train.md

SPICE training

A step-by-step training tutorial for STL10 datast is as follows.

1. Pretrain representation learning model, i.e., MoCo, assuming 4 GPUs available.

python tools/train_moco.py

2. Precompute embedding features

python tools/pre_compute_embedding.py

3. Train SPICE-Self

python tools/train_self_v2.py

4. Determine reliable images

python tools/local_consistency.py

5. Train SPICE-Semi, assuming 4 GPUs available.

python ./tools/train_semi.py --unlabeled 1 --num_classes 10 --num_workers 4 --dist-url tcp://localhost:10001 --label_file ./results/stl10/eval/labels_reliable_0.983136_6760.npy --save_dir ./results/stl10/spice_semi --save_name 098_6760 --batch_size 64  --net WideResNet_stl10 --data_dir ./datasets/stl10 --dataset stl10

Note that --label_file and --save_name should be changed according to your generated reliable label file.

TODO: More training descriptions on other datasets will be added, and some training steps will be merged.