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:

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