Getting Started with MNIST using JAX, NNX, and Grain

MNIST Dataset Preparation

• 📥 Download the MNIST dataset from HuggingFace, storing it in parquet format.
• 🐼 Use the pandas library to read the training and testing data files into dataframes.
• 🖼️ Implement a Dataset class using grain to handle data loading, extracting image bytes, and converting them into normalized NumPy arrays.

CNN Model Definition with NNX

• 🏗️ Define a Convolutional Neural Network (CNN) model using NNX, featuring two convolutional layers, average pooling, ReLU activation, and two linear layers.
• 🔍 Instantiate the CNN model and use nnx.display to visualize its architecture.
• 🧪 Pass a dummy input to the model to verify its structure and output before training.

Optimizer and Metrics Setup

• ⚙️ Configure the AdamW optimizer from Optx, utilizing the nnx.param argument to explicitly target learnable parameters.
• 📊 Visualize the optimizer setup using nnx.display.
• 📈 Define training and evaluation step functions, including a cross-entropy loss function and gradient computation.

Training and Evaluation Loop

• 🚀 Decorate training and evaluation functions with nnx.jit for significant performance improvements.
• 🔁 Implement the main training loop, iterating through the dataset, calling the train_step function, and computing metrics at intervals.
• 🎯 Execute the eval_step on the test dataset to record test metrics and reset them.

Model Prediction and Visualization

• 📉 Switch the model to evaluation mode using model.eval().
• 🔮 Define a predict_step function to obtain predicted labels for test images.
• 📊 Plot a grid of test images with their corresponding predicted labels to visually assess model performance.