-
Our code is based on MMSegmentation. Please install all the required packages and prepare datasets (at least Cityscapes) following their docs.
-
We have changed the following code files on the basis of MMSegmentation.
/mmsegmentation-master/mmseg/models/backbones/resnet.py
/mmsegmentation-master/mmseg/segmentors/encoder_decoder.py
/mmsegmentation-master/mmseg/apis/train.py
- We also add some customized configs in
/mmsegmentation-master/mmseg/configs
- If you hope to implement InfoPro for other models or datasets, please adapt all aforementioned files and configs carefully (and also follow the guidelines of changing models & datasets in MMSegmentation).
- Train DeepLabV3 on Cityscapes with 512x1024 crop sizes, batch size = 8 (4 GPUs)
CUDA_VISIBLE_DEVICES=0,1,2,3 PORT=29500 bash ./tools/dist_train.sh ./configs/deeplabv3/infopro_deeplabv3_r101-d8_512x1024_40k_cityscapes.py 4
- Train DeepLabV3 on Cityscapes with 512x1024 crop sizes, batch size = 12 (4 GPUs)
CUDA_VISIBLE_DEVICES=0,1,2,3 PORT=29500 bash ./tools/dist_train.sh ./configs/deeplabv3/infopro_deeplabv3_r101-d8_512x1024_bs12_40k_cityscapes.py 4
- Train DeepLabV3 on Cityscapes with 640x1280 crop sizes, batch size = 8 (4 GPUs)
CUDA_VISIBLE_DEVICES=0,1,2,3 PORT=29500 bash ./tools/dist_train.sh ./configs/deeplabv3/infopro_deeplabv3_r101-d8_640x1280_40k_cityscapes.py 4
- Evaluate pre-trained models (4 GPUs)
CUDA_VISIBLE_DEVICES=0,1,2,3 bash ./tools/dist_test.sh ./configs/deeplabv3/deeplabv3_r101-d8_512x1024_40k_cityscapes.py PATH_TO_PRE_TRAINED_MODELS 4 --eval mIoU
- Measured by mean Intersection over Union (mIoU).
| Model | Single Scale (SS) | Multi Scale (MS) | MS + Flip | Link |
|---|---|---|---|---|
| E2E, 40k, bs=8, 512x1024 | 79.12 | 79.81 | 80.02 | Tsinghua Cloud / Google Drive |
| E2E, 60k, bs=8, 512x1024 | 79.32 | 79.95 | 80.07 | Tsinghua Cloud / Google Drive |
| InfoPro* (K=2), 40k, bs=8, 512x1024 | 79.37 | 80.53 | 80.54 | Tsinghua Cloud / Google Drive |
| InfoPro* (K=2), 40k, bs=12, 512x1024 | 79.99 | 81.09 | 81.20 | Tsinghua Cloud / Google Drive |
| InfoPro* (K=2), 40k, bs=8, 640x1280 | 80.25 | 81.33 | 81.42 | Tsinghua Cloud / Google Drive |
