the Question for loss source 3

[cw1091293482]

Hi, Thanks for sharing this code. However, it seems that there is no L2 loss of intermediate features in this code. This is my question about loss source 3 in original paper, which is ||F_{i} - F_{C} ||_{2}^{2} in equation (4).

The features F_{i} and F come from different levels of the network, the Width and Height in feature maps might be different. In this scenario, how to perform the L2 loss?

Best

[luanyunteng]

Hi, Thanks for sharing this code. However, it seems that there is no L2 loss of intermediate features in this code. This is my question about loss source 3 in original paper, which is ||F_{i} - F_{C} ||_{2}^{2} in equation (4).

The features F_{i} and F come from different levels of the network, the Width and Height in feature maps might be different. In this scenario, how to perform the L2 loss?

Best

We add a bottleneck block to the early exit, in order to change the feature map to the same shape (4 x 4 x 2048). So we could apply the L2 loss easily. And I will add these code as soon as possible.