FANzhaoxin666 / MOCO-Subtomograms

Subtomograms classification with unsupevised learning

Generate Validation Datasets

There are 10 classes with 500 samples (subtomogram_mrc & json_label) per class in ./Datasets/train

run python generate_val.py for random sampling 50 samples ((subtomogram_mrc & json_label)) per class into ./Datasets/val

Train Moco (main_moco.py)

# RB3D
python main_moco.py -a RB3D --lr 0.003 --schedule 300 400 500 --batch-size 32 --moco-k 128 --moco-dim 128 --workers 4 --epochs 600 --dist-url "tcp://localhost:10001" --multiprocessing-distributed --world-size 1 --rank 0 --gpu 2 --aug-plus ./Datasets/ 2>&1 | tee ./logs/arch-RB3D_epochs600_bs32_lr0.003_moco-k128_moco-dim128.log

# DSRF3D_v2
python main_moco.py -a DSRF3D_v2 --lr 0.03 --schedule 300 400 500 --batch-size 32 --moco-k 128 --moco-dim 128 --workers 4 --epochs 600 --dist-url "tcp://localhost:10001" --multiprocessing-distributed --world-size 1 --rank 0 --gpu 0 --aug-plus ./Datasets/ 2>&1 | tee ./logs/arch-DSRF3D_v2_epochs600_bs32_lr0.03_moco-k128_moco-dim128.log

Visualizaiton (main_cluster.py)

The visualization results and the extracted numpy features will be saved in ./Figures.

Notes: As long as the extracted numpy features exist in the ./Figures path, the features will not be re-extracted

• Visualizing the train datasets:

  python main_cluster.py -a DSRF3D_v2 --batch-size 32 --workers 8 --moco-dim 128 --pretrained ./main_moco_checkpoint/Nonorm-DSRF3D_v2_epochs120_bs32_lr0.3_moco-k128_moco-dim128.pth.tar --dist-url "tcp://localhost:10001" --multiprocessing-distributed --world-size 1 --rank 0 --gpu 0 ./Datasets/

• Visualizing the val datasets:

  python main_cluster.py -a DSRF3D_v2 --batch-size 32 --workers 8 --moco-dim 128 --pretrained ./main_moco_checkpoint/Nonorm-DSRF3D_v2_epochs120_bs32_lr0.3_moco-k128_moco-dim128.pth.tar --dist-url "tcp://localhost:10001" --multiprocessing-distributed --world-size 1 --rank 0 --gpu 0 --evaluate ./Datasets/

Train a linear classifier (main_lincls.py)

python main_lincls.py -a RB3D --lr 30 --schedule 150 --batch-size 32 --workers 8 --epochs 200 --pretrained ./main_moco_checkpoint/arch-RB3D_epochs600_bs32_lr0.003_moco-k128_moco-dim128.pth.tar --dist-url "tcp://localhost:10001" --multiprocessing-distributed --world-size 1 --rank 0 --gpu 0 ./Datasets/ 2>&1 | tee ./logs/arch-RB3D_epochs600_bs32_lr0.003_moco-k128_moco-dim128_lincls_lr30.log

python main_lincls.py -a DSRF3D_v2 --lr 3 --schedule 30 60 90 --batch-size 32 --workers 8 --epochs 120 --pretrained ./main_moco_checkpoint/Nonorm-DSRF3D_v2_epochs600_bs32_lr0.003_moco-k128_moco-dim128.pth.tar --dist-url "tcp://localhost:10001" --multiprocessing-distributed --world-size 1 --rank 0 --gpu 0 ./Datasets/ 2>&1 | tee ./logs/Nonorm-DSRF3D_v2_epochs600_bs32_lr0.003_moco-k128_moco-dim128_lincls_lr3_epochs120.log