DeepCatnDogRecognition

It's all about being a:

0: Cat
1: Dog

Introduction

This project proposes three differents Deep Neural Network (DNN) to face the well known Kaggle Dog and Cat Classification challenge. You can get more details about it at https://www.kaggle.com/c/dogs-vs-cats.

The DNN here used for the purpose of beating the challenge are the following:

LeNet
DeepConvNet
AlexNet (pre-trained model)

You can find more information about these DNN here: https://medium.com/@sidereal/cnns-architectures-lenet-alexnet-vgg-googlenet-resnet-and-more-666091488df5

Each of these DNN is really sensitive to several hyperparameters, which in turn provide a better or worse fine tuning.

Requirements

python 2.7
tensorflow 1.9.0 or greater
opencv

Usage

First of all, download the repository by running:

git clone https://github.com/DavideNardone/DeepCatnDogRecognition.git

unzip DeepCatnDogRecognition-master.py

Before training or testing any of the DNN available in this project, you must create some directories under the root project by running the following commands:

mkdir plots
mkdir results
mkdir ckpt
mkdir npys

Once you've done this, you might need to set up some paths, located in the following python files:

src/train_net src/configs src/create_test_submission tools/loader

To train the network run:

python train_net.py <MODEL>

where the MODEL parameter specify the kind of DNN you want to train (e.g., LE-NET, DEEP-CONVNET, ALEX-NET)

The latter network is a pre-trained network and it works by loading its proper weights which can be downloaded at the following link:
http://www.cs.toronto.edu/~guerzhoy/tf_alexnet/ that need to be located at the root of the project folder.

To create the submission test useful for classifying the test set, run the following command:

python create_test_submission.py <MODEL>

Dataset

The dataset used for train and test the DNN can be downloaded at the following link: https://www.kaggle.com/c/dogs-vs-cats-redux-kernels-edition/data

Once you've downloaded it, you must unzip and locate it under the root project in order to make everything run smoothly.

Results

The models accuracy/loss achieved both on train and validation are shown below:

TRAINING_SET

LE-NET (~98 epoches)
+/- 0.78 ACC
+/- 0.5 LOSS

DEEP-CONVNET (~20 epoches)
+/- 0.86 ACC
+/- 0.4 LOSS

ALEX-NET (~15 epoches)
+/- 0.93 ACC
+/- 0.9 LOSS

VALIDATION SET

LE-NET
+/- 0.70 ACC
+/- 0.5 LOSS

DEEP-CONVNET
+/- 0.93 ACC
+/- 0.2 LOSS

ALEX-NET
+/- 0.9 ACC
+/- 0.9 LOSS

NB: As you might know, DNN are ruled by several hyperparameters that in turn affect the fine tuning process. The results presented here might not be the best one since a better fine tuning process and a more efficient image pre-processing might led to better results.

Troubleshootings

In case something goes wrong, make sure:

You've created all the required directories
Any configuration path is rightly set up
You've installed the right library versions

Authors

Davide Nardone, University of Naples Parthenope, Science and Techonlogies Departement,
Msc Applied Computer Science
https://www.linkedin.com/in/davide-nardone-127428102

Contacts

For any kind of problem, questions, ideas or suggestions, please don't esitate to contact me at:

davide.nardone@live.it