模組小隊

Mod-Squad：將專家混合設計為模組化多工學習器

這是 Mod-Squad: Designing Mixtures of Experts As Modular Multi-Task Learners 論文的 PyTorch/GPU 實作：

 @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}
}

• 實作基於Pytorch==1.10.2並在powerpc伺服器上測試。

準備

資料集：Taskonomy

從微小子集下載的範例

 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]

請將資料放入./data

預設模型將保存到./work_dir，日誌將保存到./log_dir

環境： timm==0.3.2 pytorch==1.10.2

安裝教育部模組：

 cd parallel_linear
pip3 install .

火車

 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