(command_s)
subprocess.run(shlex.split(command))
save_model_card(save_dir=output_dir,
base_model=base_model,
instance_prompt=instance_prompt,
test_prompt=validation_prompt,
test_image_dir='test_images')
message = 'Training completed!'
print(message)
if upload_to_hub:
upload_message = self.model_uploader.upload_model(
folder_path=output_dir.as_posix(),
repo_name=output_model_name,
upload_to=upload_to,
private=use_private_repo,
delete_existing_repo=delete_existing_repo)
print(upload_message)
message = message + '\n' + upload_message
if remove_gpu_after_training:
space_id = os.getenv('SPACE_ID')
if space_id:
self.api.request_space_hardware(repo_id=space_id,
hardware='cpu-basic')
return message
|