Spaces:
Running
Running
Commit
·
ee07cd3
1
Parent(s):
134ce68
ADD: fix train_model
Browse files- train_model.py +172 -56
train_model.py
CHANGED
@@ -4,41 +4,75 @@ from datasets import load_dataset
|
|
4 |
from PIL import Image, ImageOps, ImageFilter
|
5 |
from tqdm import tqdm
|
6 |
import random
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
|
8 |
def preprocess_image(image, target_size=512, quality_threshold=0.7):
|
9 |
"""Preprocess image with various enhancements"""
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
return None
|
18 |
-
|
19 |
-
# Center crop to square if not already
|
20 |
-
if width != height:
|
21 |
-
size = min(width, height)
|
22 |
-
left = (width - size) // 2
|
23 |
-
top = (height - size) // 2
|
24 |
-
image = image.crop((left, top, left + size, top + size))
|
25 |
-
|
26 |
-
# Resize to target size
|
27 |
-
image = image.resize((target_size, target_size), Image.Resampling.LANCZOS)
|
28 |
-
|
29 |
-
# Enhance image quality
|
30 |
-
# Slightly sharpen
|
31 |
-
image = image.filter(ImageFilter.UnsharpMask(radius=0.5, percent=120, threshold=3))
|
32 |
-
|
33 |
-
# Auto-adjust levels
|
34 |
-
image = ImageOps.autocontrast(image, cutoff=1)
|
35 |
-
|
36 |
-
return image
|
37 |
|
38 |
def clean_prompt(prompt):
|
39 |
"""Clean and normalize prompts"""
|
40 |
if not prompt:
|
41 |
-
return
|
42 |
|
43 |
# Remove excessive whitespace
|
44 |
prompt = ' '.join(prompt.split())
|
@@ -56,53 +90,131 @@ def clean_prompt(prompt):
|
|
56 |
|
57 |
def prepare_dreambooth_data():
|
58 |
# Load dataset
|
59 |
-
|
60 |
-
|
61 |
|
62 |
# Create directory structure
|
63 |
-
data_dir = "./
|
64 |
os.makedirs(data_dir, exist_ok=True)
|
65 |
|
66 |
valid_samples = 0
|
|
|
|
|
|
|
|
|
67 |
|
68 |
# Process images with preprocessing
|
69 |
-
for idx, sample in enumerate(tqdm(
|
70 |
-
|
71 |
-
|
72 |
-
if image is None:
|
73 |
-
continue
|
74 |
|
75 |
-
|
76 |
-
|
77 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
78 |
continue
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
79 |
|
80 |
-
|
81 |
-
image_path = os.path.join(data_dir, f"image_{valid_samples:04d}.jpg")
|
82 |
-
image.save(image_path, "JPEG", quality=95, optimize=True)
|
83 |
|
84 |
-
# Save
|
85 |
-
|
|
|
|
|
|
|
86 |
with open(caption_path, 'w', encoding='utf-8') as f:
|
87 |
f.write(prompt)
|
88 |
-
|
89 |
-
valid_samples += 1
|
90 |
|
91 |
-
print(f"
|
92 |
return data_dir
|
93 |
|
94 |
-
#
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
|
|
99 |
accelerate launch \\
|
100 |
--deepspeed_config_file ds_config.json \\
|
101 |
diffusers/examples/dreambooth/train_dreambooth.py \\
|
102 |
--pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5" \\
|
103 |
--instance_data_dir="{data_dir}" \\
|
104 |
-
--instance_prompt="a
|
105 |
-
--output_dir="./
|
106 |
--resolution=512 \\
|
107 |
--train_batch_size=1 \\
|
108 |
--gradient_accumulation_steps=1 \\
|
@@ -115,8 +227,12 @@ accelerate launch \\
|
|
115 |
--checkpointing_steps=100 \\
|
116 |
--checkpoints_total_limit=1 \\
|
117 |
--report_to="tensorboard" \\
|
118 |
-
--logging_dir="./
|
119 |
"""
|
|
|
|
|
|
|
|
|
120 |
|
121 |
-
|
122 |
-
|
|
|
4 |
from PIL import Image, ImageOps, ImageFilter
|
5 |
from tqdm import tqdm
|
6 |
import random
|
7 |
+
import requests
|
8 |
+
import io
|
9 |
+
import time
|
10 |
+
|
11 |
+
def download_image(url, timeout=10, retries=2):
|
12 |
+
"""Download image from URL with retry mechanism"""
|
13 |
+
for attempt in range(retries):
|
14 |
+
try:
|
15 |
+
headers = {
|
16 |
+
'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36'
|
17 |
+
}
|
18 |
+
response = requests.get(url, timeout=timeout, headers=headers)
|
19 |
+
|
20 |
+
if response.status_code == 200:
|
21 |
+
image = Image.open(io.BytesIO(response.content))
|
22 |
+
return image
|
23 |
+
else:
|
24 |
+
return None
|
25 |
+
|
26 |
+
except Exception as e:
|
27 |
+
if attempt == retries - 1: # Last attempt
|
28 |
+
print(f"Failed to download {url}: {e}")
|
29 |
+
return None
|
30 |
+
time.sleep(0.5) # Brief pause before retry
|
31 |
+
|
32 |
+
return None
|
33 |
|
34 |
def preprocess_image(image, target_size=512, quality_threshold=0.7):
|
35 |
"""Preprocess image with various enhancements"""
|
36 |
+
if image is None:
|
37 |
+
return None
|
38 |
+
|
39 |
+
try:
|
40 |
+
# Convert to RGB if needed
|
41 |
+
if image.mode != 'RGB':
|
42 |
+
image = image.convert('RGB')
|
43 |
+
|
44 |
+
# Filter out low quality images
|
45 |
+
width, height = image.size
|
46 |
+
if min(width, height) < target_size * quality_threshold:
|
47 |
+
return None
|
48 |
+
|
49 |
+
# Center crop to square if not already
|
50 |
+
if width != height:
|
51 |
+
size = min(width, height)
|
52 |
+
left = (width - size) // 2
|
53 |
+
top = (height - size) // 2
|
54 |
+
image = image.crop((left, top, left + size, top + size))
|
55 |
+
|
56 |
+
# Resize to target size
|
57 |
+
image = image.resize((target_size, target_size), Image.Resampling.LANCZOS)
|
58 |
+
|
59 |
+
# Enhance image quality
|
60 |
+
# Slightly sharpen
|
61 |
+
image = image.filter(ImageFilter.UnsharpMask(radius=0.5, percent=120, threshold=3))
|
62 |
+
|
63 |
+
# Auto-adjust levels
|
64 |
+
image = ImageOps.autocontrast(image, cutoff=1)
|
65 |
+
|
66 |
+
return image
|
67 |
+
|
68 |
+
except Exception as e:
|
69 |
+
print(f"Error preprocessing image: {e}")
|
70 |
return None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
71 |
|
72 |
def clean_prompt(prompt):
|
73 |
"""Clean and normalize prompts"""
|
74 |
if not prompt:
|
75 |
+
return None
|
76 |
|
77 |
# Remove excessive whitespace
|
78 |
prompt = ' '.join(prompt.split())
|
|
|
90 |
|
91 |
def prepare_dreambooth_data():
|
92 |
# Load dataset
|
93 |
+
print("Loading LAION dataset...")
|
94 |
+
dataset = load_dataset("laion/laion2B-en-aesthetic", split="train", streaming=True)
|
95 |
|
96 |
# Create directory structure
|
97 |
+
data_dir = "./laion_dataset"
|
98 |
os.makedirs(data_dir, exist_ok=True)
|
99 |
|
100 |
valid_samples = 0
|
101 |
+
processed_count = 0
|
102 |
+
max_samples = 1000 # Limit total samples to process
|
103 |
+
|
104 |
+
print(f"Starting to process up to {max_samples} samples...")
|
105 |
|
106 |
# Process images with preprocessing
|
107 |
+
for idx, sample in enumerate(tqdm(dataset, desc="Processing LAION samples")):
|
108 |
+
if processed_count >= max_samples:
|
109 |
+
break
|
|
|
|
|
110 |
|
111 |
+
processed_count += 1
|
112 |
+
|
113 |
+
try:
|
114 |
+
# Get URL and text from LAION format
|
115 |
+
image_url = sample.get('URL', '')
|
116 |
+
text_prompt = sample.get('TEXT', '')
|
117 |
+
|
118 |
+
if not image_url or not text_prompt:
|
119 |
+
continue
|
120 |
+
|
121 |
+
# Clean prompt first
|
122 |
+
prompt = clean_prompt(text_prompt)
|
123 |
+
if prompt is None:
|
124 |
+
continue
|
125 |
+
|
126 |
+
# Download image from URL
|
127 |
+
print(f"Downloading image {valid_samples + 1}: {image_url[:50]}...")
|
128 |
+
image = download_image(image_url)
|
129 |
+
if image is None:
|
130 |
+
continue
|
131 |
+
|
132 |
+
# Preprocess downloaded image
|
133 |
+
processed_image = preprocess_image(image)
|
134 |
+
if processed_image is None:
|
135 |
+
continue
|
136 |
+
|
137 |
+
# Save processed image
|
138 |
+
image_path = os.path.join(data_dir, f"image_{valid_samples:04d}.jpg")
|
139 |
+
processed_image.save(image_path, "JPEG", quality=95, optimize=True)
|
140 |
+
|
141 |
+
# Save cleaned caption
|
142 |
+
caption_path = os.path.join(data_dir, f"image_{valid_samples:04d}.txt")
|
143 |
+
with open(caption_path, 'w', encoding='utf-8') as f:
|
144 |
+
f.write(prompt)
|
145 |
+
|
146 |
+
valid_samples += 1
|
147 |
+
|
148 |
+
# Optional: Add metadata file
|
149 |
+
metadata_path = os.path.join(data_dir, f"image_{valid_samples-1:04d}_meta.txt")
|
150 |
+
with open(metadata_path, 'w', encoding='utf-8') as f:
|
151 |
+
f.write(f"URL: {image_url}\n")
|
152 |
+
f.write(f"Aesthetic: {sample.get('aesthetic', 'N/A')}\n")
|
153 |
+
f.write(f"Width: {sample.get('WIDTH', 'N/A')}\n")
|
154 |
+
f.write(f"Height: {sample.get('HEIGHT', 'N/A')}\n")
|
155 |
+
|
156 |
+
# Stop if we have enough samples
|
157 |
+
if valid_samples >= 100: # Adjust this number as needed
|
158 |
+
break
|
159 |
+
|
160 |
+
except Exception as e:
|
161 |
+
print(f"Error processing sample {idx}: {e}")
|
162 |
continue
|
163 |
+
|
164 |
+
print(f"Processed {processed_count} samples, saved {valid_samples} valid images to {data_dir}")
|
165 |
+
return data_dir
|
166 |
+
|
167 |
+
def create_demo_dataset():
|
168 |
+
"""Create demo dataset as last resort"""
|
169 |
+
print("Creating demo dataset...")
|
170 |
+
|
171 |
+
data_dir = "./demo_dataset"
|
172 |
+
os.makedirs(data_dir, exist_ok=True)
|
173 |
+
|
174 |
+
demo_prompts = [
|
175 |
+
"a beautiful landscape with mountains",
|
176 |
+
"portrait of a person with detailed features",
|
177 |
+
"abstract colorful digital artwork",
|
178 |
+
"modern architecture building design",
|
179 |
+
"natural forest scene with trees",
|
180 |
+
"urban cityscape at sunset",
|
181 |
+
"artistic oil painting style",
|
182 |
+
"vintage photography aesthetic",
|
183 |
+
"minimalist geometric composition",
|
184 |
+
"vibrant surreal art piece"
|
185 |
+
]
|
186 |
+
|
187 |
+
for idx, prompt in enumerate(demo_prompts):
|
188 |
+
# Create gradient background
|
189 |
+
color1 = (random.randint(50, 200), random.randint(50, 200), random.randint(50, 200))
|
190 |
+
color2 = (random.randint(100, 255), random.randint(100, 255), random.randint(100, 255))
|
191 |
|
192 |
+
image = Image.new('RGB', (512, 512), color1)
|
|
|
|
|
193 |
|
194 |
+
# Save files
|
195 |
+
image_path = os.path.join(data_dir, f"image_{idx:04d}.jpg")
|
196 |
+
image.save(image_path, "JPEG", quality=95)
|
197 |
+
|
198 |
+
caption_path = os.path.join(data_dir, f"image_{idx:04d}.txt")
|
199 |
with open(caption_path, 'w', encoding='utf-8') as f:
|
200 |
f.write(prompt)
|
|
|
|
|
201 |
|
202 |
+
print(f"Created {len(demo_prompts)} demo samples")
|
203 |
return data_dir
|
204 |
|
205 |
+
# Main execution with fallback
|
206 |
+
def main():
|
207 |
+
data_dir = prepare_dreambooth_data()
|
208 |
+
|
209 |
+
# Generate training command
|
210 |
+
training_command = f"""
|
211 |
accelerate launch \\
|
212 |
--deepspeed_config_file ds_config.json \\
|
213 |
diffusers/examples/dreambooth/train_dreambooth.py \\
|
214 |
--pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5" \\
|
215 |
--instance_data_dir="{data_dir}" \\
|
216 |
+
--instance_prompt="a high quality image" \\
|
217 |
+
--output_dir="./laion-model" \\
|
218 |
--resolution=512 \\
|
219 |
--train_batch_size=1 \\
|
220 |
--gradient_accumulation_steps=1 \\
|
|
|
227 |
--checkpointing_steps=100 \\
|
228 |
--checkpoints_total_limit=1 \\
|
229 |
--report_to="tensorboard" \\
|
230 |
+
--logging_dir="./laion-model/logs"
|
231 |
"""
|
232 |
+
|
233 |
+
print(f"\n✅ Dataset prepared in: {data_dir}")
|
234 |
+
print("🚀 Run this command to train:")
|
235 |
+
print(training_command)
|
236 |
|
237 |
+
if __name__ == "__main__":
|
238 |
+
main()
|