Comparison between KERAS library and FAST.AI library using convolution neural network(image classification) model
Authors
Uppari, Rahul Ramchandra
Issue Date
2020
Degree
MSc in Data Analytics
Publisher
Dublin Business School
Rights holder
Rights
Items in eSource are protected by copyright. Previously published items are made available in accordance with the copyright policy of the publisher/copyright holder.
Abstract
In past, many Data Scientists did research and developed model for Image Classification using Machine Learning and Deep Neural Network algorithms and through their it was found out Convolution Neural Network was and still the best Deep Neural Network Layer model for Image Segmentation, Image Classification and Object Detection. And to build CNN Model for image classification they used many Libraries such as KERAS, TENSORFLOW, SKLEARN, CAFFE, THEANO, NILEARN, FAST.AI and many more. Out of these libraries have choose only two libraries that has very good support of additional API Framework like KERAS uses TENSORBOARD which is the most used API Framework for Visualization and developing applications. Another library is FAST.AI which is uses PyTorch as support for specifically developing applications such as Computer Vision and Natural Language Processing etc. Both libraries have their own advantages, where KERAS has support of the world’s most used and high-level TENSORFLOW API Framework to run smooth on CPU and GPU. On other side FAST.AI uses pre-trained models and using them a basic image classification can be developed within few lines of codes which makes the coder work easy. To find out which library is best for Image Classification using Convolution Neural Network I developed this practical way to compare and find the answers. This project is implemented in a way, where the comparison can be done basis on the similar architecture, dataset, default Hyperparameters values, Tune Hyperparameter values, Epochs and Learning Rate. Both the models are built and have been effectively trained on 87000 images of American Sign Language dataset.