CCC-123
About the Loss function.
In the paper, the author said "... , which the cross-entropy loss L^attr is applied for pedestrain attribute recognition." However, in your code, the loss function is BCEWithLogitsLoss. Could you give the reason about it?
Sorry, I asked a stupid question. The BCEWithLogitsLoss in Pytorch is according with the SigmoidCrossEntropyLoss in Caffe. But I have another question. Why are the weights in your paper those values? How do you figure out the weights?
The weight I calculated is as follows:
[0.63371877 0.98601126 0.39297374 0.9494238 0.69839181 0.75289839
0.69963256 0.95700181 0.83599094 0.83235188 0.82602933 0.86278694
0.99040879 0.55656295 0.66098443 0.95031431 0.85723987 0.89706675
0.95695396 0.99496273 0.98239432 0.97219391 0.49214224 0.84040194
0.88946288 0.9938316 ]
My calculation is as follows:
pa_100k = sio.loadmat(matfile)
positive = np.zeros(26)
for label in pa_100k['train_label']: positive = positive + label
positive = positive / len(train_labels)
weight = np.exp(-positive)
print(weight)