dual focal loss

What is the equivalent of Adaptive class weight layer in the original article? The output of this layer should be the input of the softmax here: pred = torch.softmax(logits, dim=1) but none treatment were made for this input.

[CoinCheung]

I have just thought of that, if I have not misunderstood what ACW does, it seems that when the logits output of the model is a [N, M] tensor(N samples and each sample has M output values as logits of each class), ACW provides a weight vector of shape [1, M] that should be multiplied into the logits of each sample as the learned remedy for the knowledge learned from the unbalanced dataset. For each row of the [N, M] logits tensor, the weight vector is multiplied to it elementwisely, which I thought should be similar behavior of depthwise convolutions. For example, your model structure might be like this:

class Model(nn.Module):

    def __init__(self):
        self.backbone = some_fcn()
        self.acw = nn.Conv2d(in_chan, out_chan, 1, 1, 0, groups=in_chan)
        self.criteria = dual_focal_loss()

    def forward(self, x):
        feat = self.backbone(x)
        avg_pool = torch.mean(feat, dim=(2,3), keepdim=True)
        logits = self.acw(avg_pool)
        if self.training:
            out = self.criteria(logits)
        else:
            out = avg_pool
        return out

This nn.Conv2d with groups as same as input channels should act similarly as the acw layer.

[CoinCheung]

Hi, I have no idea if you are still interested in implementing acw layer, but I have just thought that maybe you can use a 1x1 depth-wise convolution layer following the logits, which should work in the same way as the acw layer.