search

jfzhang95 / pytorch-deeplab-xception

DeepLab v3+ model in PyTorch. Support different backbones.
MIT License
2.91k stars 781 forks source link

question about the training details #108

Open SaberSong opened 5 years ago

SaberSong commented 5 years ago

I have tried to train model with the scripts "train_voc.sh". however the best mIoU only achieved 73%, and it overfitting suddenly after 15 epoch, which will not happen without using SBD(no overfitting, best mIoU about 74%). I noticed that the deeplab-resnet model you provided is acquired at 58 epoch, so the default hyper-parameters can not get the best performance. I want to know the training details to get the best models, thank you! And there are another question about the validation, I noticed that you using the center crop for the val image, which may ignore the targets near the boundary. I wander whether its the same way as the author of deeplab done. If not, how the deeplab author deal with the val. I have tried the original image or the resize the long side to 513, both made the mIoU decreased over 2 percents. Looking forward to your reply, thank you.@jfzhang95

mcever commented 5 years ago

I have this same question. With default hyperparameters, the best I am seeing is 75 mIOU. Further, in the paper they achieve over 80% mIOU on VOC. Does your code allow for that? Do you support multi-scale and flipping out of the box?

Thanks for the code!

GuoleiSun commented 5 years ago

Same problem. I also got ~75 mIOU using "train_voc.sh".

hanfeng0409 commented 4 years ago

Same problem, I got 75.98 mIOU in VOC without other data. In the Paper they use additional data to train, this is link to datasets http://home.bharathh.info/pubs/codes/SBD/download.html.

kungkook commented 4 years ago

Can I know how the parameters in train.py are set when you train the voc data set? I want to train VOC2007 with xception as the main dry network, but it is always wrong :train.py:error: unrecognized arguments:Ture