Computes gradient moment for the filter of a Conv2D operation. Assumes ``Z=tf.nn.conv2d(A, f)``, where ``f`` is a ``[h_f, w_f, c_in, c_out]`` convolution filter and ``A`` are the ``[n, h_in, w_in, c_in]`` activations of the previous layer (``n`` being the batch size). ``out_grad`` is the gr
(op, f, out_grad, batch_size, mom=2)
source not stored for this graph (policy: none)