[GitHub] [incubator-mxnet] sxjscience commented on issue #16705: Dropout inconsistency bug

[GitBox]

sxjscience commented on issue #16705: Dropout inconsistency bug
URL: 
https://github.com/apache/incubator-mxnet/issues/16705#issuecomment-549533626
 
 
   @DickJC123 The answer should be different because these two dropouts should 
share the same internal random number generator and the random state will be 
updated accordingly.
   
   For the inconsistency bug mentioned in this issue, it's not exactly related 
to the seeding problem.
   
   For example, consider the following script:
   ```python
   import mxnet as mx
   mx.random.seed(123)
   x = mx.nd.ones((10, 10))
   
   y = mx.nd.Dropout(x, cudnn_off=True)
   with mx.autograd.record():
      y = mx.nd.Dropout(x, cudnn_off=True)
   ```
   The first `y = mx.nd.Dropout(x, cudnn_off=True)` is not surrounded by 
`autograd`, and should not update the random state. However, in the current 
implementation 
(https://github.com/apache/incubator-mxnet/blob/bb6305d11d4383af2022e53ad94d6a1d5d93cb00/src/operator/nn/dropout-inl.h#L495),
 the `rand()` function will still be called when the node is constructed.. 
Thus, running `y = mx.nd.Dropout(x, cudnn_off=True)` outside the `train` loop 
will still interfere the random state.
   
   This means, the following two code snippets will obtain different results:
   - Case 1
   ```python
   import mxnet as mx
   mx.random.seed(123)
   x = mx.nd.ones((3, 3), ctx=mx.gpu())
   
   y = mx.nd.Dropout(x, cudnn_off=True)
   with mx.autograd.record():
      y = mx.nd.Dropout(x, cudnn_off=True)
   print(y)
   ```
   ```
   [[0. 2. 0.]
    [0. 0. 2.]
    [0. 2. 0.]]
   <NDArray 3x3 @gpu(0)>
   ```
   
   - Case 2
   ```python
   import mxnet as mx
   mx.random.seed(123)
   x = mx.nd.ones((3, 3), ctx=mx.gpu())
   
   with mx.autograd.record():
      y = mx.nd.Dropout(x, cudnn_off=True)
   print(y)
   ```
   ```
   [[0. 0. 2.]
    [0. 0. 2.]
    [0. 2. 0.]]
   <NDArray 3x3 @gpu(0)>
   ```

----------------------------------------------------------------
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:
us...@infra.apache.org


With regards,
Apache Git Services