Deep Bayesian Active Learning with Image Data

Unofficial implementation of "Deep Bayesian Active Learning with Image Data" by Yarin Gal, Riashat Islam, Zoubin Ghahramani (ICML 2017) using Pytorch.

About The Paper

In this paper, Gal et al. combine recent advances in Bayesian deep learning into the active learning framework in a practical way -- an active learning framework for high dimensional data, a task which has been extremely challenging so far.

By taking advantage of specialised models such as Bayesian convolutional neural network, the proposed technique obtains a significant improvement on existing active learning approaches.

Methodology

Compare various acquisition functions: Bayesian Active Learning by Disagreement (BALD, Houlsby et al., 2011),Variation Ratios (Freeman, 1965), Max Entropy (Shannon, 1948), Mean STD (Kampffmeyer et al., 2016;Kendall et al., 2015) and baseline Random relying on Bayesian CNN uncertainty with simple image classification benchmark. All acquisition functions are assessed with same model structure:

Convolution-relu-convolution-relu-max pooling-dropout-dense-relu-dropout-dense-softmax

With 32 convolution kernels, 4x4 kernel size, 2x2 pooling, dense layer with 128 units and dropout probabilities 0.25 and 0.5.

All models are trained on MNIST dataset with random initial training set of 20 datapoints and a validation set of 100 points on optimised weight decay. A standard test set of 10K is used and the rest of the points are used as pool set. The test error of each model and each acquisition function is assessed after each acquisition using dropout approximation at test time.

Monte Carlo dropout is used to decide which datapoints to query next. Repeat the acquisition process for 100 times and acquiring 10 points that maximise the functions for each time. (Total acq points=1000)

Getting Started

This repo consists of 4 experiments conducted in the paper which are:

1. Comparison of various acquisition functions
2. Importance of model uncertainty
3. Comparison to current active learning technqiues with image data (Minimum Bayes Risk, MBR)
4. Comparison to semi-supervised learning

Prerequisites

• Python 3.5 or later

In Ubuntu, you can install Python 3 like this:

$ sudo apt-get install python3 python3-pip

For Windows and MacOS, please refer to https://www.python.org/getit/.

Installation

Use pip3 install -r requirements.txt to install:

• pytest (for testing)
• modAL (modular active learning framework for Pytorch)
• skorch (a scikit-learn wrapper for Pytorch)
• pytorch
• numpy
• matplotlib
• scipy
• scikit-learn

Pytest Test Suite

1. To run test:

$ pytest

Launch

Run

$ python3 main.py --batch_size 128 \
                  --epochs 50 \
                  --lr 1e-3 \
                  --seed 369 \
                  --experiments 3 \
                  --dropout_iter 100 \
                  --query 10 \
                  --acq_func 0 \
                  --val_size 100 \
                  --result_dir result_npy

Or use --help for more info.

Note

• --determ is set as False by default for Experiment 1, add this to run Experiment 2.
• --val_size is set as 100 by default for Experiment 1. To run Experiment 4, please set this to 5000.
• In this implementation, acqusition_iterations = dropout_iterations = 100
• For Experiment 3, please refer to comparison_to_MBR.ipynb or Google Colab link here.

Results

1. Comparison of various acquisition functions

Number of acquired images to get model error of %: (the lower the better)

Techniques	10% error (Paper: Keras)	10% error (Experiment: Pytorch)	5% error (Paper: Keras)	5% error (Experiment: Pytorch)
Random (Baseline)	255	250	835	517
Mean STD	230	100	695	295
BALD	145	150	335	296
Var Ratios	120	143	295	283
Max Entropy	165	163	355	310

Average running time for each experiment (100 validation points) on Google Colab (Tesla T4 15GB)

To further reduce computational time, 2000 random points subset will be used instead of whole pool. These datapoints are selected randomly from pool data points using

np.random.choice(range(len(X_pool)), size=2000, replace=False)

• Random: ~2m 17s
• BALD: ~10m 52s
• Var Ratios: ~10m 58s
• Max Entropy: ~10m 39s
• Mean STD: ~10m 40s

Best 2 models: Mean STD, Var Ratios

2. Importance of model uncertainty

BALD

Var Ratios

Max Entropy

3. Comparison to current active learning techniques

Note:

This experiment is run on Binary Classification test. (MNIST two digit classification)

4. Comparison to semi-supervised learning

Test error on MNIST with 1000 acquired images, using 5000 validation points:

Technique	Test error (Paper: Keras)	Test error (Experiment: Pytorch)
Random(Baseline)	4.66%	3.73%
Mean STD	-	1.81%
BALD	1.80%	1.81%
Max Entropy	1.74%	1.66%
Var Ratios	1.64%	1.57%

Best 2 models: Var Ratios, Max Entropy

Acknowledgements

1. https://github.com/Riashat/Deep-Bayesian-Active-Learning
2. https://github.com/damienlancry/DBAL
3. https://github.com/modAL-python/modAL/blob/master/examples/pytorch_integration.py