Reproducibility on VOC12

[arnike]

Hi Vladimir,

I was wondering about the training specifics to reproduce 80.1% IoU on validation, which you report. My initial model is deeplab_resnet_init.ckpt, which I reckon is a MS-COCO pre-trained model, but otherwise hasn't seen VOC. With train.py I get 82.6% / 74.8% on training / validation. train_msc.py yields 83.4% / 75.9%. I don't change anything set by default (i.e. 20000 iterations, learning rate, etc.). Is there any detail I missed? Another question, did you also observe overfitting in your experiments?

Great job and thanks, Nikita

[DrSleep]

it might be that the TF version is different. Otherwise, if you carefully followed README, you should have reached the numbers mentioned