Some questions of paper and codes

[WangDeyu]

Some questions of paper and codes

[y0ast]

1. The updating mechanism is described in equation 4,5,6 of the paper: https://arxiv.org/abs/2003.02037
2. See page 11 of: https://arxiv.org/pdf/1711.00937.pdf

Is the gradient of x just for calculating gradient penalty? How does the loss of l_gradient_penalty * gp backpropagate?

Yes, that's right. It's added to the loss, which is followed by loss.backward()

Could you please explain the process of model.update_embeddings ? What’s the meaning of self.N and self.m?

model.update_embeddings is called after each model update step. self.N is the number of points of a particular class in a minibatch, self.m is the sum of the embeddings per class. Both are exponentially averaged.

This is explained below equation 6 in the paper (see 1. above).