CCNet implemented by Jittor

Preprocess

Dataset

Please download ADE dataset and make sure the file structures are as follow:

ADE20K
|-images/
|-index_ade20k.mat
|-index_ad220k.pkl
|-objects.txt

You need to run python3 edit_dataset.py to generate data list for ADE20K dataset, after that the file structures will be:

ADE20K
|-images/
|-datalist/
|-index_ade20k.mat
|-index_ad220k.pkl
|-objects.txt

Pretrain

ResNet101 Backbone

Please download MIT imagenet pretrained resnet101-imagenet.pth and store it at pretrain/resnet101-imagenet.pth, then run python3 edit_pretrain_resnet.py to get pretrain/resnet101_base.pth for Jittor reloading.

VAN_b2 Backbone

Please download pytorch version pretrained van_b2.pth and store it at pretrain/van_b2.pth, then run python3 edit_pretrain_van.py to get pretrain/van_b2_base.pth for Jittor reloading.

Run

Training

In order to train xxx model (xxx is either resnet or van), please run the following instruction, and the output log will be stored at log/xxx/yyyymmdd_hhmmss.log

./train_xxx.sh

In case of machine crash or remote disconnection, we recommend that you use tmux to execute training or testing.