Trouble replicating the results

[ykotseruba]

Hi,
I was trying to replicate the results presented in your paper and ran into some problems.
As far as I understand, you more or less follow the training procedure for DeepLab_LargeFOV network outlined here but with 2 classes instead of the original 21 and without the final CRF refinement layer (correct me if this is wrong).
I am training on the 10,582 images from the augmented PASCAL dataset, initialize the weights with VGG trained on ImageNet and use learning parameters as specified in the paper, however the results are not nearly as good. I haven't run the full set of tests yet, but on the example images results produced by my network are far worse than the published pixelobjectness model.
Also suspiciously, the loss fluctuates significantly and training for 10000 iterations takes about 3 hours on NVIDIA Titan X instead of 8 hours stated in the paper.
Below is the solver I've been using, could you please let me know if I'm missing something?

`lr_policy: "step"
gamma: 0.1
stepsize: 2000
base_lr: 0.001

display: 10
max_iter: 10000
momentum: 0.9
weight_decay: 0.0005`