Problems in Converting PNASNet_mobile to pytorch

[tzzcl]

Hi, I just want to use your code to convert PNASNet_mobile models to PyTorch.

After modifiing your code with both PNASNet.pytorch and PNASNet.TF,

I find that the accuracy is slow on converted models with PNASNet.pytorch
Test: [50000/50000] Prec@1 0.710 Prec@5 0.901
But In PNASNet.TF, the accuracy is 74.2

The shape and dict test is pass. But the Check 2 failed.

Could you give me some hints on converting the PNASNet-mobile models?

Thanks

[chenxi116]

Hmm, I am not sure where it went wrong. I suggest running Check 2 at intermediate layers, to progressively find where the conversion began to fail.