Open ManjeeraJagiri opened 6 years ago
Theoretically is equivalent. But it is recommended to use sigmoid instead of softmax because you will reduce the computation time. Let p1 be the probability of success (class = '1') and p2=1-p1 the probability of class '0'. With sigmoid activation u need to compute only p1 (so u know p2=1-p1), but with softmax u will compute p1 and p2 together (p2 is redundant) . So imo if you want to be computationally efficient, u should use the sigmoid version.
These are the changes I have made: In SSD_training.py, I changed
NUM_CLASSES = 2
to include my class +background In get_data_from_XML.py, I changedself.num_classes = 1
and _to_one_hot function asIs it okay to encode labels, when there is only one class present? I am still training the model(No gpu, it will take time), so I am not sure if it will work. Any help is greatly appreciated!