Mod-Squad: Designing Mixtures of Experts As Modular Multi-Task Learners

This is a PyTorch/GPU implementation of the paper Mod-Squad: Designing Mixtures of Experts As Modular Multi-Task Learners:

@article{chen2022modsquad,
            title={Mod-Squad: Designing Mixtures of Experts As Modular Multi-Task Learners},
            author={Zitian Chen and Yikang Shen and Mingyu Ding and Zhenfang Chen and Hengshuang Zhao and Erik Learned-Miller and Chuang Gan},
            journal={CVPR},
            year={2023}
}

• The implementation was based on Pytorch==1.10.2 and test on powerpc server.

Prepare

Dataset: Taskonomy

An example of the download from tiny subset

omnitools.download class_object class_scene depth_euclidean depth_zbuffer edge_occlusion edge_texture keypoints2d keypoints3d nonfixated_matches normal points principal_curvature reshading rgb segment_semantic segment_unsup2d segment_unsup25d --components taskonomy --subset tiny --dest ./taskonomy_tiny/   --connections_total 40 --agree --name [your name] --email [your email]

Please put the data in ./data

Default model will save to ./work_dir, logs will be save to ./log_dir

Environment: timm==0.3.2 pytorch==1.10.2

Install MoE module:

cd parallel_linear
pip3 install .

Train

python -m torch.distributed.launch --nnodes=1 --nproc_per_node=2 --master_port 44875 main_mt.py \
        --batch_size 6 \
        --epochs 100 \
        --input_size 224 \
        --blr 4e-4 --weight_decay 0.05 \
        --warmup_epochs 10 \
        --model mtvit_taskgate_att_mlp_base_MI_twice \
        --drop_path 0.1 \
        --scaleup \
        --exp-name scaleup_mtvit_taskgate_att_mlp_base_MI_twice \