feevos commented on issue #9822: gluon HybridBlock wrapper of constant 
nd.array, is it possible?
URL: 
https://github.com/apache/incubator-mxnet/issues/9822#issuecomment-372877890
 
 
   Thanks @jmacglashan  you are right. On mxnet [discuss 
forum](https://discuss.mxnet.io/t/custom-layer-infer-shape-after-first-forward-pass/585/7)
 I was given a solution that does exactly what you describe:
   ```Python
   class CustomConv(HybridBlock):
       def __init__(self, const_ndarray, use_bias = True, **kwargs):
           super(CustomConv, self).__init__(**kwargs)
           self.use_bias = use_bias 
           
           with self.name_scope():
               self.weight = self.params.get('weight', 
                                             shape=(100, 100, 3, 3),
                                             allow_deferred_init=True)
               self.bijkl = self.params.get(
                   'bijkl', 
                   shape=const_ndarray.shape,
                   init=mx.init.Constant(const_ndarray.asnumpy().tolist()), 
                   differentiable=False)
   
           if self.use_bias:
                   self.bias = self.params.get(
                       'bias',
                       allow_deferred_init=True,
                       init = mx.init.Zero(),
                       shape=(100,))
           
       def hybrid_forward(self, F, x, weight, bijkl, bias=None):
           proj_weight = F.sum(F.dot(weight, bijkl), axis=[2, 3])
           
           if self.use_bias:
               return F.Convolution(data=x, weight=proj_weight, bias=bias, 
num_filter=100, kernel=(5, 5))
               
           else:
           
               return F.Convolution(data=x, weight=proj_weight, no_bias=True, 
num_filter=100, kernel=(5, 5))
   
   ```
   
   Only issue still left is that **I cannot (yet) infer the shape (of some 
dimension) during first run**.  Getting there ... 
   
   

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org


With regards,
Apache Git Services

Reply via email to