Initialization issues in decoder of deeplabv3plus

[Justin0111]

Hi, thanks for sharing this nice repository. However, I found that you comment the initialization for convolutional weights in deeplabv3plus.py. Do you deliberately comment this part since this setting can result in better performance?

[YudeWang]

Hi @Justin0111, I remember using the this part of initialization leads to NaN in training. It is really strange and I comment it.

[Justin0111]

Thanks for your reply. Did you compare the segmentation results of using this part of initialization? I found that using initialization slightly reduce the performance, which is very confusing and I don't know whether it was caused by NaN as you mentioned. By the way, I found that you choose fixed aspp_rate for both os 8 and 16, which is different from that of the paper. Is it a typo?

[YudeWang]

Hi @Justin0111 , For initialization, I suggest comment it. As for ASPP_rate, it's okay to times 2 on them for os=8. Personally I think the receptive field is large enough and it is not necessary to increase dilation rate in aspp module anymore.

[Justin0111]

Hi @YudeWang, Thank you for your reply, I will have a try.