zrl4836 / acfnet Goto Github PK

Line 59&60 in loss/criterion.py
pred = preds[0] + preds[1]
loss4 = lovasz_softmax(F.softmax(pred, dim=1), target, ignore=self.ignore_index)

这里的 pred 没有做upsample, 然后直接给了lovasz_softmax, 而 target 是原图的size大小，这样会造成 pred 与 target 的size 不匹配。

Hi,
Thanks for your contribution! I wanna know what's the final mIoU of your training code? Is it the same as the paper?
Meanwhile, I notice that the learning rate you set is 1e-2 while it is 1e-3 in the original paper. Have you tried 1e-3?
Thank you again! All the best!

hi,good work! I only have two 2080ti cards. Is it enough to run the code ?

你好，请问你是用几张gpu跑出的效果啊，

# multiply & normalize ->(B, N, C')
energy = torch.bmm(proj_query, proj_key)
energy_new = torch.max(energy, -1, keepdim=True)[0].expand_as(energy) - energy

I think the normalization implemented in acfnet.py is different from the equation (2) in the original paper.
And I wonder why you implement this way.

As mentioned in title. I didn't see this operation in paper. Could you show me some guidance?

when i use acf module in unet, the train loss will be nan