Tackling Cats vs Dogs kaggle challenge using transfer-learning and pytorch

Introduction

The challenge provides a dataset of

• 12,500 Cat photos

• 12,500 Dog photos

and then asks us to predict the category of 12,500 test images (Cat or Dog). The submission is evaluated based on log loss (smaller is better):

Approach

As shown in the figure below, my approach will be as follows:

1. Use a pretrained imagenet model (e.g. resnet-152)

2. Replace the last fully-connected layer with a new one with binary output, while freezing the rest of the layers and train on the given image dataset.

3. Unfreeze a few preceding layers and retrain for further fine-tuning.

As in some other projects, I'll use Pytorch , which I am increasingly a fan of.

Here's the link to the notebook: https://nbviewer.jupyter.org/github/sangeetkar/cats_vs_dogs_kaggle/blob/master/cnet.ipynb