dogbreed-CNN

Dog breed classification using Transfer learning and Convolutional Neural Networks (CNN). This project was done as part of the Udacity Deep Learning nanodegree program.

Overview

The goal of this project is to apply the knowledge gained on how to create a CNN from scratch and then train it to recognize dog breeds from an image. After my initial attempt at building a simple CNN network, transfer learning will be used to improve the network's predictions and increase the accuracy of the recognized breeds. Addictionally, open-cv is used to detect faces in images and I'll be using the network to determine what that human looks like (in terms of dog breeds) 😆

Methodology

Outlined below are the high level steps I used. Important to remark that PyTorch and OpenCV were used for the entire project

1. Prepare datasets

   • Download the dog dataset and add it to the root folder of your project (/dogImages)
   • Download the human dataset (/lfw)
   • Get the full classes dictionary from ImageNet to associate class index with breeds
   • Prepare training, test and validation batches, making sure to augment your training data to avoid overfitting

2. Building CNN

   • First I'll use a simple CNN architechture, using 3 CNN layers for feature detection and 3 fully connected MLP to classify the images. This network has a low accuracy, but it's important to understand how different architecture choices and hyperparameter settings affect the result of your network
   • After this attempt I use transfer learning using the pretrained VGG16 model. The feature detection layers will be used as they come and I will focus on training the classifier layers, changing the last fully connected layer in order to fit our target case.
   • Setup optimization, loss function and related hyperparameters in order to achieve better results

3. App development

   • Using the newly trained network along with OpenCV for face recognition, I built a function that takes as input a path to an image and returns:
     • Results if a human face or a dog was detected
     • Apply model in order to predict dog breed or human-dog resemblance
     • If no faces or dogs found in image, do nothing

Results

After playing around with the model this is what I obtained with a couple sample images

Tools

For this project I used an AWS Deep Learning AMI (Ubuntu 18.04). I highly recommend using VS Code and remote-SSH into your instance (Make sure you have a security group that allows port 22 from your IP). I worked in pytorch_latest_p37 virtual env that comes with the forementioned instance, which already contains the libs open-cv and pytorch