[GitHub] feevos opened a new issue #9316: bug in parameter default values of gluon.loss.Loss

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

feevos opened a new issue #9316: bug in parameter default values of gluon.loss.Loss
URL: https://github.com/apache/incubator-mxnet/issues/9316
 
 
   ## Description
   
   (I think!) there is a bug in the default values of the parameters ```weight``` and ```batch_axis``` of the module ```gluon.loss.Loss```. In particular when I try to create a custom loss function derived from Loss I get the following error: 
   
   **implementation that produces error**
   ```Python
   
   from mxnet.gluon.loss import Loss
   
   class jaccard(Loss):
       """
       Jaccard loss coefficient. Adopted from tensorlayer: 
       https://github.com/zsdonghao/tensorlayer/blob/master/tensorlayer/cost.py
       INPUT:
           tensor of size (Nbatch, Nclasses, W,H)
       OUTPUT:
             The average (over batch) of the average value (over classes). Can I do better? 
             TODO: Use a weight for each class?  See example in tests/Notebooks/custom_loss.ipynb
       """
   
       def __init__(self, _smooth=1.0e-3, _axis=[2,3], _weight = None, _batch_axis= 0, **kwards):
           Loss.__init__(self, **kwards)
   
           with self.name_scope():
               self.smooth = _smooth
               self.axis = _axis
   
   
       def hybrid_forward(self,F,_label, _output):
   
           itrs = F.sum(_label*_output, axis=self.axis)
           l = F.sum(_label*_label, axis=self.axis)
           r = F.sum(_output*_output, axis=self.axis)
   
           IoU = (2.0*itrs+self.smooth)/(l+r+self.smooth)
   
           meanIoU = F.mean(IoU,axis=1)
           meanTot = F.mean(meanIoU)
   
           return meanTot
   
   
   jacc_loss = jaccard()
   ```
   
   ## Error Message:
   
   ```Python
   TypeError                                 Traceback (most recent call last)
   <ipython-input-5-8b8b7a97ea46> in <module>()
   ----> 1 jacc_loss = jaccard()
   
   <ipython-input-4-fed5e9fc1f68> in __init__(self, _smooth, _axis, _weight, _batch_axis, **kwards)
        11 
        12     def __init__(self, _smooth=1.0e-3, _axis=[2,3], _weight = None, _batch_axis= 0, **kwards):
   ---> 13         Loss.__init__(self, **kwards)
        14 
        15         with self.name_scope():
   
   TypeError: __init__() takes exactly 3 arguments (1 given)
   
   ```
   
   ## What have you tried to solve it?
   
   If I  modify the constructor of Loss by providing the default values of the Loss (as described in the [documentation](https://mxnet.incubator.apache.org/api/python/gluon/loss.html) :
   
   ```Python
   class jaccard(Loss):
       """
       Jaccard loss coefficient. Adopted from tensorlayer: 
       https://github.com/zsdonghao/tensorlayer/blob/master/tensorlayer/cost.py
       INPUT:
           tensor of size (Nbatch, Nclasses, W,H)
       OUTPUT:
             The average (over batch) of the average value (over classes). Can I do better? 
             TODO: Use a weight for each class?  See example in tests/Notebooks/custom_loss.ipynb
       """
   
       def __init__(self, _smooth=1.0e-3, _axis=[2,3], _weight = None, _batch_axis= 0, **kwards):
           Loss.__init__(self,weight=_weight, batch_axis = _batch_axis, **kwards)
   
           with self.name_scope():
               self.smooth = _smooth
               self.axis = _axis
   
   
       def hybrid_forward(self,F,_label, _output):
   
           itrs = F.sum(_label*_output, axis=self.axis)
           l = F.sum(_label*_label, axis=self.axis)
           r = F.sum(_output*_output, axis=self.axis)
   
           IoU = (2.0*itrs+self.smooth)/(l+r+self.smooth)
   
           meanIoU = F.mean(IoU,axis=1)
           meanTot = F.mean(meanIoU)
   
           return meanTot
   
   ```
   the problem is solved. I guess there is an inconsistency between documentation and implementation (there are no default values as described in the docs)? Hope this helps. 
   

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