Binary classification inconsistencies

[aiqc]

• nn.Sigmoid = float
• nn.BCELoss() = float
• torchmetrics.Accuracy() for binary = long

so i had to implement a flip_type function with a try statement while looping in epochs

meanwhile

• torchmetrics.Accuracy() for multi-label = float

https://torchmetrics.readthedocs.io/en/stable/classification/accuracy.html

looking at the Accuracy docs, it seems like comparing equal length tensors is reserved for ordinal ints. seems like a symptom of the nn.Softmax(dim=1) problem