In this project, we dive into the world of deep learning by utilizing the Keras library to classify fashion articles in the Fashion MNIST dataset. Keras is a high-level neural networks API, written in Python and capable of running on top of TensorFlow, CNTK, or Theano. It is user-friendly, modular, and extensible, which makes it a go-to library for both beginners and experienced practitioners in the field of deep learning.
The dataset we are working with is the Fashion MNIST dataset, which is a collection of 70,000 grayscale images across 10 categories. Each image is a 28x28 pixel representation of a fashion item, such as a shirt, dress, sneaker, etc. The dataset is divided into 60,000 training samples and 10,000 testing samples. This dataset is intended to serve as a direct drop-in replacement for the original MNIST dataset for benchmarking machine learning algorithms, as it is a more challenging dataset compared to the original MNIST.
The images are reshaped from a 2D array of 28x28 pixels to a flat 1D array of 784 pixels. Data normalization is performed by converting the pixel values from integers to floats and scaling them to the range [0, 1].
The following tools and libraries are utilized in this project:
Python Keras TensorFlow NumPy Matplotlib
The neural network model built in this project is a simple sequential model with the following layers:
- A dense layer with 784 inputs (flattened image pixels) and 100 neurons, using ReLU activation function.
- A softmax output layer with 10 neurons, corresponding to the 10 classes of the dataset.
The code is structured as follows:
- Import necessary libraries.
- Load and preprocess the Fashion MNIST dataset.
- Define the neural network model architecture.
- Compile the model (the code for this is not included in the provided snippet).
achieving an exceptional accuracy of 99% on the Fashion MNIST dataset. This project demonstrated my ability to preprocess data, fine-tune neural network architectures, and rigorously evaluate model performance to ensure high precision in image classification tasks.
This project serves as a fundamental step into the domain of image classification using neural networks. The Fashion MNIST dataset provides a more challenging alternative to the classic MNIST dataset and serves as a good benchmark for algorithms.
Please find the complete code on my GitHub repository at [Project-DL-Keras.ipynb].