GenMnist

Mnist, but backward. Lol. Instead of classifying digits, the model is trained to generate images from an input digit between 1-9.

Data

Calling loader.load() should just work. First it will download the imgs.zip.

Model Architecture

The model is a dumb little MLP with these layers:

• Input Layer: One-hot encoded representation of the digit (dimension: 10).
• Hidden Layer: 128 neurons with ReLU activation.
• Output Layer: A flattened representation of the MNIST image (dimension: 784).

Usage

1. Instantiate the model:

model = Model()

2. Train the model:

model.train()

3. Generate images for a list of digits or one digit:

model.generate(list(range(10)))
model.generate(4)

Visualization

Generated images are visualized in a grid. Definitely room for improvement but not back for a ~2s training loop.