Suddenly lists about skorch HOT 5 CLOSED

That would be the safest solution. We could also patch pytorch arrays to have a shape attribute and would be fine ;)

from skorch.

I have investigated the reason and it is caused by sklearn's safe_indexing:

import torch
from sklearn.utils import safe_indexing
xx = torch.zeros((10, 3))
safe_indexing(xx, [0, 1])

# returns:
# [
#   0
#   0
#   0
#  [torch.FloatTensor of size 3], 
#   0
#   0
#   0
#  [torch.FloatTensor of size 3]]

To sum up, safe_indexing is often used to index data in sklearn. But this function only works well with pandas DataFrames (via iloc) or numpy arrays (via shape). If the array is neither, sklearn defaults to [X[idx] for idx in indices], which works for lists but not for torch arrays.

If we want to solve this without changing sklearn or pytorch (if only pytorch tensors had a shape attribute intsead of size ...), I would suggest that we require the data to be in the numpy world if the user wants to benefit from sklearn goodies.

from skorch.

So in essence everywhere we communicate with sklearn we have to make sure that the data is formatted as numpy arrays?

from skorch.

Update: Now torch tensors have a shape attribute, but it still doesn't work (because of pytorch/pytorch#2305).

>>> from sklearn.utils import safe_indexing
>>> import torch
>>> xx = torch.zeros((10, 3))
>>> safe_indexing(xx, [0, 1])
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/bbossan_dev/anaconda3/envs/inferno/lib/python3.6/site-packages/sklearn/utils/__init__.py", line 112, in safe_indexing
    return X[indices]
TypeError: indexing a tensor with an object of type list. The only supported types are integers, slices, numpy scalars and torch.LongTensor or torch.ByteTensor as the only argument.

However, this works:

>>> safe_indexing(xx, ([0, 1],))
 0  0  0
 0  0  0
[torch.FloatTensor of size 2x3]

A workaround could thus be to always wrap list indices in a tuple, since it works for pytorch and doesn't interfere with numpy.

from skorch.

The issue regarding GridSearchCV is solved with this solution. Therefore I'm closing this.

from skorch.