hideandseek 0.1.19

library for deep learning and privacy preserving deep learning

hideandseek

Highly modularized deep learning training library.

Why use hideandseek?

• 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

Currently prettifying codes. (30.10.2022.)

import torch
import torch.nn as nn

# Generate data
x = torch.rand(200,1)
y = 5*x+2

network = nn.Linear(1,1)
dataset = torch.utils.data.TensorDataset(x, y)
criterion = nn.MSELoss()
cfg = {
'lr': 1e-2,
'batch_size': 32,
'epoch': 10 # optional
}

# Training configuration. All you need to train a neural network
kwargs = {
'network':network,
'dataset':dataset,
'cfg_train':cfg,
'criterion':criterion,
'name': 'Test' # optional
}
trainer = hs.N.Node(**kwargs)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
trainer.network.to(device)

# Train for predefined number of epochs
trainer.train() # Train for predefined number of epochs
trainer.train(5) # Train for specified number of epochs
trainer.train(epoch=5) # Same thing with trainer.train(5)
trainer.train(step=500) # Train for specified number of updates

trainer.network.cpu()

and simply run multiple batch of experiments with a single line command such as:

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
• GUI for generating experiment scripts when conducting variable sweeps