hideandseek 0.3.27

library for deep learning and privacy preserving deep learning

hideandseek

Highly modularized deep learning training library.

Why use hideandseek?

• Only code the experiment specific parts of ANN experiment (dataset, nn.Module definition, criterion, forward pass)

• Define every ANN experiment parameters in keyword argument, for easy experimental control

• Takes care of every other logistics (logging, device type matching, amp, random batching control)

• Easy training & saving deep learning models along with other modules (ex: preprocessing modules) required in inference

• Run multiple deep learning experiments in parallel on multiples GPUs (powered by hydra, and python multiprocessing)

• Design and analyze experiments scientifically by modifying variables (powered by hydra)

• Modularized machine learning pipeline allows using the same script for all types of experiments

• The same training code can be run in privacy preserving setting by minimal modifications

Look at Simple/train.py for simple run cases

Single run:

python train.py lr=1e-3 batch_Size=32 random_seed=0

Multirun with batch of experiments (Hyperparameter sweep):

python train.py -m lr=1e-3,1e-2 batch_size=32,64 "random_seed=range(0,5)" \
hydra/launcher=joblib hydra.launcher.n_jobs=8
# Runs total of 2*2*5=40 batch of experiments, with 8 processes at a time. Experiment results are stored in hydra.sweep.dir which can be overridden.

To do

• Draw figures to explain hideandseek
• .py based tutorial
• `.ipynb' based tutorial
• GUI for generating experiment scripts when conducting variable sweeps