search

nxsEdson / MLCR

Multi-label Co-regularization for Semi-supervised Facial Action Unit Recognition (NeurIPS 2019)
63 stars 18 forks source link

A question for your codes? #1

Open Keysmis opened 4 years ago

Keysmis commented 4 years ago

I'm wondering what the three values mean? Is_begin_weight,begin_weight1,begin_weight2 thx~

nxsEdson commented 4 years ago

We need the pre-trained weights as the initialization of GCN. These values are used for controlling the initialization.

Keysmis commented 4 years ago

and what process method for the label is 999?

Keysmis commented 4 years ago

and the variable self.weight is how to get it? is it get from the annotated labels by computing the overall positive/negative of each AU?

nxsEdson commented 4 years ago

For unlabeled data, I just ignore them and do not back pass the loss. For self.weight, do you mean the adjacent map for GCN? I calculate the map based on the whole database.

Keysmis commented 4 years ago

@nxsEdson Thanks for your great work!, yeah, i already compute that weight in whole dataset, but i have another question......， i couldn't found the variable of self.select_sample in computing the JS dis... ， what does that mean? and the weights of BCE_sigmoid_negtive_bias_all loss function, is it compute the pos/neg ratio in whole dataset?