Fast-SharpCF-pytorch

Fast Implementation of Adversarial Collaborative Filtering for Free (RecSys23)

Fast Version of SharpCF (RecSys23). Slow version of Pre-Post Processing is based on BPR-pytorch.
I assume that the paper used pre/pos precessing of BPR-pytorch, because the Training Time(MovieLens-1M, 17.1s) is similar to BPR-pytorch.

Contribution Points (Running Time Improvement: 11.8663x)

1. Remove DataLoader used in BPR-pytorch (Improvement: 46.4108x)
   • Remove IPC(Inter Process Communiation) Overhead that Multi-Process DataLoader has
   • Remove Data Loading Process(CPU->GPU) Every Epoch
2. Negative Sampling (Improvement: 429.898x)
   • Negative Sampling at Every Epoch is Done within only GPU, not CPU
3. HR, nDCG Calculation (Improvement: 125.357x)
   • HR and nDCG is done within only GPU, not CPU

	01.train()	02.neg_sample	03.get_batch	04.zero_grad	05.backward	05.forward	06.step	07.eval()	08.hr_ndcg	total
slow(origin)	3.49839e-05	14.3219	2.68599	0.0379853	0.976263	0.37763	0.177941	0.000172035	2.54794	21.1258
fast(optimized)	3.73626e-05	0.0333146	0.0578742	0.0405509	1.00162	0.424448	0.20211	3.7194e-05	0.0203255	1.78032
improvements(fast over slow)	0.936335	429.898	46.4108	0.936733	0.974684	0.889698	0.880417	4.62535	125.357	11.8663
paper	nan	nan	nan	nan	nan	nan	nan	nan	nan	17.1
improvements(fast over paper)	nan	nan	nan	nan	nan	nan	nan	nan	nan	9.60503

Hardward Specs - CPU: Virtual Machine - GPU: RTX-3090-ti

Latency Breakdown (Optimized Version vs Slow Version)

nDCG over Epochs (Trajactory Loss is used after 40-th epoch, nDCG@10)

Data

• Dataset is used in https://github.com/hexiangnan/adversarial_personalized_ranking/tree/master/Data
• To Select Sepecific Dataset('ml-1m', 'pinterrest-20', 'yelp'), modify dataset variable in config.py

Prepare Dependencies and Dataset

git submodule sync
git submodule update --init --recursive
pip install -r requiremetns.txt
python ./download_dataset.py

Reproduce Results in README.md

Fast Version of SharpCF (warm epoch: 40)

python ./main.py --factor_num=128 --batch_size 4096 --epochs 1000 --free_warm 40 --free_lambda 4096 --lr 0.01 --top_k 10 --chk_period 100 --out_dir ./out_report_post

Fast Version of BPR (warm epoch: 1000)

python ./main.py --factor_num=128 --batch_size 4096 --epochs 1000 --free_warm 1000 --free_lambda 4096 --lr 0.01 --top_k 10 --chk_period 100 --out_dir ./out_report_post

Slow Version of SharpCF (warm epoch: 40) which Might Seem like Paper's Implementation.

python ./slow_main.py --factor_num=128 --batch_size 4096 --epochs 1000 --free_warm 40 --free_lambda 4096 --lr 0.01 --top_k 10 --chk_period 100 --out_dir ./out_slow_report_post

Generate Plotly Plot (*.html) in out_dir

python ./generate_plot.py --sharp_dir ./out_report_post/batch_4096_lr_0.01_epochs_1000_freewarm_40_freelambda_4096.0/ --bpr_dir ./out_report_post/batch_4096_lr_0.01_epochs_1000_freewarm_1000_freelambda_4096.0/ --slow_dir ./out_slow_report_post/batch_4096_lr_0.01_epochs_1000_freewarm_40_freelambda_4096.0/ --out_dir ./out_plot