Closed roywei closed 5 years ago
@sandeep-krishnamurthy could you help take a look? would like this to go in before MXNet 1.5.0 release. Now MXNet backend has more strick partial shape inference. So at this time, softmax shape is not inferred yet and should not go into the logic of ndim==1
I am concerned with too many specialization for MXNet backend.
Thanks for the review, agree and created issue https://github.com/awslabs/keras-apache-mxnet/issues/233
https://github.com/apache/incubator-mxnet/pull/14869 fix nightly test failure after unknown dim changed from 0 to -1 in MXNet. This fix softmax shape check in keras after the change. requires https://github.com/apache/incubator-mxnet/pull/14869 to go into nightly pip packages to pass CI.