Transfer learning

This repository contains code for transfer learning using pretrained Inception network on CIFAR-10 dataset.

Installation

Required packages

1. Tensorflow: https://www.tensorflow.org/install/install_linux#InstallingAnaconda
2. Tensorpack - used for data loading and preprocessing:
   pip install -U git+https://github.com/ppwwyyxx/tensorpack.git
3. Skiimage - used for extracting HOG features from images:
   pip install scikit-image
4. GPyOpt - used for Bayesian Optimisation of hyper-parameters:
   pip install gpyopt

Pretrained CNN codes

Pretrained CNN codes are available here: https://drive.google.com/open?id=0B6fInPVjwoO1d0kzUTY1OVBtbDg

Testing

To test the setup run:
python -m unittest tests

Scripts

CNN codes extraction

Extract CNN codes for the test examples:
python extract.py --mode test

Extract CNN codes for the training examples:
python extract.py --mode train

Visualisations

Examples from test dataset, output stored in examples.png:
python plotting.py --mode plot_examples

Examples from test dataset and their HOG features, outputs stored in examples.png and examples_hog.png:
python plotting.py --mode plot_hog

T-SNE embedding of test examples with class segmentation and original images:
python plotting.py --mode plot_cnn

Classification

Available classification methods include: softmax_raw, svm_raw, svm_hog, svm_hog_kern, svm_cnn, svm_cnn_kern.

Example:
python classify.py --mode svm_cnn

Method	Accuracy
Linear SVM on HOG features	46.9%
Kernelized SVM on HOG features	31.4%
Linear SVM on CNN codes	89.6%
Kernelized SVM on CNN codes	86.6%
Optimized Linear SVM on CNN codes	90.6%

Bayesian Optimisation

For tunning the hyperparamters. Example:
python bayes_opt.py --mode linear