It could be that I've forgot to specify it, but the blocks syntax should allow for the user to specify that a dimension stretches to the size of the input. In Caffe we do this with negative numbers, e.g. if block_c = -1 is given by the user, then at runtime we replace it with block_c = <number of input channels>. I believe we should adopt the same syntax for Tensorflow, with a slight improvement: in python, the operator should allow lists of length 2, and automatically expand the list from [block_h block_w] to [-1 block_h block_w] -- this should be for both the low-level Tensorflow python's ops, as well as in Keras. The same logic should be applied to the padding and strides syntax as well.
It could be that I've forgot to specify it, but the blocks syntax should allow for the user to specify that a dimension stretches to the size of the input. In Caffe we do this with negative numbers, e.g. if
block_c = -1is given by the user, then at runtime we replace it withblock_c = <number of input channels>. I believe we should adopt the same syntax for Tensorflow, with a slight improvement: in python, the operator should allow lists of length 2, and automatically expand the list from[block_h block_w]to[-1 block_h block_w]-- this should be for both the low-level Tensorflow python's ops, as well as in Keras. The same logic should be applied to the padding and strides syntax as well.