Udacity Data Science Nanodegree

Capstone Project - Dog Breed Classifier

The goal of this project is to detect the breed of a dog using Convolutional Neural Network (CNN) with at least 60% accuracy.

Software Requirements

Python 3.x Numpy Pandas Matplotlib Scikit-Learn Keras Tensorflow OpenCV

Metrics:

For this project, the metric used is the accuracy score.

File structure and desciption:

|   .gitattributes
|   dog_app.ipynb
|   extract_bottleneck_features.py
|   LICENSE
|   README.md
|   
+---haarcascades
|       haarcascade_frontalface_alt.xml
|       
+---images
|       American_water_spaniel_00648.jpg
|       Brittany_02625.jpg
|       Curly-coated_retriever_03896.jpg
|       Labrador_retriever_06449.jpg
|       Labrador_retriever_06455.jpg
|       Labrador_retriever_06457.jpg
|       sample_cnn.png
|       sample_dog_output.png
|       sample_human_2.png
|       sample_human_output.png
|       Welsh_springer_spaniel_08203.jpg
|       
+---saved_models
|       InceptionV3_model.weights.best.hdf5
|       weights.best.from_scratch.hdf5
|       weights.best.VGG16.hdf5
|       
\---test images
        Dodo.jpg
        Grate Dane.jpg
        Labrador Retriver.jpg
        Ryan Gosling.jpg
        Ryan Reynolds.jpg
        Scooby Doo.jpg

The above tree describes all the files as part of this project.

dog_app.ipynb The main file for this project, it contains all the codes and also has the outputs.
extract_bottleneck_features.py : it is used as a module in the main file to extract bottleneck features
haar cascades: Contains the haarcascade XML file to be used by cascade classifiers for face detection.
saved_models: contains the saved best model weights of pre-trained models. It doesn't include InceptionV3 weights because of its large size to upload. It can be generated by training the model. Steps in the next section.
images and test images: images to test the model

Steps to replicate the project

All the dependencies are in requirement.txt and requirements-gpu.txt. A new virtual environment is preferable. Using python python -m venv myenv

using Anaconda distribution conda create --name myenv

Then all the dependencies can be installed using pip pip install -r requirements.txt
Clone the repo in new venv
Download the dog image dataset. Unzip the folder at location path/to/dog-project/dogImages.
Download the human image dataset. Unzip the folder at location path/to/dog-project/lfw.
Download the VGG-16 and InceptionV3 bottleneck features for the dog dataset. Place it in the repo, at location path/to/dog-project/bottleneck_features.
Run the jupyter notebook

Takeaways and final thoughts

The InceptionV3 model achieves an accuracy of 80% after 20 epochs. The accuracy can be improved in various ways like increasing the number of epochs, trying other pre-trained models, etc. The next step for me is to create a web app that takes an image as an input, then gives the result as a breed of dog. I have also written an accompanying medium blog https://medium.com/@akash.ashu17/dog-breed-classifier-6a0d5fb34e80

Acknowledgement and Licensing terms

I am thankful to Udacity for providing such wonderful tutorials. All the datasets are provided by Udacity. Kindly refer to Udacity Terms of Service for further information.

akashashu1799 / dogbreed