File size: 17,452 Bytes
1f2e262 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 | """
ONNX Export Script - Stage 11
Export PyTorch model to ONNX format with verification and benchmarking.
Hardware: RTX 3050 6GB (CUDA preferred but not required for export)
ONNX export is more stable on CPU, but verification can use GPU.
"""
import argparse
import sys
import time
from pathlib import Path
from typing import Optional, Tuple
import numpy as np
import onnx
import onnxruntime as ort
import torch
import torch.nn as nn
# Add src to path
sys.path.insert(0, str(Path(__file__).parent))
from models import create_model, get_model_info
from utils import load_config, set_seed
def parse_args() -> argparse.Namespace:
"""Parse command line arguments."""
parser = argparse.ArgumentParser(
description="Export brain tumor classifier to ONNX with verification"
)
parser.add_argument(
"--checkpoint",
type=str,
required=True,
help="Path to checkpoint file (.pth)"
)
parser.add_argument(
"--config",
type=str,
required=True,
help="Path to YAML config file (e.g., configs/efficientnet_b0.yaml)"
)
parser.add_argument(
"--output",
type=str,
default=None,
help="Output ONNX path (default: models/<model_name>.onnx)"
)
parser.add_argument(
"--opset-version",
type=int,
default=17,
help="ONNX opset version (default: 17)"
)
parser.add_argument(
"--simplify",
action="store_true",
help="Simplify ONNX model using onnx-simplifier"
)
parser.add_argument(
"--benchmark",
action="store_true",
help="Run performance benchmark (100 iterations)"
)
return parser.parse_args()
def setup_export(config: dict) -> Tuple[str, Path]:
"""
Setup export environment.
Args:
config: Configuration dictionary
Returns:
Tuple of (device, models_dir)
"""
print("=" * 70)
print("Brain Tumor MRI Classifier - ONNX Export")
print("=" * 70)
# Check if CUDA is available (not required, but preferred for verification)
cuda_available = torch.cuda.is_available()
device = "cuda" if cuda_available else "cpu"
if cuda_available:
print(f"GPU: {torch.cuda.get_device_name(0)}")
print(f"CUDA Version: {torch.version.cuda}")
else:
print("GPU: Not available (CPU-only mode)")
print(f"PyTorch Version: {torch.__version__}")
print(f"ONNX Version: {onnx.__version__}")
print(f"ONNX Runtime Version: {ort.__version__}")
print()
# Set seed for reproducibility
seed = config.get('seed', 42)
set_seed(seed)
# Create models directory
workspace_root = Path.cwd()
models_dir = workspace_root / "models"
models_dir.mkdir(parents=True, exist_ok=True)
print(f"Models directory: {models_dir}")
print()
return device, models_dir
def load_model_from_checkpoint(
checkpoint_path: str,
config: dict,
device: str
) -> nn.Module:
"""
Load model from checkpoint.
Args:
checkpoint_path: Path to checkpoint file
config: Configuration dictionary
device: Device to load model on (for verification)
Returns:
Model loaded with checkpoint weights
"""
print("=" * 70)
print("Loading Model")
print("=" * 70)
print(f"Checkpoint: {checkpoint_path}")
# Load checkpoint
# weights_only=False: our own checkpoints contain numpy arrays in `metrics`.
checkpoint = torch.load(checkpoint_path, map_location='cpu', weights_only=False)
# Get config from checkpoint if available
model_config = checkpoint.get('config', config)
# Create model
model = create_model(model_config)
# Load state dict
model.load_state_dict(checkpoint['model_state_dict'])
# Model info
info = get_model_info(model)
print(f"Model: {model_config['model']['name']}")
print(f"Parameters: {info['total_params']:,}")
print(f"Model size: {info['model_size_mb']:.2f} MB")
if 'epoch' in checkpoint:
print(f"Checkpoint epoch: {checkpoint['epoch']}")
if 'metrics' in checkpoint:
val_metrics = checkpoint['metrics']
if 'macro_recall' in val_metrics:
print(f"Validation macro recall: {val_metrics['macro_recall']:.4f}")
print()
# Set to eval mode
model.eval()
return model
def export_to_onnx(
model: nn.Module,
output_path: Path,
opset_version: int,
config: dict
) -> None:
"""
Export PyTorch model to ONNX format.
Args:
model: PyTorch model to export
output_path: Path to save ONNX model
opset_version: ONNX opset version
config: Configuration dictionary
"""
print("=" * 70)
print("Exporting to ONNX")
print("=" * 70)
# Move model to CPU for stable export
model = model.cpu()
model.eval()
# Get input size from config
img_size = config['data']['img_size']
# Create dummy input
dummy_input = torch.randn(1, 3, img_size, img_size)
print(f"Input shape: {list(dummy_input.shape)}")
print(f"Output path: {output_path}")
print(f"Opset version: {opset_version}")
print()
# Dynamic axes for batch flexibility
dynamic_axes = {
'input': {0: 'batch_size'},
'output': {0: 'batch_size'}
}
try:
# Export to ONNX via the legacy TorchScript exporter.
# NOTE: torch >= 2.14 defaults to dynamo=True, which produced a
# 0.49 MB graph missing most weights (parity check failed with
# max_diff ~59). dynamo=False is the stable path for this model.
# dynamo=False also supports opset 17 natively, without auto
# version conversion.
print("Exporting model...")
torch.onnx.export(
model,
(dummy_input,),
str(output_path),
export_params=True,
opset_version=opset_version,
do_constant_folding=True,
input_names=['input'],
output_names=['output'],
dynamic_axes=dynamic_axes,
verbose=False,
dynamo=False,
)
print(f"✓ Export successful: {output_path}")
# Get file size
file_size_mb = output_path.stat().st_size / (1024 ** 2)
print(f" ONNX model size: {file_size_mb:.2f} MB")
print()
except Exception as e:
print(f"✗ Export failed: {e}")
raise
def simplify_onnx(onnx_path: Path) -> None:
"""
Simplify ONNX model using onnx-simplifier.
Args:
onnx_path: Path to ONNX model (will be overwritten)
"""
print("=" * 70)
print("Simplifying ONNX Model")
print("=" * 70)
try:
from onnxsim import simplify
print("Loading ONNX model...")
onnx_model = onnx.load(str(onnx_path))
print("Simplifying...")
model_simplified, check = simplify(onnx_model)
if check:
# Save simplified model
onnx.save(model_simplified, str(onnx_path))
print(f"✓ Simplification successful")
# Get new file size
file_size_mb = onnx_path.stat().st_size / (1024 ** 2)
print(f" Simplified model size: {file_size_mb:.2f} MB")
else:
print("✗ Simplification check failed (keeping original)")
print()
except ImportError:
print("✗ onnx-simplifier not installed. Skipping simplification.")
print(" Install with: pip install onnx-simplifier")
print()
except Exception as e:
print(f"✗ Simplification failed: {e}")
print(" Keeping original ONNX model.")
print()
def verify_onnx(
onnx_path: Path,
pytorch_model: nn.Module,
device: str,
config: dict,
verbose: bool = True
) -> bool:
"""
Verify ONNX model correctness.
Checks:
1. ONNX model is valid
2. ONNX Runtime can load and run it
3. Outputs match PyTorch model (within tolerance)
Args:
onnx_path: Path to ONNX model
pytorch_model: Original PyTorch model
device: Device for PyTorch inference
config: Configuration dictionary
verbose: Whether to print detailed results
Returns:
True if verification passed, False otherwise
"""
print("=" * 70)
print("Verifying ONNX Model")
print("=" * 70)
try:
# 1. Check ONNX model validity
if verbose:
print("1. Checking ONNX model validity...")
onnx_model = onnx.load(str(onnx_path))
onnx.checker.check_model(onnx_model)
if verbose:
print(" ✓ ONNX model is valid")
# 2. Load with ONNX Runtime
if verbose:
print("2. Loading with ONNX Runtime...")
# Choose execution provider
providers = []
if device == "cuda" and "CUDAExecutionProvider" in ort.get_available_providers():
providers.append("CUDAExecutionProvider")
if verbose:
print(" Using CUDAExecutionProvider")
providers.append("CPUExecutionProvider")
ort_session = ort.InferenceSession(str(onnx_path), providers=providers)
if verbose:
print(f" ✓ ONNX Runtime session created")
print(f" Active provider: {ort_session.get_providers()[0]}")
# 3. Compare outputs
if verbose:
print("3. Comparing outputs with PyTorch model...")
# Create test input.
# NOTE: synthetic inputs (gaussian or uniform) are a bad parity
# probe for this model — the trained BatchNorm statistics amplify
# OOD activations so much that tiny CPU/GPU convolution
# differences blow the logits to ~±3000, causing flaky failures.
# Real val-split images (test split stays locked) agree to <5e-3,
# verified separately in src/debug_realimage.py.
import pandas as pd
from data.transforms import get_val_transforms
img_size = config['data']['img_size']
processed_dir = Path.cwd() / config['data']['processed_dir']
manifest = pd.read_csv(processed_dir / "manifest.csv")
val_df = manifest[manifest["split"] == "val"].head(4)
transform = get_val_transforms(config)
test_input = torch.stack([
transform(np.load(processed_dir / fp))
for fp in val_df["filepath"]
])
if verbose:
print(f" Parity inputs: {len(val_df)} real val images")
# PyTorch inference
pytorch_model = pytorch_model.to(device)
pytorch_model.eval()
with torch.no_grad():
test_input_device = test_input.to(device)
pytorch_output = pytorch_model(test_input_device)
pytorch_output = pytorch_output.cpu().numpy()
# ONNX Runtime inference
ort_inputs = {ort_session.get_inputs()[0].name: test_input.numpy()}
ort_output = ort_session.run(None, ort_inputs)[0]
# Compute differences
max_diff = np.abs(pytorch_output - ort_output).max()
mean_diff = np.abs(pytorch_output - ort_output).mean()
if verbose:
print(f" Max absolute difference: {max_diff:.6f}")
print(f" Mean absolute difference: {mean_diff:.6f}")
# Check parity
# For float32, use 1e-4 tolerance
# For mixed precision models, use 1e-2 tolerance (more lenient)
amp_enabled = config.get('training', {}).get('use_amp', False)
tolerance = 1e-2 if amp_enabled else 1e-4
if max_diff < tolerance:
if verbose:
print(f" ✓ Output parity check passed (tolerance: {tolerance})")
print()
print("✓ Verification successful")
print()
return True
else:
if verbose:
print(f" ✗ Output parity check failed (max_diff={max_diff:.6f} > {tolerance})")
print()
print("✗ Verification failed: outputs do not match")
print()
return False
except Exception as e:
print(f"✗ Verification failed with error: {e}")
print()
return False
def benchmark_models(
onnx_path: Path,
pytorch_model: nn.Module,
device: str,
config: dict,
num_iterations: int = 100
) -> None:
"""
Benchmark PyTorch vs ONNX Runtime inference speed.
Args:
onnx_path: Path to ONNX model
pytorch_model: Original PyTorch model
device: Device for inference
config: Configuration dictionary
num_iterations: Number of iterations for benchmarking
"""
print("=" * 70)
print("Benchmarking Performance")
print("=" * 70)
print(f"Running {num_iterations} iterations on each model...")
print()
# Get input size
img_size = config['data']['img_size']
# Create test input
test_input = torch.randn(1, 3, img_size, img_size)
# --- PyTorch Benchmark ---
pytorch_model = pytorch_model.to(device)
pytorch_model.eval()
test_input_device = test_input.to(device)
# Warmup
with torch.no_grad():
for _ in range(10):
_ = pytorch_model(test_input_device)
if device == "cuda":
torch.cuda.synchronize()
# Benchmark
pytorch_times = []
with torch.no_grad():
for _ in range(num_iterations):
start = time.perf_counter()
_ = pytorch_model(test_input_device)
if device == "cuda":
torch.cuda.synchronize()
end = time.perf_counter()
pytorch_times.append(end - start)
pytorch_avg = np.mean(pytorch_times) * 1000 # Convert to ms
pytorch_std = np.std(pytorch_times) * 1000
# --- ONNX Runtime Benchmark ---
providers = []
if device == "cuda" and "CUDAExecutionProvider" in ort.get_available_providers():
providers.append("CUDAExecutionProvider")
providers.append("CPUExecutionProvider")
ort_session = ort.InferenceSession(str(onnx_path), providers=providers)
ort_input_name = ort_session.get_inputs()[0].name
test_input_numpy = test_input.numpy()
# Warmup
for _ in range(10):
_ = ort_session.run(None, {ort_input_name: test_input_numpy})
# Benchmark
onnx_times = []
for _ in range(num_iterations):
start = time.perf_counter()
_ = ort_session.run(None, {ort_input_name: test_input_numpy})
end = time.perf_counter()
onnx_times.append(end - start)
onnx_avg = np.mean(onnx_times) * 1000 # Convert to ms
onnx_std = np.std(onnx_times) * 1000
# --- Results ---
print(f"PyTorch ({device.upper()}):")
print(f" Average: {pytorch_avg:.2f} ± {pytorch_std:.2f} ms")
print(f" Throughput: {1000/pytorch_avg:.1f} images/sec")
print()
provider = ort_session.get_providers()[0]
print(f"ONNX Runtime ({provider}):")
print(f" Average: {onnx_avg:.2f} ± {onnx_std:.2f} ms")
print(f" Throughput: {1000/onnx_avg:.1f} images/sec")
print()
# Speedup
speedup = pytorch_avg / onnx_avg
if speedup > 1.0:
print(f"Speedup: {speedup:.2f}x faster with ONNX Runtime")
elif speedup < 1.0:
print(f"Speedup: {1/speedup:.2f}x slower with ONNX Runtime")
else:
print(f"Speedup: equivalent performance")
print()
# Model size
onnx_size_mb = onnx_path.stat().st_size / (1024 ** 2)
print(f"ONNX model size: {onnx_size_mb:.2f} MB")
print()
def main():
"""Main entry point."""
# Parse arguments
args = parse_args()
# Load config
config = load_config(args.config)
# Setup
device, models_dir = setup_export(config)
# Determine output path
if args.output is None:
model_name = config['model']['name']
output_path = models_dir / f"{model_name}.onnx"
else:
output_path = Path(args.output)
output_path.parent.mkdir(parents=True, exist_ok=True)
# Load model from checkpoint
model = load_model_from_checkpoint(args.checkpoint, config, device)
# Export to ONNX
try:
export_to_onnx(model, output_path, args.opset_version, config)
except Exception as e:
print(f"Export failed: {e}")
sys.exit(1)
# Simplify if requested
if args.simplify:
simplify_onnx(output_path)
# Verify ONNX model
verification_passed = verify_onnx(output_path, model, device, config)
if not verification_passed:
print("Warning: Verification failed. ONNX model may not produce correct results.")
print()
# Benchmark if requested
if args.benchmark:
benchmark_models(output_path, model, device, config)
# Summary
print("=" * 70)
print("Export Complete")
print("=" * 70)
print(f"ONNX model saved to: {output_path}")
print(f"Verification: {'✓ Passed' if verification_passed else '✗ Failed'}")
print()
print("Done!")
if __name__ == "__main__":
main()
|