mnist-cnn-augmentation

A PyTorch implementation of a Convolutional Neural Network (CNN) for classifying the MNIST dataset, featuring data augmentation and real-time training visualizations.

Download the code and run it in the command line to download the MNIST dataset and train your neural networks.

Prerequisites

At the beginning of this project I would advise creating a virtual environment for the project to run in:

python3 -m venv dl_env  # Create a virtual environment
source dl_env/bin/activate  # Activate it (Linux/macOS)
# On Windows, use: dl_env\Scripts\activate

You should see (dl_env) to the left of your shell prompt now.

Additionally, ensure you have the following installed:

• Python 3.x
• PyTorch
• Matplotlib

To install the necessary dependencies, run the following command:

pip install torch torchvision torchaudio numpy pandas matplotlib scikit-learn

Feedforward Neural Network for the first test

Run with the following command:

python3 mnist_pytorch.py

This will:

1. Download the MNIST dataset
2. Train a simple 2-layer neural network
3. Print loss after each epoch

CNN Neural Network

CNNs are a much more effective at image classification as well shall see.

python3 mnist_cnn.py

After you run the second file in this Github as shown above you will see that this CNN is much more effective at minimizing our loss function. To summarize, this cnn code is more effective as it utilizes:

1. Local feature detection: CNNS use convolutional layers that apply filter sto small regions of the input, so they can detect local patterns more effectively
2. Weight Sharing: This CNN model uses the same filter (same weights) across the entire image so it reduces overall parameter number.
3. Spatial Hierarchy: CNNs preserve spatial structure of images, so they can recognize more complex patterns our Feedforward wouldn't
4. Pooling Layers: CNNS include pooling layers that reduce the spatial size of the input.
5. Also, we use data augmentation in the cnn code which artificially increases the size of the data by roatating the images, scaling images, etc.

CNN Neural Network with Visuals

Thus far, we only can see training epochs and percentages outputted to our command line. Some pictures would make things more interesting and informative to boot. Run the mnist_cnn_visuals.py code in the github.

python3 mnist_cnn_visuals.py

Now, after you finish running your cnn take a look at what files are in your directory.

ls

You should see different images you can open up and take a look at. I am using xdg for this but there are different ways to open it up.

xdg-open sample_predictions_epoch_2.png

You can now observe the MNIST dataset and see how your model does on classification.

Further Research

And that's it!

If you would like to continue to research this topic:

Pytorch documentation: https://pytorch.org/docs/stable/index.html

Paper on Document Recognition: http://vision.stanford.edu/cs598_spring07/papers/Lecun98.pdf

3Blue1Brown CNN video: https://www.youtube.com/watch?v=KuXjwB4LzSA