For classification of an imbalanced image dataset I tried to apply weights and use the CrossEntropy loss function:
_, class_counts = np.unique(data.y.items, return_counts=True)
weights = np.sum(class_counts)/class_counts
weights = tensor(weights).float().cuda()
learn.loss_func = CrossEntropyFlat(weight=weights)
After the above, the learn class can be used to fit like normally. The problem is running the
ClassificationInterpretation. This throws an error.
RuntimeError: Expected object of backend CPU but got backend CUDA for argument #3 'weight'
It seems the ClassificationInterpretation can’t handle the weights passed into the CrossEntropy loss function???
This is a pity because I’d like to see the change in the confusion matrix after applying weights.