ChimeraMix: Image Classification on Small Datasets via Masked Feature Mixing

This is the official implementation of the paper ChimeraMix: Image Classification on Small Datasets via Masked Feature Mixing (IJCAI 2022).

Abstract

Deep convolutional neural networks require large amounts of labeled data samples. For many real-world applications, this is a major limitation which is commonly treated by augmentation methods. In this work, we address the problem of learning deep neural networks on small datasets. Our proposed architecture called ChimeraMix learns a data augmentation by generating compositions of instances. The generative model encodes images in pairs, combines the features guided by a mask, and creates new samples. For evaluation, all methods are trained from scratch without any additional data. Several experiments on benchmark datasets, e.g., ciFAIR-10, STL-10, and ciFAIR-100, demonstrate the superior performance of ChimeraMix compared to current state-of-the-art methods for classification on small datasets.

Citation

bibtex @inproceedings{chimeramix, title = {ChimeraMix: Image Classification on Small Datasets via Masked Feature Mixing}, author = {Reinders, Christoph and Schubert, Frederik and Rosenhahn, Bodo}, booktitle = {Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, {IJCAI-22}}, publisher = {International Joint Conferences on Artificial Intelligence Organization}, editor = {Lud De Raedt}, pages = {1298--1305}, year = {2022}, month = {7}, note = {Main Track}, doi = {10.24963/ijcai.2022/181}, url = {https://doi.org/10.24963/ijcai.2022/181}, }

Installation

The Anaconda environment can be created as follows.

bash conda env create -f environment.yaml conda activate chimeramix

Experiments

Training ChimeraMix on ciFAIR-10, ciFAIR-100, and STL-10 is shown in the following sections. You can set the number of examples per class via the max_labels_per_class parameter. The experiments require a single GPU with 10GB of memory, such as the NVIDIA GeForce GTX 1080 Ti.

ciFAIR-10

```bash

ChimeraMix+Grid

python traingenerator.py +dataset=cifair10 +experiment=chimeramixgrid maxlabelsperclass=5 python trainclassifier.py +dataset=cifair10 +experiment=chimeramixgrid maxlabelsperclass=5

ChimeraMix+Seg

python traingenerator.py +dataset=cifair10 +experiment=chimeramixsegmentation maxlabelsperclass=5 python trainclassifier.py +dataset=cifair10 +experiment=chimeramixsegmentation maxlabelsperclass=5 ```

ciFAIR-100

```bash

ChimeraMix+Grid

python traingenerator.py +dataset=cifair100 +experiment=chimeramixgrid maxlabelsperclass=5 python trainclassifier.py +dataset=cifair100 +experiment=chimeramixgrid maxlabelsperclass=5

ChimeraMix+Seg

python traingenerator.py +dataset=cifair100 +experiment=chimeramixsegmentation maxlabelsperclass=5 python trainclassifier.py +dataset=cifair100 +experiment=chimeramixsegmentation maxlabelsperclass=5 ```

STL-10

```bash

ChimeraMix+Grid

python traingenerator.py +dataset=stl10 +experiment=chimeramixgrid maxlabelsperclass=5 python trainclassifier.py +dataset=stl10 +experiment=chimeramixgrid maxlabelsperclass=5

ChimeraMix+Seg

python traingenerator.py +dataset=stl10 +experiment=chimeramixsegmentation maxlabelsperclass=5 python trainclassifier.py +dataset=stl10 +experiment=chimeramixsegmentation maxlabelsperclass=5 ```

Experiment Sweeps

To reproduce all main experiments on your Slurm cluster, execute the following two commands. Replace <SLURM PARTITION> with the name of your Slurm partition.

bash python train_generator.py "+dataset=glob(*)" "+experiment=glob(*)" "max_labels_per_class=5,10,20,30,50,100" "seed=range(0,5)" "hydra.launcher.partition=<SLURM PARTITION>" --multirun python train_classifier.py "+dataset=glob(*)" "+experiment=glob(*)" "max_labels_per_class=5,10,20,30,50,100" "seed=range(0,5)" "hydra.launcher.partition=<SLURM PARTITION>" --multirun