fnbalves / zero_shot_learning Goto Github PK

This project's objective is to reproduce the results from

https://static.googleusercontent.com/media/research.google.com/pt-BR//pubs/archive/41473.pdf

However, we use a smaller dataset (CIFAR-100) with lower resolution images. In order to do so, new cost functions needed to be created.

To use this code

1. Get CIFAR-10 and CIFAR-100 from:

https://www.cs.toronto.edu/~kriz/cifar.html

Put cifar-100 test, train and meta files inside a folder called cifar-100-python/ (read_cifar100 conde uses them)

2. Download the glove model from

http://nlp.stanford.edu/data/glove.6B.zip

And extract the files into a folder called glove.6B

3. Install python dependencies by using pip and the requirements.txt file:

sudo pip install requirements.txt

4. Create a folder called pickle files and run read_cifar100 to create all datasets

5. To train the composite model, run the train_composite file

6. To visualize the TSNE plots, run the visualize_results file (Change the indicated vars on the code)

7. To compute quantitative results, run the compute_quantitative_results file and use the functions (Change the indicated vars on the code)