MiniTorch

My Detailed Notes: minitorch | 十派的玩具箱

A compact deep learning systems project based on the MiniTorch teaching framework. This repository focuses on implementing core autodiff and tensor operators, then using them to train CNN models for sentiment classification and image classification.

Features

• Implemented MiniTorch core components such as autodiff, tensor operations, module/parameter management, and fast operators.
• Implemented neural network operators including Conv1D, Conv2D, pooling, dropout, and logsoftmax.
• Ran end-to-end training on:
  • SST-2 sentiment classification
  • MNIST digit classification
• Added modern training scripts for easier reproduction with PyTorch and Hugging Face.

Structure

minitorch/
|- minitorch/                  # framework core
|- project/                    # original MiniTorch training scripts
|  |- run_sentiment.py
|  |- run_mnist_multiclass.py
|  |- app.py
|  |- data/
|- scripts/                    # modern training scripts
|  |- train_sentiment_hf.py
|  |- train_mnist_torch.py
|- tests/
|- requirements.txt
|- requirements-modern.txt

Setup

conda create -n myminitorch python=3.11
conda activate myminitorch

pip install -r requirements.txt
pip install -e .

If you want to use the modern training scripts:

pip install -r requirements-modern.txt

Data

MNIST

Put the following files under project/data/:

• train-images-idx3-ubyte
• train-labels-idx1-ubyte

Example:

mkdir -p project/data
cd project/data

wget -c https://storage.googleapis.com/cvdf-datasets/mnist/train-images-idx3-ubyte.gz
wget -c https://storage.googleapis.com/cvdf-datasets/mnist/train-labels-idx1-ubyte.gz

gunzip -kf train-images-idx3-ubyte.gz
gunzip -kf train-labels-idx1-ubyte.gz
cd ../..

SST-2 and GloVe

For the original MiniTorch sentiment script, it is recommended to set:

export HF_HOME=/root/shared-nvme/minitorch/project/data/hf_cache
export EMBEDDINGS_ROOT=/root/shared-nvme/minitorch/project/data

This stores:

• SST-2 cache under project/data/hf_cache/
• GloVe files under project/data/glove/

The modern Hugging Face sentiment script only needs HF_HOME; it does not use GloVe.

Run

Original MiniTorch scripts

python project/run_sentiment.py | tee sentiment.txt
python project/run_mnist_multiclass.py | tee mnist.txt

Modern scripts

python scripts/train_sentiment_hf.py \
  --output-dir outputs/sst2-hf \
  --max-train-samples 2000 \
  --max-eval-samples 500

python scripts/train_mnist_torch.py \
  --data-dir project/data \
  --output-dir outputs/mnist-torch \
  --epochs 5

Visualization

streamlit run project/app.py

Resume Summary

Built and extended a MiniTorch-based deep learning systems project, implementing autodiff, tensor operators, and CNN modules, then validating the framework on SST-2 sentiment classification and MNIST image classification.