ShiroOnigami23/MyTorch-MNIST-Dataset

📊 MyTorch Refined MNIST Dataset

Curated by Aryan Singh Chandel (Shiro) at Rustamji Institute of Technology (RJIT).

This dataset contains the refined version of the MNIST handwritten digit database, specifically pre-processed for compatibility with the MyTorch deep learning framework.

🏗️ Dataset Structure

The data is stored in a compressed NumPy format (mnist_raw.npz) containing:

• X_train / y_train: 60,000 samples for training.
• X_test / y_test: 10,000 samples for final validation.

⚡ Pre-processing Specification

Every image in this dataset has undergone the following transformation logic used in the MyTorch 98.59% accuracy run:

1. Flattening: $28 \times 28$ spatial grids converted to $\mathbb{R}^{784}$ feature vectors.
2. Global Scaling: Pixel values $P \in [0, 255]$ normalized to $P' \in [0, 1]$ via $P' = P / 255.0$.
3. Refinement: Standardized tensor shapes for high-speed matrix multiplication in NumPy-based linear layers.

🚀 Usage

To load this dataset into your own MyTorch project:

import numpy as np
data = np.load('mnist_raw.npz')
X_train, y_train = data['X_train'], data['y_train']

🎓 Citation If you use this refined dataset in your research, please attribute: Chandel, A. S. (2026). MyTorch: Deep Learning from Scratch at RJIT.