[GitHub] ThomasDelteil opened a new issue #10766: Bug Cannot save/load params with Gluon model

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

ThomasDelteil opened a new issue #10766: Bug Cannot save/load params with Gluon model
URL: https://github.com/apache/incubator-mxnet/issues/10766
 
 
   See the issue reported here:
   https://discuss.mxnet.io/t/problem-when-loading-saved-params/982/11
   
   ```python
   self.model.load_params(self.model_path, ctx=ctx)
   File “/home/name/virtualEnv/local/lib/python2.7/site-packages/mxnet/gluon/block.py”, line 317, in load_params
   self.prefix)
   File “/home/name/virtualEnv/local/lib/python2.7/site-packages/mxnet/gluon/parameter.py”, line 676, in load
   self[name]._load_init(arg_dict[name], ctx)
   File “/home/name/virtualEnv/local/lib/python2.7/site-packages/mxnet/gluon/parameter.py”, line 209, in _load_init
   assert set(ctx) == set(self._deferred_init[1]), 
   IndexError: tuple index out of range
   ```
   
   Code [here](https://github.com/ZhengzheYang/Reading-Comprehension/blob/19c591c767f7022730e8ec3aee6502f91f613266/model.py#L232)
   
   Reproduce bug:
   ```python
   embedding_size = 400
   vocab_size = 100
   batch_size = 1
   ctx = mx.gpu()
   ​
   # Create first model
   test = RCv1(nd.ones((vocab_size, embedding_size)), vocab_size=vocab_size, batch_size=batch_size)
   # Initialize params
   test.model.collect_params().initialize(mx.init.Xavier(magnitude=2.24, rnd_type='gaussian'), ctx=ctx)
   # Runs one batch
   print('Model 1', test.model(nd.ones((batch_size, 25, 94, embedding_size), ctx), nd.ones((batch_size, 126), ctx)))
   # Save params
   test.model.save_params('test.params')
   # Create second model
   test2 = RCv1(nd.ones((vocab_size, embedding_size)), vocab_size=vocab_size, batch_size=batch_size)
   # Load params
   test2.model.load_params('test.params')
   ```
   ​
   I was able to reproduce. Initializing the parameters from the newly created network prior to loading the parameters solved the issue. Couldn't figure out exactly what is wrong. I feel it could be related to the `embedding` layer being shared by the LSTM and CNN.
   
   @piiswrong any ideas?

----------------------------------------------------------------
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