Learning with Self-Supervised Regularization

Update: this repo exists only for the reproduction of the experiments presented in the paper below. The user is encouraged to check the PyTorch version of this repo for practical semi-supervised image classification on large realistic datasets using modern CNN backbones, along with the latest developments and features. https://github.com/FlyreelAI/sesemi

This repository contains a Keras implementation of the SESEMI architecture for supervised and semi-supervised image classification, as described in the NeurIPS'19 LIRE Workshop paper:

Tran, Phi Vu (2019) Exploring Self-Supervised Regularization for Supervised and Semi-Supervised Learning.

Approach

The training and evaluation of the SESEMI architecture for supervised and semi-supervised learning is summarized as follows:

Separate the input data into labeled and unlabeled branches. The unlabeled branch consists of all available training examples, but without ground truth label information;
Perform geometric transformations on unlabeled data to produce six proxy labels defined as image rotations belonging in the set of {0,90,180,270} degrees along with horizontal (left-right) and vertical (up-down) flips;
Apply input data augmentation and noise to each branch independently;
At each training step, sample two mini-batches having the same number of unlabeled and labeled examples as inputs to a shared CNN backbone. Note that labeled examples will repeat in a mini-batch because the number of unlabeled examples is much greater;
Compute the supervised cross-entropy loss using ground truth labels and the self-supervised cross-entropy loss using proxy labels generated from image rotations and flips;
Update CNN parameters via stochastic gradient descent by minimizing the sum of supervised and self-supervised loss components;
At inference time, take the supervised branch of the network to make predictions on test data and discard the self-supervised branch.

Requirements

The code is tested on Ubuntu 16.04 with the following components:

Software

Anaconda Python 3.6;
Keras 2.2.4 with TensorFlow GPU 1.12.0 backend;
CUDA 9.1 with CuDNN 7.1 acceleration.

Hardware

This reference implementation loads all data into system memory and utilizes GPU for model training and evaluation. The following hardware specifications are highly recommended:

At least 64GB of system RAM;
NVIDIA GeForce GTX TITAN X GPU or better.

Usage

For training and evaluation, execute the following bash commands in the same directory where the code resides. Ensure the datasets have been downloaded into their respective directories.

# Set the PYTHONPATH environment variable.
$ export PYTHONPATH="/path/to/this/repo:$PYTHONPATH"

# Train and evaluate SESEMI.
$ python train_evaluate_sesemi.py
	--network <network_str>
	--dataset <dataset_str>
	--labels <nb_labels>
	--gpu <gpu_id>

# Train and evaluate SESEMI with unlabeled extra data from Tiny Images.
$ python train_evaluate_sesemi_tinyimages.py
	--network <network_str>
	--extra <nb_extra>
	--gpu <gpu_id>

The required flags are:

<network_str> refers to one of convnet, wrn, or nin architecture;
<dataset_str> refers to one of three supported datasets svhn, cifar10, and cifar100;
<nb_labels> is an integer denoting the number of labeled examples;
<nb_extra> denotes the amount of unlabeled extra data to sample from Tiny Images;
<gpu_id> is a string denoting the GPU device ID, defaults to 0 if not specified.

vuptran / sesemi