PyTorch to ONNX no batch normalization layer

1

I have some very standard CNN-BatchNorm-relu combinations in my model, after I use torch.onnx.export(), the BatchNorm layer doesn’t exist any more in onnx model, I carefully checked the model and found that BN has been fused in CNN layer.
This happens after I update my pytorch to 1.7, my code used to work in 1.6. Is there a way to keep BN layer even after it is converted to onnx model?

2 Likes
2

Same problem, the BatchNorm Layers disappear after exporting to onnx format…
How did you find BN has been fused in CNN layer? In the source code of PyTorch?

3

See This Issue

2 Likes