Switch ops in functional_ops, math_ops, state_ops, string_ops,
contrib/quantization, and contrib/layers to use C++ shape inference functions. Implement Betainc C++ shape inference function; it's a little different from the python one, mainly because propagating an unknown input (a_shape in this case) when all are unknown is not correct for C++ (in case we later backwards bind the value), since it could end up being a broadcasted scalar. Change ReduceJoin's C++ shape inference function to match python. Fix SymbolicGradient's shape function in C++ - there are actually more inputs than outputs. Changes QuantizedBiasAdd to be more precise (the C++ shape fn invokes the common bias_add to compute the shape of the first output). Change: 132688147
Loading
Please sign in to comment