Hi there, I've been reading the paper and checking your code, and I tried to apply it to another custom dataset for OnTAL.
I have a few questions regarding the code:
in the constructor of TRNTHUMOSDataLayer, you have
So, you create tmp_data and concatenate it to video_feature, so that video_feature.shape[0] is a multiple of self.stride. However, immediately below you're doing the same for target:
but at this point, len(video_feature) % self.stride = 0, because you just concatenated it with tmp_data. Shouldn't you save the original len of video_feature, and use that one instead?
I don't understand why you consider the entry a TP only by checking if the corresponding GT value is equal to 1, it's like you're expecting all the elements in this_cls_prob to be positives.
I thought this_cls_prob was an array of probabilities that that specific snippet belongs to a class, so that you could have something like (assume using a threshold of 0.7):
Hi there, I've been reading the paper and checking your code, and I tried to apply it to another custom dataset for OnTAL. I have a few questions regarding the code:
So, you create tmp_data and concatenate it to video_feature, so that video_feature.shape[0] is a multiple of self.stride. However, immediately below you're doing the same for target:
but at this point, len(video_feature) % self.stride = 0, because you just concatenated it with tmp_data. Shouldn't you save the original len of video_feature, and use that one instead?
can't you just do:
can't you just do:
I don't understand why you consider the entry a TP only by checking if the corresponding GT value is equal to 1, it's like you're expecting all the elements in this_cls_prob to be positives.
I thought this_cls_prob was an array of probabilities that that specific snippet belongs to a class, so that you could have something like (assume using a threshold of 0.7):
And in this case, I would expect you should use something like:
and so on, but maybe I just misinterpreted the meaning of the arrays?