Improving Few-shot learning for Image Classification using data augmentation and aggregation

This is a BSc Thesis that explores and seeks to improve the findings from the paper Revisiting Local Descriptor based Image-to-Class Measure for Few-shot Learning by Wenbin Li, Lei Wang, Jinglin Xu, Jing Huo, Yang Gao and Jiebo Luo. In CVPR 2019.

Prerequisites

• Unix system
• Python 3.8 - 3.10
• CUDA 10.0 - 12
• All library dependencies from requirements.txt (pip virtual environment recommended)

Getting Started

Installation

• Clone this repo:

git clone https://github.com/d33dler/avq2c.git

• Install requirements.txt dependencies with pip:

pip install -r requirements.txt

• cd executables/ and run python execute.py --help to see the available options.

Datasets

• miniImageNet.
• StanfordDog.
• StanfordCar. (Not tested)
• CUB-200.
  

For training and testing on different conditions you must edit the model root config. Refer to models/architectures/config_blueprints/CONFIG_DOCUMENTATION for the model config documentation.

You may also re-use config samples from models/architectures/config_blueprints/ to reproduce the results in the paper.

Training & testing

We run the script from the exec directory of the project.

cd executables

• Run

  python execute.py [--jobs JOB [JOB ...]] [--jobfile JOBFILE] [--job_id JOB_ID] [--test]

  Options:

  • jobs JOB [JOB ...] Paths(s) to the model config file(s). Each file should be in YAML format and contain the configuration parameters required for a job.

  • jobfile JOBFILE Path to a file containing a list of job arrays. The file should be in YAML format and list configurations or paths to configurations for multiple jobs.

  • job_id JOB_ID An integer representing the index of the job to be launched from the job array specified in the jobfile. Indexing starts at 1.

  • test Run the script in test mode. If this flag is set, the jobs will be launched in test mode.

• Results

• Plots