README.md

---
language: 
- en
tags:
- pytorch
- mnist
- neural-network
license: apache-2.0
datasets:
- mnist
---

# Model Card for MyNeuralNet

## Model Description

`MyNeuralNet` is a simple, fully connected neural network designed for classifying the handwritten digits of the MNIST dataset. The model consists of three linear layers with ReLU activation functions, followed by a final layer with a softmax output to predict probabilities across the 10 possible digits (0-9).

## How the Model Was Trained

The model was trained using the MNIST dataset, which consists of 60,000 training images and 10,000 test images. Each image is a 28x28 grayscale representation of a handwritten digit. Training was conducted over 32 epochs with a batch size of 32. The SGD optimizer was used with a learning rate of 0.01.

### Training Script

The model training was carried out using a custom PyTorch script, similar to the following pseudocode:

```python
for epoch in range(n_epochs):
    for images, labels in dataloader:
        # Forward pass
        predictions = model(images)
        loss = loss_function(predictions, labels)
        
        # Backward pass and optimization
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
```

## Using the Model

Below is a simple example of how to load `MyNeuralNet` and use it to predict MNIST images:

```python
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
from huggingface_hub import hf_hub_download


# Ensure the device selection logic is centralized
def get_device():
    return torch.device("cuda" if torch.cuda.is_available() else "cpu")


# Define the neural network architecture
class MyNeuralNet(nn.Module):
    def __init__(self):
        super(MyNeuralNet, self).__init__()
        self.Matrix1 = nn.Linear(28 * 28, 100)
        self.Matrix2 = nn.Linear(100, 50)
        self.Matrix3 = nn.Linear(50, 10)
        self.R = nn.ReLU()

    def forward(self, x):
        x = x.view(-1, 28 * 28)
        x = self.R(self.Matrix1(x))
        x = self.R(self.Matrix2(x))
        x = self.Matrix3(x)
        return x.squeeze()


# Define the custom dataset class
class CTDataset(Dataset):
    def __init__(self, filepath, device):
        # Add 'device' as a parameter to the class constructor
        x, y = torch.load(filepath)
        self.x = x.float().div(255).to(device)  # Use the passed 'device' for tensor operations
        self.y = F.one_hot(y, num_classes=10).float().to(device)

    def __len__(self):
        return self.x.shape[0]

    def __getitem__(self, ix):
        return self.x[ix], self.y[ix]


def load_model():
    device = get_device()
    model_state_dict = torch.load(hf_hub_download(repo_id="Svenni551/may-mnist-digits", filename="model.pth"),
                                  map_location=torch.device(device))
    model = MyNeuralNet().to(device)
    model.load_state_dict(model_state_dict)
    model.eval()
    return model


def predict(input_data):
    device = get_device()
    model = load_model()
    if isinstance(input_data, str):  # Assuming filepath to dataset
        dataset = CTDataset(input_data, device)  # Pass 'device' as an argument
        loader = DataLoader(dataset, batch_size=32, shuffle=False)
        predictions = []
        with torch.no_grad():
            for batch, _ in loader:
                yhat = model(batch).argmax(axis=1).cpu().numpy()
                predictions.extend(yhat)
        return predictions
    elif isinstance(input_data, torch.Tensor):
        if len(input_data.shape) == 3:  # Single image
            input_data = input_data.unsqueeze(0)  # Add batch dimension
        input_data = input_data.to(device)
        with torch.no_grad():
            prediction = model(input_data).argmax(axis=1).item()
        return prediction
    else:
        raise ValueError("Unsupported input type. Provide a file path to a dataset or a PyTorch Tensor.")

# Example usage:
# prediction = predict('path/to/your/dataset.pt')
# or for an image:
# prediction = predict(your_image_tensor)

# print(prediction)
```

## Performance

<div style="display: flex; justify-content: space-between;">
    <img src="img/output.png" alt="Training Plot 1" style="width: 45%;">
    <img src="img/output2.png" alt="Training Plot 2" style="width: 45%;">
</div>

## Testing Model

<div style="display: grid; grid-template-rows: auto auto; grid-template-columns: 50% 50%; gap: 20px;">
    <img src="img/myplot1.png" alt="Training Plot 1" style="width: 100%;">
    <div style="display: grid; grid-template-columns: 50% 50%;">
        <img src="img/myplot2.png" alt="Training Plot 2" style="width: 100%;">
        <img src="img/myplot3.png" alt="Training Plot 3" style="width: 100%;">
    </div>
</div>

## Limitations and Ethics

This model was solely trained on the MNIST dataset and is optimized only for recognizing handwritten digits. Its application in other contexts has not been tested and might lead to inaccurate results.

## License

The MyNeuralNet model is made available under the Apache-2.0 license. For more details, please refer to the LICENSE file in the repository.