Created by: pangyoki

PR types

Bug fixes

PR changes

APIs

Describe

Fix the sample method of Uniform and Normal class .

For Uniform class： The shape of inputs of elementwise_add op in sample method is not matching. In detail, in the following code: output = uniform_random_tmp * (zero_tmp + self.high - self.low) + self.low the shape of zero_tmp is [batch_shape, sample_shape]， the shape of self.high is [batch_shape]. When zero_tmp and self.high doing elementwise_add operation, it will cause "Broadcast dimension dismatch".

To solve this problem, before doing elementwise_add operation, reshape zero_tmp. That is, change the shape of zero_tmp from [batch_shape, sample_shape] to [sample_shape, batch_shape]. At the same time, reshape uniform_random_tmp to [sample_shape, batch_shape].

For Normal class, it has the same problem, using the reshape method too.