Result

The result on ADE20K validation set.

The result on COCO-Stuff validation set.

Visualization on ADE20K validation set.

Usage

The supernet training and searching process costs about one week on 8 Nvidia V100 GPUs. So if you want to reproduce the result, make sure you have enough GPU resources.

Please see MMSegmentation for dataset prepare.

For supernet pre-training, run:

cd classification;

then run:

bash run_supernet_fix_backbone.sh;

and

bash run_supernet_fix_trans.sh;

For supernet fine-tuning, run:

cd segmentation;

then run:

bash tools/dist_train_bignas.sh configs/hess_dynamic_supernet/base/hess_4x8_ade20k_fix_backbone.py.py 8 --work-dir /path;

and

bash tools/dist_train_bignas.sh configs/hess_dynamic_supernet/base/hess_4x8_ade20k_fix_trans.py.py 8 --work-dir /path;

To search for the optimal subnet, run:

python tools/search_alternately.py configs/hess_dynamic_supernet/base/hess_4x1_ade20k_search.py --eval mIoU;

To retrain the subnet on ImageNet, run:

cd classification;

then run:

bash retrain_subnet_base.sh

Then run:

cd retrain; sh tools/dist_train.sh local_configs/HESS/base/<config-file> <num-of-gpus-to-use> --work-dir /path/to/save/checkpoint

to fine-tune the subnet on segmentation dataset.

To evaluate, run:

sh tools/dist_test.sh local_configs/HESS/<config-file> <checkpoint-path> <num-of-gpus-to-use>

To test the inference speed in mobile device, please refer to tnn_runtime.

Acknowledgement

The implementation is based on MMSegmentation.

The checkpoint can be found in Google Drive