[BUG] Interoperability of cuML (confusion matrix) with torch arrays

[divyegala]

Splitting from #3567. Bug has got to do with interoperability of cuml's confusion matrix with torch arrays. However, testing for this is likely to go beyond just confusion matrix.

Adding to this bug:

Code:

from cuml.metrics import confusion_matrix
print(confusion_matrix(y_true=labels.type(dtype=torch.int32).to(device),
                                   y_pred=pred.argmax(1).type(dtype=torch.int32).to(device)
                                   ))
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-23-b32704b34c7b> in <module>
     20 
     21     print(confusion_matrix(y_true=labels.type(dtype=torch.int32).to(device),
---> 22                            y_pred=pred.argmax(1).type(dtype=torch.int32).to(device)))


/opt/conda/envs/rapids/lib/python3.7/site-packages/cuml/internals/api_decorators.py in inner_with_getters(*args, **kwargs)
    464 
    465                 # Call the function
--> 466                 ret_val = func(*args, **kwargs)
    467 
    468             return cm.process_return(ret_val)

/opt/conda/envs/rapids/lib/python3.7/site-packages/cuml/metrics/confusion_matrix.py in confusion_matrix(y_true, y_pred, labels, sample_weight, normalize)
     67 
     68     if labels is None:
---> 69         labels = sorted_unique_labels(y_true, y_pred)
     70         n_labels = len(labels)
     71     else:

/opt/conda/envs/rapids/lib/python3.7/site-packages/cuml/metrics/utils.py in sorted_unique_labels(*ys)
     22     labels."""
     23     ys = (cp.unique(y) for y in ys)
---> 24     labels = cp.unique(cp.concatenate(ys))
     25     return labels

/opt/conda/envs/rapids/lib/python3.7/site-packages/cupy/manipulation/join.py in concatenate(tup, axis, out)
     55         tup = [m.ravel() for m in tup]
     56         axis = 0
---> 57     return core.concatenate_method(tup, axis, out)
     58 
     59 

cupy/core/_routines_manipulation.pyx in cupy.core._routines_manipulation.concatenate_method()

cupy/core/_routines_manipulation.pyx in cupy.core._routines_manipulation.concatenate_method()

/opt/conda/envs/rapids/lib/python3.7/site-packages/cuml/metrics/utils.py in <genexpr>(.0)
     21     """Extract an ordered array of unique labels from one or more arrays of
     22     labels."""
---> 23     ys = (cp.unique(y) for y in ys)
     24     labels = cp.unique(cp.concatenate(ys))
     25     return labels

/opt/conda/envs/rapids/lib/python3.7/site-packages/cupy/manipulation/add_remove.py in unique(ar, return_index, return_inverse, return_counts, axis)
    110         aux = ar[perm]
    111     else:
--> 112         ar.sort()
    113         aux = ar
    114     mask = cupy.empty(aux.shape, dtype=cupy.bool_)

cupy/core/core.pyx in cupy.core.core.ndarray.sort()

cupy/core/core.pyx in cupy.core.core.ndarray.sort()

cupy/core/_routines_sorting.pyx in cupy.core._routines_sorting._ndarray_sort()

cupy/cuda/thrust.pyx in cupy.cuda.thrust.sort()

RuntimeError: radix_sort: failed on 1st step: cudaErrorInvalidDevice: invalid device ordinal
Code that does work by converting torch.tensor with .cpu():

from sklearn.metrics import confusion_matrix
 
 print(confusion_matrix(y_true=labels.type(dtype=torch.int32).cpu(),
                                    y_pred=pred.argmax(1).type(dtype=torch.int32).cpu()
                                    ))
                                    
[[1      2]
 [  3      4]]

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[github-actions]

This issue has been labeled inactive-90d due to no recent activity in the past 90 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed.