hqhz1817

Why do you change d.trainable twice after d.compile during train?

1

It seems you can only change one model's trainable before model.compile, if you compile one model, your change of trainable will not work. So, why do you change model.trainable twice...

I think alpha should be min(negative CAMS). CAM = Conv(weight * Feature), Feature must be >= 0, but weight maybe < 0, so min(CAM) maybe < 0.

https://github.com/dontLoveBugs/Deformable_ConvNet_pytorch/blob/aea497d1a5a6090940c8f24a47b97ea09073ae98/network/deform_conv/deform_conv.py#L50 This code just change the order from [x1, y1, x2, y2...] to [x1, x2,...y1,y2...]， not from [x1, x2,...y1,y2...] to [x1, y1, x2, y2...].

https://github.com/zbwxp/SegVit/blob/85616388cd8983e2d4daafd971540b210e956666/decode_heads/atm_single_head.py#L277 semseg = mask_cls.softmax(-1) @ mask_pred.sigmoid() which will result in: **1>= semseg >= 0** When we compute [loss_func](https://github.com/zbwxp/SegVit/blob/85616388cd8983e2d4daafd971540b210e956666/losses/criterion.py#L58), we set semseg as a **logits**, and compute prob by semseg.sigmoid() If...