Fix in vjpConv2DBackpropFilter

[Shashi456]

The differentiation is wrt (x, input):

/// TensorFlow builtin conv2d gradient helper for the filter.
@differentiable(wrt: (x, input), vjp: _vjpConv2DBackpropFilter)
@usableFromInline

While the order of the arguments returned in the vjp is input, x.
Another fix is in the input function the shape needed is the input shape and the one being used is filter shape.

[Shashi456]

@rxwei, Actually this is the case for depthwise and conv3d as well, Should i reflect the change there as well?

[dan-zheng]

@rxwei, Actually this is the case for depthwise and conv3d as well, Should i reflect the change there as well?

Yes, could you please change those as well?

Please also add differentiation tests for the updated operators in Tests/TensorFlowTests/TensorAutoDiffTests.swift. 🙂

You can check correctness against Python TF as is done here.

[Shashi456]

can this be tested and merged? Conv2d Differentiation test was added in #443, I'll do the same for conv3d and add a differentiation test today in another PR.

[rxwei]

Please do a git pull so that the test that was added will run as part of CI.

[Shashi456]

This can be merged.