RickardKarl/bc_learning_image

PyTorch implementation and reproduction of between-class learning for image classification https://arxiv.org/abs/1711.10284

Reproduction of BC learning for images

This repository is a reproduction study of Between-class Learning for Image Classification by Yuji Tokozume, Yoshitaka Ushiku, and Tatsuya Harada.

This reproduction was made by Rickard Karlsson, Qian Bai and Erik Handberg. Our project and our results are presented in detail in this blog post.

Between-class (BC) learning

Below is how the authors of the original study describe BC learning in their repository:

Between-class (BC) learning:

• We generate between-class examples by mixing two training examples belonging to different classes with a random ratio.
• We then input the mixed data to the model and train the model to output the mixing ratio.
• Original paper: Learning from Between-class Examples for Deep Sound Recognition by us (github)

Our reproduction

From the original study, we have tried to reproduce selected parts, not the entire study. We have trained an 11-layer CNN to do standard, BC and BC+ learning on the CIFAR-10 dataset.

Data augmentation

We have trained an 11-layer CNN to do standard and BC learning without data augmentation on the CIFAR-10 dataset.

Ablation analysis

We have trained an 11-layer CNN on the CIFAR-10 dataset to do all suggested ablation analysis in the original study (Mixing method, Label, # mixed classes, Where to mix).

Pytorch

The original study created their model using Chainer. We ported this code to Pytorch. Therefore, the setup for running our model is slightly different.

Setup

• Install Pytorch on a machine with CUDA GPU, or use Google Colab.
• Run download_cifar10.py to download the CIFAR-10 dataset.

Training

• Example commands:

  • Standard learning on CIFAR-10:

      python main.py --dataset cifar10 --netType convnet --data path/to/dataset/directory/
    

  • BC learning on CIFAR-10:

      python main.py --dataset cifar10 --netType convnet --data path/to/dataset/directory/ --BC
    

  • BC+ learning on CIFAR-10:

      python main.py --dataset cifar10 --netType convnet --data path/to/dataset/directory/ --BC --plus
    

• Notes:

  • By default, the training runs 10 times. You can specify the number of trials by using --nTrials command.
  • Please check opts.py for other command line arguments.
  • Please checkout to the respective branch for ablation analysis training.

Results

Error rate for 11-layer CNN on CIFAR-10

Learning	Original study	Our result
Standard	6.07	6.59
BC	5.40	5.67
BC+	5.22	5.51

• For more of our results, please see our blog post
• For more of the results from the original study, please see the paper: Between-class Learning for Image Classification

Caltech101

We also trained an 11-layer CNN on the Caltech-101 dataset. Please checkout branch caltech101_experiment for the code used for training.

---