search

donnyyou / torchcv

TorchCV: A PyTorch-Based Framework for Deep Learning in Computer Vision
https://pytorchcv.com
Apache License 2.0
2.25k stars 378 forks source link

mIoU of SFNet is much lower than 78.9% #162

Closed perrying closed 4 years ago

perrying commented 4 years ago

I trained SFNet with default setting of torchcv/scripts/seg/cityscapes/run_sfnet_res18_cityscapes.sh except for NGPUS (I changed NGPUS from 8 to 4, and --train_batch_size from 2 to 4). But I got about 73% of mIoU for single scale inference. It may be because pretrained model, 3x3resnet18-imagenet.pth, is not given. Although I also trained SFNet (ResNet-101) using pretrained model, 3x3resnet101-imagenet.pth, given by this repository, I got about 78% of mIoU for multi-scale inference.

How can I reproduce the paper results?

seekFire commented 4 years ago

@perrying Hi, Have you achieved the result shown in the paper yet?

perrying commented 4 years ago

I haven't reproduced it yet.

seekFire commented 4 years ago

@perrying Thank you for your reply! I was interested in this work and planned to re-implement it, but now it seems that my plan should be postponed...

lxtGH commented 4 years ago

@seekFire @perrying We will release the trained model. Also, we will release the better models in another repo. https://github.com/lxtGH/SFSegNets

seekFire commented 4 years ago

@lxtGH Thank you for your response, we are looking forward to it!

lxtGH commented 4 years ago

@donnyyou Please give the trained ckpt for reproducing results.

donnyyou commented 4 years ago

You could email your training logs to "youansheng@pku.edu.cn", and I will check them for you. @perrying @seekFire

donnyyou commented 4 years ago

The logs you email me seems right for res101, but wrong for res18. The pretrained res18 model is missing for initialing the weights. You could download the deepbase res18 pretrained model to fix this.

donnyyou commented 4 years ago

@perrying

donnyyou commented 4 years ago

https://github.com/CSAILVision/semantic-segmentation-pytorch/blob/master/mit_semseg/models/resnet.py The pretrained imagenet models could be downloaded here! Note that the resnet version is different from the torchvision version by replacing 7x7 convolution with three 3x3 convolutions.

donnyyou commented 4 years ago

Not using deeply supervised losses, that res18-based model uses, might get better results for training the sf models with deeper resnets, such as res50, res101, etc..

donnyyou commented 4 years ago

More questions please email me at youansheng@pku.edu.cn.

donnyyou commented 4 years ago

You should also change the loss type to fpndsnohemce_loss series for training res18-based sf model. @perrying

seekFire commented 4 years ago

@donnyyou So you mean when training the sf model with deeper resnets, we should not use fpndsnohemce_loss series as its loss, isn't it?

perrying commented 4 years ago

@donnyyou Thank you for the clarification!

donnyyou commented 4 years ago

@donnyyou So you mean when training the sf model with deeper resnets, we should not use fpndsnohemce_loss series as its loss, isn't it?

You are right, and dsnce_loss might get better results.

seekFire commented 4 years ago

@perrying OK, thanks for your answer! BTW, if I select other backbone, such as HRNet, MobileNet and so on, which loss would you recommend?

donnyyou commented 4 years ago

deeply supervised losses for smaller models. @seekFire

seekFire commented 4 years ago

@donnyyou OK! I see. Thank you very much!