Spaces:
Running
Running
File size: 12,841 Bytes
32d4a86 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 |
"""
Model Optimization Script for Production Deployment
Reduces model size and improves inference speed through:
1. Quantization (INT8)
2. TorchScript compilation
3. Model pruning (optional)
4. State dict optimization
"""
import torch
import os
import argparse
from pathlib import Path
# Import model classes BEFORE loading checkpoints (needed for unpickling)
# This ensures PyTorch can find the class definitions when loading saved objects
# Note: resnet_train.py has module-level code that loads COCO data, which may fail
# if training files aren't present. We'll handle this in the functions.
def quantize_model(checkpoint_path, output_path, model_type='resnet'):
"""
Quantize model to INT8 for 4x size reduction and faster inference.
Note: Slight accuracy loss (usually <1%)
"""
print(f"Quantizing {model_type} model...")
device = torch.device('cpu') # Quantization typically done on CPU
# Import classes before loading (required for unpickling)
# resnet_train.py now handles missing training data gracefully
if model_type == 'resnet':
# Import the module itself so we can update vocab later
import resnet_train
from resnet_train import EncoderCNN, DecoderRNN, Vocabulary
# Make Vocabulary available in __main__ for unpickling
# This handles cases where checkpoint was saved with Vocabulary from __main__
import __main__
if not hasattr(__main__, 'Vocabulary'):
__main__.Vocabulary = Vocabulary
elif model_type == 'efficientnet':
from efficient_train import Encoder, Decoder, ImageCaptioningModel
from transformers import AutoTokenizer
# Load checkpoint (now all classes are available for unpickling)
checkpoint = torch.load(checkpoint_path, map_location=device, weights_only=False)
if model_type == 'resnet':
# For ResNet, quantize encoder and decoder separately
# IMPORTANT: Update vocab from checkpoint before creating DecoderRNN
# The decoder uses len(vocab.word2idx) in its __init__, so we need the full vocab
if 'vocab' in checkpoint and checkpoint['vocab'] is not None:
# Update the vocab in resnet_train module (DecoderRNN.__init__ references resnet_train.vocab)
resnet_train.vocab = checkpoint['vocab']
print(f" Updated vocab size: {len(checkpoint['vocab'].word2idx)}")
else:
raise ValueError("Checkpoint does not contain 'vocab' key. Cannot proceed.")
encoder = EncoderCNN()
decoder = DecoderRNN() # Now uses the correct vocab size from checkpoint
encoder.load_state_dict(checkpoint['encoder'])
decoder.load_state_dict(checkpoint['decoder'])
# Set to eval mode
encoder.eval()
decoder.eval()
# Prepare for quantization (dummy input)
dummy_input = torch.randn(1, 3, 224, 224)
# Quantize encoder (only Linear and Conv2d layers)
encoder_quantized = torch.quantization.quantize_dynamic(
encoder, {torch.nn.Linear, torch.nn.Conv2d}, dtype=torch.qint8
)
# Quantize decoder (only Linear layers - Embedding requires special config)
# Embeddings are typically small and don't benefit much from quantization
decoder_quantized = torch.quantization.quantize_dynamic(
decoder, {torch.nn.Linear}, dtype=torch.qint8
)
# Save quantized model
quantized_checkpoint = {
'encoder': encoder_quantized.state_dict(),
'decoder': decoder_quantized.state_dict(),
'vocab': checkpoint.get('vocab'),
'quantized': True
}
elif model_type == 'efficientnet':
# Classes already imported above before loading checkpoint
tokenizer = AutoTokenizer.from_pretrained('gpt2')
tokenizer.pad_token = tokenizer.eos_token
special_tokens = {'additional_special_tokens': ['<start>', '<end>']}
tokenizer.add_special_tokens(special_tokens)
encoder = Encoder(model_name='efficientnet_b3', embed_dim=512)
decoder = Decoder(
vocab_size=len(tokenizer),
embed_dim=512,
num_layers=8,
num_heads=8,
max_seq_length=64
)
model = ImageCaptioningModel(encoder, decoder)
# Load state dict - handle both 'model_state' key and direct state dict
if 'model_state' in checkpoint:
model.load_state_dict(checkpoint['model_state'])
else:
model.load_state_dict(checkpoint)
model.eval()
# Quantize the full model
model_quantized = torch.quantization.quantize_dynamic(
model, {torch.nn.Linear, torch.nn.Conv2d}, dtype=torch.qint8
)
quantized_checkpoint = {
'model_state': model_quantized.state_dict(),
'quantized': True
}
torch.save(quantized_checkpoint, output_path)
# Compare sizes
original_size = os.path.getsize(checkpoint_path) / (1024 * 1024) # MB
quantized_size = os.path.getsize(output_path) / (1024 * 1024) # MB
reduction = (1 - quantized_size / original_size) * 100
print(f"✓ Quantization complete!")
print(f" Original size: {original_size:.2f} MB")
print(f" Quantized size: {quantized_size:.2f} MB")
print(f" Size reduction: {reduction:.1f}%")
return output_path
def optimize_state_dict(checkpoint_path, output_path):
"""
Remove unnecessary metadata and optimize state dict for smaller size.
"""
print(f"Optimizing state dict...")
# Import classes before loading (required for unpickling)
try:
from resnet_train import Vocabulary
# Make Vocabulary available in __main__ for unpickling
import __main__
if not hasattr(__main__, 'Vocabulary'):
__main__.Vocabulary = Vocabulary
except ImportError:
pass
checkpoint = torch.load(checkpoint_path, map_location='cpu', weights_only=False)
# Create optimized checkpoint with only essential data
optimized = {}
for key, value in checkpoint.items():
if key not in ['optimizer', 'scheduler', 'epoch', 'loss', 'metrics']:
optimized[key] = value
# Save with highest compression
torch.save(optimized, output_path, _use_new_zipfile_serialization=True)
original_size = os.path.getsize(checkpoint_path) / (1024 * 1024)
optimized_size = os.path.getsize(output_path) / (1024 * 1024)
reduction = (1 - optimized_size / original_size) * 100
print(f"✓ State dict optimized!")
print(f" Original: {original_size:.2f} MB")
print(f" Optimized: {optimized_size:.2f} MB")
print(f" Reduction: {reduction:.1f}%")
return output_path
def create_torchscript(checkpoint_path, output_path, model_type='resnet'):
"""
Convert model to TorchScript for faster loading and inference.
Note: Requires example input for tracing.
"""
print(f"Creating TorchScript model...")
device = torch.device('cpu')
# Import classes before loading (required for unpickling)
if model_type == 'resnet':
import resnet_train
from resnet_train import EncoderCNN, DecoderRNN, Vocabulary
# Make Vocabulary available in __main__ for unpickling
import __main__
if not hasattr(__main__, 'Vocabulary'):
__main__.Vocabulary = Vocabulary
elif model_type == 'efficientnet':
from efficient_train import Encoder, Decoder, ImageCaptioningModel
from transformers import AutoTokenizer
checkpoint = torch.load(checkpoint_path, map_location=device, weights_only=False)
if model_type == 'resnet':
# Update vocab from checkpoint before creating DecoderRNN
if 'vocab' in checkpoint and checkpoint['vocab'] is not None:
resnet_train.vocab = checkpoint['vocab']
print(f" Updated vocab size: {len(checkpoint['vocab'].word2idx)}")
else:
raise ValueError("Checkpoint does not contain 'vocab' key. Cannot proceed.")
encoder = EncoderCNN().eval()
decoder = DecoderRNN().eval() # Now uses the correct vocab size
encoder.load_state_dict(checkpoint['encoder'])
decoder.load_state_dict(checkpoint['decoder'])
# Trace encoder
dummy_image = torch.randn(1, 3, 224, 224)
encoder_traced = torch.jit.trace(encoder, dummy_image)
# For decoder, we need to trace with proper inputs
# This is more complex due to RNN structure
print(" ⚠ TorchScript for RNN decoder may require manual scripting")
print(" ✓ Encoder traced successfully")
torch.jit.save(encoder_traced, output_path.replace('.pth', '_encoder.pt'))
elif model_type == 'efficientnet':
# Classes already imported above
tokenizer = AutoTokenizer.from_pretrained('gpt2')
tokenizer.pad_token = tokenizer.eos_token
special_tokens = {'additional_special_tokens': ['<start>', '<end>']}
tokenizer.add_special_tokens(special_tokens)
encoder = Encoder(model_name='efficientnet_b3', embed_dim=512)
decoder = Decoder(
vocab_size=len(tokenizer),
embed_dim=512,
num_layers=8,
num_heads=8,
max_seq_length=64
)
model = ImageCaptioningModel(encoder, decoder).eval()
model.load_state_dict(checkpoint['model_state'])
# Trace encoder only (decoder has dynamic inputs)
dummy_image = torch.randn(1, 3, 224, 224)
encoder_traced = torch.jit.trace(model.encoder, dummy_image)
torch.jit.save(encoder_traced, output_path.replace('.pth', '_encoder.pt'))
print(" ✓ Encoder traced successfully")
print(f"✓ TorchScript saved to {output_path}")
return output_path
def main():
parser = argparse.ArgumentParser(description='Optimize models for production deployment')
parser.add_argument('--model', type=str, choices=['resnet', 'efficientnet', 'both'],
default='both', help='Model to optimize')
parser.add_argument('--method', type=str, choices=['quantize', 'optimize', 'torchscript', 'all'],
default='all', help='Optimization method')
parser.add_argument('--resnet-path', type=str, default='resnet_best_model.pth',
help='Path to ResNet checkpoint')
parser.add_argument('--efficientnet-path', type=str, default='efficient_best_model.pth',
help='Path to EfficientNet checkpoint')
parser.add_argument('--output-dir', type=str, default='optimized_models',
help='Output directory for optimized models')
args = parser.parse_args()
# Create output directory
os.makedirs(args.output_dir, exist_ok=True)
models_to_process = []
if args.model in ['resnet', 'both']:
if os.path.exists(args.resnet_path):
models_to_process.append(('resnet', args.resnet_path))
else:
print(f"⚠ Warning: {args.resnet_path} not found, skipping ResNet")
if args.model in ['efficientnet', 'both']:
if os.path.exists(args.efficientnet_path):
models_to_process.append(('efficientnet', args.efficientnet_path))
else:
print(f"⚠ Warning: {args.efficientnet_path} not found, skipping EfficientNet")
if not models_to_process:
print("❌ No models found to optimize!")
return
for model_type, model_path in models_to_process:
print(f"\n{'='*60}")
print(f"Processing {model_type.upper()} model")
print(f"{'='*60}")
base_name = Path(model_path).stem
output_base = os.path.join(args.output_dir, f"{model_type}_{base_name}")
if args.method in ['quantize', 'all']:
quantized_path = f"{output_base}_quantized.pth"
quantize_model(model_path, quantized_path, model_type)
if args.method in ['optimize', 'all']:
optimized_path = f"{output_base}_optimized.pth"
optimize_state_dict(model_path, optimized_path)
if args.method in ['torchscript', 'all']:
torchscript_path = f"{output_base}_torchscript.pt"
create_torchscript(model_path, torchscript_path, model_type)
print(f"\n{'='*60}")
print("✓ Optimization complete!")
print(f"Optimized models saved to: {args.output_dir}")
print(f"{'='*60}")
if __name__ == '__main__':
main()
|