jkvt2 / hierloss Goto Github PK

Hi, thanks for the loss implementation.
I am bit confused here. I know that labels are one hot encoded before we calculate the loss, I would also like to know if the logits are also in the form of one-hot encoding?

PS: My final class prediction is equal to the number of classes, and I can use cross_entropy to calculate a loss in the normal case. If I use this hierarchical loss how are the predictions expected for the loss calculation? Is it a single value or a len(num_classes)?

Hi @jkvt2

Sorry for bothering you again.
I know that raw_probs contains the probability for each class. What does probs contain? I don't quite understand what exactly is done here?

@jkvt2 Thanks for the loss functionality.

Could you please let me know which yolo9000 implementation did you follow? I don't find any yolo9k implementation. Thanks.

Hi @jkvt2
I am trying to have this loss function in pytorch but somehow I am stuck at this point and can't figure out what's the issue.

I get this error: RuntimeError: Index tensor must have the same number of dimensions as input tensor in below line. Do you have any idea about how this works in pytorch?

Thanks

@jkvt2
I am here again with one more doubt.
While testing, why do you just pass the children? Don't we need predictions for parent objects as well? or it's handled in the code which I don't understand?

Great jobs！
Which yolo9000 implementation to work/test with？