Implementation of Consistency Loss

[michaelku1]

Hello, so I noticed that there is a slight difference between the paper's fornulation of consistency loss and the actual implementation. Please correct me if I am wrong. As can be seen, the actual formulation are as follows:

However, I noticed that for the final consistency loss, the actual implementation uses a sum instead of an average across layers as indicated in equation 12 (weight_dict for this type of loss is 1 by defualt, therefore notion of average is perhaps not incorporated). As for the per layer consistency loss, the actual implementation uses an average over all M object queries instead of just the sum, as indicated in equation 13.

I am wondering what are the right formulations to follow? Thanks.

[encounter1997]

Hi, as shown in Eq. 12, the loss is divided by the number of decoder layers, which equals len(pred_boxes_all).

As for the summation in Eq. 13, we are sorry for the mistake, it should be an average over M predictions in an individual decoder layer.

[encounter1997]

Thanks for pointing it out, we will correct it and upload a new file to Arxiv.

[michaelku1]

Thanks for your reply. For equation 12, it shows that it is normalised over number of decoder layers, but I think len(pred_boxes_all) gives the number of object queries (batch_size*300), which is suggested here:

[encounter1997]

Each element in the list pred_boxes_all has a shape of (B*300, 4). Afterward, they are stacked into a tensor with a shape of (num_layers, B*300, 4), as shown in this line.

[michaelku1]

I think you are right. Sorry I have not run the code yet myself so was just reading along. Thanks.