i-c-karakozis / zeroshot_learning

Zero-shot Learning for Animal Classification

Zero-shot Learning

This repository hosts a pytorch project that performs Zero-shot Animal Classification on the "Animals with Attributes" dataset. For details and references, refer to the PDF report found at the top directory level of the repository.

Environmental Setup

This project requires Python 3.6.7. To install the python dependencies, run the following command (we recommend using a virtual environment that supports Python 3.6.7):

pip install -r requirements.txt

Dataset Download

The dataset annotations can be found in the annotations directory. To download the dataset, run the setup script from the top directory level of the repository as follows:

chmod +x setup.sh
./setup.sh

This script will also create all the directories that are needed to train and test your models.

Model Training

It is recommended that you train your models on a machine with access to GPU resources. To train your models, run the following python script:

python main.py

To inspect the parameters you can configure for training, execute python main.py --help. The debug flag should only be set when you want to debug your pipeline.

For more details on model training, refer to the project report.

Model Evaluation

It is recommended that you evaluate your models on a machine with access to GPU resources. To evaluate your models, run the following python script:

python main.py --model {0}

{0} should be replaced by the name of the file in the models directory that stores the weights of the pytorch model you want to evaluate. Models are automatically saved during training.

For more details on model evaluation, refer to the project report.

Note from the Author

Feedback and comments are highly appreciated. Feel free to open an issue on the repository page to express your thoughts and suggest changes.