File size: 2,794 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 | """
ONNX parity debug - pinpoint the stage-11 parity failure.
Checks, on a fixed seeded input:
1. Initializer census: does the ONNX graph contain ~4,012,672 params?
2. Torch CPU logits vs ORT CPU logits (same input).
3. Torch CUDA logits vs ORT CUDA logits.
4. Per-output margin between argmax classes.
"""
from pathlib import Path
import numpy as np
import onnx
import torch
import onnxruntime as ort
from models import create_model
from utils import load_config
CHECKPOINT = "models/best_epoch019_macro-recall0.9929.pth"
ONNX_PATH = "models/efficientnet_b0.onnx"
CONFIG = "configs/efficientnet_b0.yaml"
def main() -> None:
torch.manual_seed(0)
dummy = torch.randn(1, 3, 224, 224)
# --- 1. Initializer census ----------------------------------------
model_proto = onnx.load(ONNX_PATH)
init_elems = sum(np.prod(list(i.dims)) for i in model_proto.graph.initializer)
print(f"Graph initializers: {len(model_proto.graph.initializer)} tensors")
print(f"Initializer element count: {init_elems:,} (expected ~4,012,672)")
print(f"Graph opset: {[op.version for op in model_proto.opset_import]}")
# --- 2. Torch CPU ---------------------------------------------------
config = load_config(CONFIG)
ckpt = torch.load(CHECKPOINT, map_location="cpu", weights_only=False)
model = create_model(ckpt.get("config", config))
model.load_state_dict(ckpt["model_state_dict"])
model.eval()
with torch.no_grad():
torch_logits = model(dummy)
print(f"\nTorch CPU logits: {torch_logits.numpy().round(3)}")
# --- 3. ORT CPU ------------------------------------------------------
session_cpu = ort.InferenceSession(str(ONNX_PATH), providers=["CPUExecutionProvider"])
onnx_cpu = session_cpu.run(None, {"input": dummy.numpy()})[0]
print(f"ORT CPU logits: {onnx_cpu.round(3)}")
print(f"CPU diff: max {np.abs(torch_logits.numpy() - onnx_cpu).max():.4f}")
# --- 4. Torch CUDA vs ORT CUDA --------------------------------------
if torch.cuda.is_available():
model_cuda = model.cuda()
with torch.no_grad():
torch_cuda = model_cuda(dummy.cuda()).cpu().numpy()
session_gpu = ort.InferenceSession(
str(ONNX_PATH), providers=["CUDAExecutionProvider", "CPUExecutionProvider"]
)
print(f"ONNX CUDA active provider: {session_gpu.get_providers()[0]}")
onnx_gpu = session_gpu.run(None, {"input": dummy.numpy()})[0]
print(f"Torch CUDA logits: {torch_cuda.round(3)}")
print(f"ORT CUDA logits: {onnx_gpu.round(3)}")
print(f"CUDA diff torch-vs-ort: max {np.abs(torch_cuda - onnx_gpu).max():.4f}")
print(f"CUDA diff cpu-vs-cuda(ort): max {np.abs(onnx_cpu - onnx_gpu).max():.4f}")
if __name__ == "__main__":
main()
|