mnist-i-jepa

MNIST I-JEPA Classifier

• I-JEPA model architecture

Model Overview

• Model Name: MNIST I-JEPA Classifier
• Model Type: Convolutional Neural Network with I-JEPA feature extraction
• Dataset: MNIST (Modified National Institute of Standards and Technology)
  • A dataset of 70,000 handwritten digits (28x28 grayscale images) of digits 0-9.
• Task: Image classification (Digit recognition)
• Framework: PyTorch
• Preprocessing:
  • Images are resized to 28x28 pixels, converted to grayscale (if necessary), and normalized before feeding into the model.

Model Architecture

• Input Layer: 28x28 grayscale images.
• Convolutional Layers:
  • Conv1: 32 filters of size 3x3, applied to the input image (1 channel).
  • Conv2: 64 filters of size 3x3, applied to the output of Conv1.
• Activation Functions:
  • ReLU activations after each convolutional layer.
• Pooling Layers:
  • Max pooling with a 2x2 window after Conv1 and Conv2 to downsample.
• Fully Connected Layers:
  • Flattened output from the convolutional layers and passed to a fully connected layer with feature_dim (default 128) neurons.
  • Final fully connected layer outputs 10 units, corresponding to the 10 possible digit classes (0-9).
• Output:
  • Softmax activation is used on the final output layer to produce class probabilities.

Training Details

• Epochs: 5
• Batch Size: (Standard MNIST batch size, 64-128 recommended)
• Optimizer: Adam optimizer with learning rate 0.001
• Loss Function: Cross-Entropy Loss (for classification)
• Hardware Used: (Google Colab T4 GPU)

Performance Metrics

• Training Accuracy: (99% on MNIST)
• Training Loss: (0.0310)

Usage Instructions

• How to Use:
  1. Clone the repository.
  2. Dependencies:

     pip install torch torchvision matplotlib pillow
     

  3. cd demo

     python demo.py
     

Model Limitations

• This model may not perform well on images that differ greatly from the MNIST dataset, such as noisy or distorted digits.
• The model is optimized for recognizing digits from the MNIST dataset and may not generalize well to other types of handwritten digits or more complex images.