m-lecuyer/adaptive_rdp

adaptive_rdp

Requirements

• CUDA 10.2+ drivers
• python3.7
• torch torchvision opacus tqdm matplotlib pandas requests torchcsprng tensorboard seaborn

Plot results from included log files

python plots.py

Or:

make plot

Re-run paper's experiments

python exp_adaptive.py
python exp_fine_tuning.py

Or:

make exp

Train a single model

For tensorbord logs:

tensorboard --logdir=/tmp/stat

Run e.g.:

# Adaptive training:
  python main.py -b 128 -na 4 --sigma 1 --filter 10 --adaptive noise --adaptive-metric train_acc --adaptive-strategy updown --maximum-factor 1. --adaptation-additive .1 --adaptation-period 10 --move-down-period 50 --lr 0.05 --epochs 300 --wd 0 --model-type ResNet9 --log-dir ResNet9_adaptive

# Fine-tuning:
python main.py -b 64 -na 8 --sigma 2 --odometer --adaptive noise --adaptive-metric train_acc --adaptive-strategy down --maximum-factor 2. --minimum_factor 1. --adaptation-additive .1 --adaptation-period 1 --enable-bn --lr 0.01 --epochs 20 --wd 0 --model-type ResNet18 --fine-tune last --log-dir ResNet18_finetuning