[GitHub] [incubator-mxnet] sxjscience commented on a change in pull request #18199: [Bug Fix] Fix GroupNorm Implementation

[GitBox <gi...@apache.org>]

sxjscience commented on a change in pull request #18199:
URL: https://github.com/apache/incubator-mxnet/pull/18199#discussion_r417932015



##########
File path: python/mxnet/gluon/nn/basic_layers.py
##########
@@ -820,26 +820,29 @@ class GroupNorm(HybridBlock):
     """
     def __init__(self, num_groups=1, epsilon=1e-5, center=True, scale=True,
                  beta_initializer='zeros', gamma_initializer='ones',
-                 prefix=None, params=None):
+                 in_channels=0, prefix=None, params=None):
         super(GroupNorm, self).__init__(prefix=prefix, params=params)
         self._kwargs = {'eps': epsilon, 'num_groups': num_groups, 'center': center, 'scale': scale}
         self._num_groups = num_groups
         self._epsilon = epsilon
         self._center = center
         self._scale = scale
         self.gamma = self.params.get('gamma', grad_req='write' if scale else 'null',
-                                     shape=(num_groups,), init=gamma_initializer,
+                                     shape=(in_channels,), init=gamma_initializer,
                                      allow_deferred_init=True)
         self.beta = self.params.get('beta', grad_req='write' if center else 'null',
-                                    shape=(num_groups,), init=beta_initializer,
+                                    shape=(in_channels,), init=beta_initializer,
                                     allow_deferred_init=True)
 
     def hybrid_forward(self, F, data, gamma, beta):
         norm_data = F.GroupNorm(data, gamma=gamma, beta=beta, num_groups=self._num_groups, eps=self._epsilon)

Review comment:
       I just realized one quick issue. Should we consider to move GroupNorm to npx? Currently, the layer won’t be usable in the new numpy interface. @zhreshold 




----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org