Implemented binary_cross_entropy_with_logits in paddle.nn.functional.loss

[KevinUli]

Close #16741

[KevinUli]

Result from paddle ground truth and other backend differs, so I set test_values=False

[xoiga123]

@KevinUli How do they differ? If you explicitly set test_values=False then you have to implement your own testing from the returned values.

[KevinUli]

@xoiga123 When running the function for torch and paddle,

>>> arr = torch.as_tensor([0.5, 0.5,], dtype=torch.float32)
>>> pos_weight = torch.as_tensor([1., 2.,], dtype=torch.float32)
>>> torch.nn.functional.binary_cross_entropy_with_logits(arr, arr, weight=arr, reduction="none", pos_weight=pos_weight)
Tensor([0.3620, 0.4806])

>>> arr = paddle.to_tensor([0.5, 0.5,], dtype="float32")
>>> pos_weight = paddle.to_tensor([1., 2.,], dtype="float32")
>>> paddle.nn.functional.binary_cross_entropy_with_logits(arr, arr, weight=arr, reduction="none", pos_weight=pos_weight)
Tensor(shape=[2], dtype=float32, place=Place(cpu), stop_gradient=True, [0.36203849, 0.54305774])

Getting rid of the pos_weight kwarg returns the same value for all frameworks
The pos_weight kwarg seems to work differently for paddle as opposed to other frameworks.
Should I simply remove pos_weight kwarg from the testing and adding a comment for it?

[xoiga123]

@KevinUli Yeah seems weird, I'll look into source and see what I can find. Thank you

[xoiga123]

@KevinUli Yeah clearly the pytorch's implementation of pos_weight is correct, while the paddle's implementation is incorrect. This seems easily fixable. Please add a comment below the test definition

# TODO: paddle's implementation of pos_weight is wrong
# https://github.com/PaddlePaddle/Paddle/blob/f0422a28d75f9345fa3b801c01cd0284b3b44be3/python/paddle/nn/functional/loss.py#L831

And then proceed to open an issue on their side, or open a PR to fix it if you want to. I'll merge this PR but please keep track of when it's fixed and add pos_weight back. Thank you 🤗

[KevinUli]

I've opened an issue on their side.

[xoiga123]

Well done 👍