Introduction:

This project is the implementation of the paper: <b>"Using filter banks in Convolutional Neural Networks for texture classification"</b> [arXiv] in pytorch. <br/> V. Andrearczyk & Paul F. Whelan

In this project, TCNN3 architecture is implemented in end to end manner from scratch(<b>No pretraining</b>) for the DTD dataset.

Architecture

<img src="https://github.com/Shantanu48114860/TCNN3/blob/master/TCNN3.png">

Dataset:

Download the raw images (train/test/val) from the following link:<br/> https://www.robots.ox.ac.uk/~vgg/data/dtd/

Prepare the train test and validation images:

Edit the ROOT_PATH variable in create_texture_train_test_val_file.py to give the path of the downloaded images from DTD dataset link. <br/>

Run the following command: <br/> <b>python3 create_texture_train_test_val_file.py</b> <br/> This will seprate out the the 10 splits of train, test and validtion files in the ./images folder. <br/>

Create pickle files for train test and validation images

Run the following command:<br/> <b>python3 create_pickle.py</b> <br/> This will create the train, test and val pickle files in the folder: ./Dataset/Texture/DTD/ <br/>

Software required:

Python(ver: 3.7)<br/> Pytorch(ver: 1.3.1)

Models:

Pretrained pytorch models for DTD dataset of TCNN3 architecture can be dowbloaded from the following link:<br/> https://uflorida-my.sharepoint.com/:f:/g/personal/shantanughosh_ufl_edu/EsslShM1m61Ji2lxzrtI9gUB-yqIhDIntbkzaVHPlYv1vQ?e=Z0CBah <br/>

Training:

python3 train.py<br/> This will create the models and place them in the ./Models folder.

Testing:

python3 test.py

Hyperparameters:

Epochs: 400<br/> Learning rate: 0.0001<br/> Batch size: 32<br/> Weight Decay: 0.0005<br/>

Accuracy on the DTD dataset:

The accuracy is replicated as 27.8 % on the DTD dataset training from scratch(end to end as specified by the authors of the paper)