Pretrained inception_v3 not matched with research paper

1

Hello,
This inception_v3 are not matching same as just mentioned in table 1. from this research paper. link
image

In first part, why it is of kernel_size=1? it should have been of kernel_size=3
In second part, where is stride value? it should have been of stride=2
In third part, there shouldn’t been no pool layer mentioned in paper? why did you even implemented it?

I think there is lot of stuff not matching to this research paper. I would like pytorch team kindly look into this matter and resolve as soon as possible.

Thanks

2

The reference for the inception v3 model in TorchVision is the TensorFlow implementation, not the paper.

See

and some more discussion about weights in

Best regards

Thomas

1 Like
3

Thanks for clearing that this implementation is not based on research paper.