Conv2d and BatchNorm need some attention while porting. I posted about the constant eps value handling in my message about porting batch normalization. I have ported resnet architectures from Tensorflow to Pytorch and there is no loss, the errors mostly creep in because of constants like eps values, shapes, stride implementation differences as I mentioned in this message. In my experience flipping is the NCH and NHC convention (N being batch, C being channel and H being height), if you flip everything while reading from TF model and setting it in Torch, it will work.