This repository is for the IROS 2021 paper ADD: A Fine-grained Dynamic Inference Architecture for Semantic Image Segmentation.
Dynamic-Auto-DeepLab performs three-stage training by firstly searching for the architecture. Second, train the model with the searched network architecture. Third, train earlier-exit-decision.
Modifiy path to your Cityscapes in mypath.py
We search for the architecture on Cityscapes
cd scripts
bash search_cityscapes.sh
The searched architecture and searching progress can be seen by:
tensorboard --logdir path-to-your-exp
One can choose network to train by modified .sh file. Note that we the batch size is #GPU/16 since we use torch.distributed
bash train_dist.sh
bash train_edm.sh
bash eval.sh
-
Pytorch version 1.0+
-
Python 3
-
tensorboardX
-
pycocotools
-
tqdm
-
apex