[GitHub] leezu opened a new issue #11206: Embedding(sparse_grad=True) + hybridization fallback issues

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

leezu opened a new issue #11206: Embedding(sparse_grad=True) + hybridization fallback issues
URL: https://github.com/apache/incubator-mxnet/issues/11206
 
 
   With the mxnet master branch below code will fail. The combination of the following causes the failure:
   
   - Use of Embedding(sparse_grad=True) + other operations in one HybridBloc
   - Combining multiple forward passes during a single autograd record session
   
   ```import mxnet as mx
   
   
   class Embedding(mx.gluon.HybridBlock):
       def __init__(self, num_tokens, embedding_size):
           super().__init__()
           self.num_tokens = num_tokens
   
           with self.name_scope():
               self.embedding = mx.gluon.nn.Embedding(
                   num_tokens, embedding_size, sparse_grad=True)
   
       def hybrid_forward(self, F, words):
           emb = self.embedding(words)
           return emb + F.ones_like(emb)
   
   
   ctx = mx.cpu()
   embedding = Embedding(1000, 300)
   embedding.initialize(ctx=ctx)
   embedding.hybridize()
   
   loss_function = mx.gluon.loss.SigmoidBinaryCrossEntropyLoss()
   with mx.autograd.record():
       emb_in = embedding(mx.nd.arange(10, ctx=ctx))
       emb_in_2 = embedding(mx.nd.arange(10, ctx=ctx))
       loss = emb_in.sum() + emb_in_2.sum()
   loss.backward()
   
   print(embedding.embedding.weight.grad().data)
   ```
   
   
   ```
    Storage type fallback detected:
   operator = add_n
   input storage types = [default, default, ]
   output storage types = [row_sparse, ]
   params = {"num_args" : 2, }
   context.dev_mask = cpu
   The operator with default storage type will be dispatched for execution. You're seeing this warning message because the operator above is unable to process the given ndarrays with specified storage types, context and parameter. Temporary dense ndarrays are generated in order to execute the operator. This does not affect the correctness of the programme. You can set environment variable MXNET_STORAGE_FALLBACK_LOG_VERBOSE to 0 to suppress this warning.
   Traceback (most recent call last):
     File "sparse_bug.py", line 30, in <module>
       print(embedding.embedding.weight.grad().data)
     File "/Users/lllausen/anaconda3/lib/python3.6/site-packages/mxnet/ndarray/sparse.py", line 728, in data
       return self._data()
     File "/Users/lllausen/anaconda3/lib/python3.6/site-packages/mxnet/ndarray/sparse.py", line 266, in _data
       self.wait_to_read()
     File "/Users/lllausen/anaconda3/lib/python3.6/site-packages/mxnet/ndarray/ndarray.py", line 1720, in wait_to_read
       check_call(_LIB.MXNDArrayWaitToRead(self.handle))
     File "/Users/lllausen/anaconda3/lib/python3.6/site-packages/mxnet/base.py", line 210, in check_call
       raise MXNetError(py_str(_LIB.MXGetLastError()))
   mxnet.base.MXNetError: [19:37:39] src/operator/contrib/../operator_common.h:493: Not implemented: operator = _backward_Embedding
   input storage types = [default, default, ]
   output storage types = [default, default, ]
   params = {"input_dim" : 1000, "output_dim" : 300, "dtype" : float32, "sparse_grad" : True, }
   context.dev_mask = cpu
   
   ```

----------------------------------------------------------------
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:
users@infra.apache.org


With regards,
Apache Git Services