Add probability to Accuracy #1354
Conversation
vcarpani
changed the title
ignite/metrics/accuracy.py
Outdated
| def __init__( | ||
| self, | ||
| output_transform: Callable = lambda x: x, | ||
| is_multilabel: bool = False, | ||
| device: Union[str, torch.device] = torch.device("cpu"), | ||
| mode: Mode = Mode.LABELS, |
There was a problem hiding this comment.
I'm not sure about correctness of name LABELS for multi-class case where we require to pass probas or logits and then take argmax.
There was a problem hiding this comment.
What about UNCHANGED or RAW_INPUT?
There was a problem hiding this comment.
Yes, something like RAW_INPUT could work. I'm thinking if we could not generalize this new options to all possible inputs and metric type: binary, multiclass, multilabel...
|
@vfdev-5 the last commit should solve your comments |
|
I'll try to implement that behaviour somewhere next week ;) |
@vfdev-5, just wanted to update you on this one, I've been quite busy recently, but I started working on it |
|
@vcarpani thanks for the head-up :) |
|
I would like to add that "logits" in the context of multiclass problems usually mean the input to the softmax (not the sigmoid) function. I am not sure if there are any use cases for the current |
Fixes #1089
Description:
Accuracyaccording to the suggestions from the issue.Still WIP, to be done:
Check list: