speedinghzl
No code implementation about discriminative loss ?
Section 3.4 in your TPAMI version paper, Learning Category Consistent Features Is there any code for this part ?
I was also wondering this, Is it possible for providing the implementation of this discriminative loss?
@LiZhenLiangLee @wqhIris Here's my guess. The original paper published in TPAMI seems not strict enough to illustrate the implement detail. In fact, from this code, i find this model is a encoder-only model rather than auto-encoder like UNet. The upsample is simple bilinear interpolation, and the upsampled feature is the final segmentation result. Therefore, this main problem to implement the loss function proposed in Learning Category Consistent Features section is there's no feature embedding to "pull" and "push"(i.e. the loss component $L_var$ and $L_dis$). Here's what i have tried. I tried to re-implement this metric loss by adding bilinear interpolation followed by a 1x1 conv to the ccnet, while also code the loss function to optimize the feature(embedding) output from 1x1 conv. In fact, i found this loss is not easy to implement, and the training loss of my task seems a little unstable(but actually in convergence). I don't know whether the author of this repo also noticed this problem.
