MiniTorch Module 4: Neural Networks and Convolutions

• Docs: https://minitorch.github.io

• Overview: https://minitorch.github.io/module4/module4/

Module 4 extends MiniTorch with convolutional neural network operations to build an image recognition system. You'll implement a version of LeNet on MNIST for digit recognition and 1D convolution for NLP sentiment classification:

• 1D and 2D convolution for feature extraction and image processing
• Pooling operations for spatial down-sampling and dimension reduction
• Advanced NN functions including softmax, dropout, and max operations
• MNIST digit classification and SST2 sentiment analysis with CNNs

Quick Start

# Install dependencies
pip install -e ".[dev,extra]"

# Set up MNIST dataset
./get_mnist.sh

# Sync files from Module 3
python sync_previous_module.py ../Module-3 .

# Run Module 4 tests
pytest -m task4_1
pytest -m task4_2
pytest -m task4_3
pytest -m task4_4

# Train CNN models
python project/run_mnist_multiclass.py  # MNIST digits
python project/run_sentiment.py         # SST2 sentiment

# Run style checks
pre-commit run --all-files

Tasks Overview

Task 4.1: 1D Convolution

File: minitorch/fast_conv.py

• Implement _tensor_conv1d() function
• Support forward and backward convolution with parallel processing

Task 4.2: 2D Convolution

File: minitorch/fast_conv.py

• Implement _tensor_conv2d() function
• Optimize for image processing with efficient memory access

Task 4.3: Pooling Operations

File: minitorch/nn.py

• Implement tile() tensor reshaping function
• Implement avgpool2d() for average pooling

Task 4.4: Advanced Neural Network Functions

File: minitorch/nn.py

• Implement max(), argmax(), softmax(), logsoftmax(), and dropout()
• Implement maxpool2d() for max pooling operations
• Add property tests and ensure gradient computation correctness

Task 4.4b: Extra Credit (CUDA Convolution)

File: minitorch/cuda_conv.py (Create this file)

• Implement conv1d and conv2d on CUDA for efficient GPU processing
• Critical for large-scale image recognition performance
• Show output on Google Colab

Task 4.5: Training an Image Classifier

Files: project/run_sentiment.py and project/run_mnist_multiclass.py

• Implement Conv1D, Conv2D, and Network for both sentiment and image classification
• Train models on SST2 sentiment data and MNIST digit classification
• Use Streamlit visualization to view hidden states of your model

Training Requirements:

• Train a model on Sentiment (SST2), add training logs as sentiment.txt showing train loss, train accuracy and validation accuracy (should achieve >70% best validation accuracy)
• Train a model on Digit classification (MNIST), add logs as mnist.txt showing train loss and validation accuracy
• Implement Conv1D, Conv2D, and Network classes for both training files

Resources

• Installation Guide: Setup instructions with MNIST
• Testing Guide: CNN testing strategies