Have you trained and evaluated the deeplabv2?

[Hutai]

I use your python scripts to evaluate deeplabv2 model, and I found that the segmenter.py need to be update, e.g. "self.set_input_arrays(img, label, data_dim)" should be converted to "self.set_input_arrays(img, label)" due to the updated caffe compared to the caffe of v1. But, when I used the updated scripts to evaluate one image, its prediction result is so bad, I don't know where I was wrong and can you give me some tips or train-deeplabv2 scripts?