The stochs library provides efficient C++ implementations of stochastic optimization algorithms for common machine learning settings, including situations with finite datasets augmented with random perturbations (e.g. data augmentation or dropout). The library is mainly used from Python through a Cython extension. Currently, SGD, (S-)MISO and (N-)SAGA are supported, for dense and sparse data. See the following reference for details:
A. Bietti and J. Mairal. Stochastic Optimization with Variance Reduction for Infinite Datasets with Finite-Sum Structure. NIPS, 2017.
The library requires Eigen >=3.3 (it will be downloaded automatically in the setup.py
script unless the folder or symlink include/Eigen
already exists)
and glog. To install glog on Ubuntu, run:
sudo apt-get install libgoogle-glog-dev
The Python package can be built with the following command (this requires Cython and a compiler with OpenMP support such as gcc, you might need to change the CC
and CXX
environment variables on a mac):
python3 setup.py build_ext -if
By default, the library is built for double precision floating point numbers (np.float64
), for single precision (np.float32
) set USE_FLOAT = 1
in setup.py
.
Example usage with dropout perturbations:
import numpy as np
import stochs
X, y, Xtest, ytest = load_some_dataset()
solver = stochs.MISO(X.shape[1], # number of features
X.shape[0], # number of datapoints
alpha=1.0, # initial step-size
lmbda=0.01, # L2 regularization
loss=b'squared_hinge', # squared hinge loss
prox=b'l1', # use L1 regularizer (by default 'none')
prox_weight=0.1, # L1 regularization weight
average=False) # no iterate averaging
n = X.shape[0]
for epoch in range(100):
if epoch == 2:
# start decaying the step-size after a few epochs
# if average=True, this also starts iterate averaging
solver.start_decay()
# pick random indexes for one epoch
idxs = np.random.choice(n, n)
# apply 10% dropout
Xperturbed = X[idxs] * np.random.binomial(1, 0.9, size=X.shape) / 0.9
# run algorithm on batch of perturbed data
solver.iterate(Xperturbed, y[idxs], idxs)
# with no perturbations, use: solver.iterate_indexed(X, y, idxs)
print(solver.compute_loss(Xtest, ytest)) # compute test loss
# access parameter vector with solver.w()
A more thorough example for sentiment analysis on the IMDB dataset is given in the examples
folder, with sparse solvers and non-uniform sampling.