Reproducing M0(2)(F(2)) (the released SSL1)

[Turlan]

I followed your training steps by using

python BDL.py --snapshot-dir ./snapshots/gta2city \ --init-weights /path/to/inital_weights \ --num-steps-stop 80000 \ --model DeepLab

I took the initial model DeepLabV2 as initial weights. Source data is the provided translated images GTA5 as CityScapes (DeepLab). Target data is the training set of Cityscapes. Then I only got 43.6 mIou while I should get 44.3.

[liyunsheng13]

I think the result might be normal. The adaptation result has some perturbation. In my experiments, I have results better than or worse than 44.3. I think 43.6 is a little low but still makes sense. You can try another time to see whether it remains the same.

[Turlan]

Thanks, I will try again.

[Turlan]

I tried again, and the result is still not right, I am training for the third time. For the second training, I draw the mIOU curve on cityscape val dataset. I find that the overfitting issue seems to be severe. For this training, the best result is at round 40k, not at the suggested 80k. I wanna know if this normal too. Is it ok that just take 40k's model as the best model for the following SSL training?

[liyunsheng13]

The overfitting problem does exist. But in my experiments, it only happens after 80000 iterations. I list all the results for every 10000 iterations below: 10000,42.77->20000,43.13->30000,45.31->40000,44.51->50000,44.03->60000,44.54->70000,43.96->80000,44.22->90000,42.35->10000,41.69->110000,41.35 I repeat the experiments for several times and find 70000 or 80000 can give a stable performance. I don't suggest you to pick up the best performance. That is not what I did when I wrote the paper. But when you do self-training, you will find there is no overfitting problem.