torch.nn.functional.binary_cross_entropy and torch.nn.functional.binary_cross_entropy_with_logits documentation wrong on description about target

[cheyennee]

📚 The doc issue

The binary_cross_entropy documentation shows that target – Tensor of the same shape as input with values between 0 and 1. However, the value of target does not necessarily have to be between 0-1, but the value of input must be between 0-1.

import torch
arg_1_tensor = torch.randint(2, 3, [5, 5], dtype=torch.float32)
arg_2_tensor = torch.randint(2, 3, [5, 5], dtype=torch.float32)
arg_3 = None
arg_4 = "mean"
res = torch.nn.functional.binary_cross_entropy(input=arg_1_tensor,target=arg_2_tensor,weight=arg_3,reduction=arg_4,)
print(res)
# res: RuntimeError: all elements of input should be between 0 and 1

The binary_cross_entropy_with_logits documentation shows that target – Tensor of the same shape as input with values between 0 and 1. However, the value of target does not necessarily have to be between 0-1.

import torch
arg_1_tensor = torch.randint(2, 4, [10, 64], dtype=torch.float32)
arg_2_tensor = torch.randint(2, 4, [10, 64], dtype=torch.float32)
arg_3 = None
arg_4_tensor = torch.rand([64], dtype=torch.float32)
arg_4 = arg_4_tensor.clone()
arg_5 = "mean"
res = torch.nn.functional.binary_cross_entropy_with_logits(input=arg_1_tensor,target=arg_2_tensor,weight=arg_3,pos_weight=arg_4,reduction=arg_5,)
print(res)
# res: tensor(-3.8518)

Suggest a potential alternative/fix

No response

cc @svekars @carljparker @albanD @mruberry @jbschlosser @walterddr @mikaylagawarecki

[MylesJP]

Mind if I take a shot at this one for a school project?

[jbschlosser]

IIRC we've tried to introduce a [0, 1] restriction for target in the past but ran into BC issues from people using values outside the range. I'm not sure we want to encourage this by changing the docs, however.

[jbschlosser]

I take it back; it looks like this was addressed and actually landed in #97814. Closing as addressed but feel free to reopen if there's more I missed.

[cheyennee]

Mind if I take a shot at this one for a school project?

Sure!