Bhadresh Dhanani

@qizhuli That was a good point. I tried and tested various learning rate. They all stuck at same loss value ```87.3365``` ``` **Learning rate: 0.000001** I1009 13:08:37.736109 86511 solver.cpp:280] Solving...

@qizhuli Sure. Here is my training proto. [train_val.txt](https://github.com/hszhao/PSPNet/files/1369691/train_val.txt) I did not change anything in BN layers. Could you guide me if there is anything I need to change.

@qizhuli Good to know the specifics. I will try to update the training prototxt according to your suggestions. Will post an update soon.

@gy12346123 Yeah I fine tuned the model using VOC 2012 pre-trained weights and the loss is bit stable now but now per class accuracy is always 0 due to class...

@zanbri Did you change the BN layer code or just the softmax. Since the BN logic for both this paper is different, its kinda hard to put just the modified...

@TheinAnh How did you solve the problem?

@ThienAnh I meant how did you tackle the loss issue?

@xhsoldier Would you mind providing train_val.prototxt? Or any script you used while training and testing?

@xhsoldier So I took one of your training model (pspnet101_VOC2012_train.prototxt) and started training on my dataset which are of size 473X473 and total classes 3. I am getting this weird...

@xhsoldier It worked. Thank you for showing me the right direction. I am pretty much following your caffe-segmentation code. Its very well organized and easy to understand. Great work. One...