Fix CrossEntropyLoss Parameters Deprecation in PyTorch

While training models in PyTorch, you might encounter deprecation warnings related to the CrossEntropyLoss configuration. These warnings appear when older parameters such as size_average or reduce are used. PyTorch is gradually deprecating these parameters in favor of a single, clearer argument reduction.

The deprecation exists to simplify the API and eliminate confusion caused by combining multiple flags that control the same behavior. Instead of toggling size_average and reduce, PyTorch now expects you to explicitly define how the loss should be aggregated.

Here is an example that triggers the deprecation warnings:

import torch

x = torch.tensor([[1.5, 2.5, 3.5], [2.0, 1.0, 0.5]])
target = torch.tensor([1, 0])

loss = torch.nn.CrossEntropyLoss(size_average=True)
y = loss(x, target)
print(y) # tensor(0.9360)

loss = torch.nn.CrossEntropyLoss(size_average=False)
y = loss(x, target)
print(y) # tensor(1.8720)

loss = torch.nn.CrossEntropyLoss(reduce=False)
y = loss(x, target)
print(y) # tensor([1.4076, 0.4644])

Output:

UserWarning: size_average and reduce args will be deprecated, please use reduction='mean' instead.
UserWarning: size_average and reduce args will be deprecated, please use reduction='sum' instead.
UserWarning: size_average and reduce args will be deprecated, please use reduction='none' instead.

To fix deprecations, use the reduction parameter directly when creating the loss function. This is the modern and forward-compatible solution:

import torch

x = torch.tensor([[1.5, 2.5, 3.5], [2.0, 1.0, 0.5]])
target = torch.tensor([1, 0])

loss = torch.nn.CrossEntropyLoss(reduction='mean')
y = loss(x, target)
print(y) # tensor(0.9360)

loss = torch.nn.CrossEntropyLoss(reduction='sum')
y = loss(x, target)
print(y) # tensor(1.8720)

loss = torch.nn.CrossEntropyLoss(reduction='none')
y = loss(x, target)
print(y) # tensor([1.4076, 0.4644])