Simplifying the code handling dim_ordering for tensorflow back-end

[MycChiu]

Tensorflow originally takes an input tensor of the shape
[batch, in_height, in_width, in_channels]
while dealing with the convolution-related operations,
and Theano takes an input of the shape
[batch, in_channels, in_height, in_width]
Therefore, currently in the master branch we are doing an additional tensor transpose when the dim_ordering complies with the theano convention.

However, since r0.8 Tensorflow has added support for different dim_ordering via the data_format argument. Right now, we can actually just pass in 'NCHW' (i.e. [n_sample,channel,height,width]) for the data_format when the dim_ordering is 'th', which will save a tf.transpose operation.

Additionally, when I run the benchmark files from soumith/convnet-benchmarks, I see a consistent 20% speed gain using 'NCHW' data_format while running on gpu. The effect seems to be caused by CuDNN using 'NCHW' data_format internally, and Tensorflow has to perform additional operation to convert the NHWC tensors to comply with CuDNN. Although I am not sure if this applies to all environments, (I am running on Ubuntu 16.04, GTX1070 with CUDA toolkit 8.0RC, CuDNN V5, keras 1.0.5, tensorflow 0.8.0 ) but if it does, I think it would be a good incentive to change the current way of dealing with different data formats.

[fchollet]

Thanks for pointing this out, but unfortunately support doesn't seem mature enough to switch at this point. In particular the CPU implementation only supports NHWC.

[ddkang]

Do you think this is worth revisiting?