I'm trying to convert an old tensorflow/keras network I have to pytorch and I'm confused as to the values I obtain of the batch_normalization (BN) weights. For example, the first BN layer of my pytorch network has two sets of elements (weight and bias) of shape [8]. In contrast, the same BN layer in tensorflow returns 4 sets of elements of shape [8].
Since I'm not planning on retraining the network once I've converted it to pytorch, I'd like to know which two sets of elements of the tensorflow BN layer are the corresponding for the pytorch BN layer.