Hash Bit Selection with Reinforcement Learning

We release all cods and configurations for HBS-RL. For details, see Hash Bit Selection with Reinforcement Learning

Requirements

- PyTorch
- tmuxp    

Preparation

Construt Bit Pool

MNIST

Feature & LSH

python3 hmethod/hash_baseline/features/mnist.py

ITQ, SpH

python3 hmethod/hash_baseline/features/hash_baseline.py --methods ITQ SpH

Cifar10

Feature

python3 hmethod/hash_baseline/features/cifar10.py

SH, ITQ, SpH

python3 hmethod/hash_baseline/features/hash_baseline.py --dataset cifar10

Merge book

python3 hmethod/hash_baseline/features/sh-sph-itq-merge.py

Imagenet & Nus-Wide

Prepare Feature and Book files like MNIST and Cifar10

Feature files

cd hmethod/dataset/feature/
python3 feature.py --dataset imagenet
python3 feature.py --dataset nuswide

Book files

Reference DPSH, GreedyHash, and CSQ with codes in ./hmethod/dataset

Training & Testing

Train HBS-RL

tmuxp load scripts/algo-hbs-rl/mnist_train.yaml

Test & Evaluation

tmuxp load scripts/algo-hbs-rl/test.yaml

Plot the entropy curves

for bit in {32,64,128};do
python3 ./scripts/plot.py ./data/checkpoint/mnist/LSH/entropy/ -cf 1 -lf progress.txt -x Episode -y MAP_TRAIN -i $bit -s 100 -o 0.05 -S 1 -L '$\alpha=0$' '$\alpha=0.01$' '$\alpha=0.02$' '$\alpha=0.03$' '$\alpha=0.04$' -T MNIST' @ '$bit' bits' -yl 'MAP (Train)' -cl 'HBS-RL (LSH):' -cL left -LL 'lower right' -p ent_$bit.pdf;done

python3 main.py --datasets mnist --hmethods lsh --num_bits 512 --num_sub_bits 64 --topK 0
python3 main_ppo.py --num_bits '512' --num_sub_bits '128' --dataset 'mnist' --hmethod 'lsh' --device 'cuda:5'

Plot Precision Recall P-R

python3 evaluate.py --plot True --dataset mnist --method LSH{PPO_NDomSet_Random}512 --num_bit 32 --topK 50000 --device cuda:0
python3 evaluate.py --plot True --dataset mnist --method SpH{PPO_NDomSet_Random}512 --num_bit 32 --topK 50000 --device cuda:0