Chenxi Liu
Chenxi Liu
Hmm, I am not sure where it went wrong. I suggest running Check 2 at intermediate layers, to progressively find where the conversion began to fail.
Sorry, no. Since I am no longer at Google, I do not even have access to the search code anymore.
Typically architecture search consists of two stages: the first stage that finds an "optimal" architecture, and the second stage that trains and tests this found "optimal" architecture (on potentially another...
Hi, you can try `python main.py --train --dataset cityscapes --exp lr1e-2_crop769_batch20_lastmult10 --epochs 480 --base_lr 0.01 --crop_size 769 --batch_size 20 --last_mult 10`, which should get >76 mIOU, which is ~1 lower...
That I have not tried.
Sounds like the main change is replacing the first parameter "3" in https://github.com/chenxi116/DeepLabv3.pytorch/blob/master/deeplab.py#L144 and https://github.com/chenxi116/DeepLabv3.pytorch/blob/master/deeplab.py#L148 to "1".
So there are some datasets that use class 255 to label pixels that are ambiguous. These pixels should not count towards IOU calculation. In utils.py I first made sure they...
I see your concern, but I think another good property of the metric would be: 0.0 for the worst algorithm in the world, and 1.0 for the best algorithm in...
The information you have provided is not enough for me to diagnose the problem for you. For example, what's your GPU memory size? How big are the images you are...
This seems to happen when a particular data batch has size 1. Then BN breaks down. Which dataset are you training on? And what is your batch size?