Wrong batch normalization axis for classifier layers, using tf

[natih]

I think there might be a bug when using tf, where batch normalization is set to bn_axis=3 in conv_block_td ann identity_block_td that serve the classifier layers in resnet50.
The classifier layers input tensor after ROI pooling is of shape [numBatches, numBoxes, poolSize, poolSize, numChannels], so bn_axis should better be set to -1.