Shushunya/mnist-cgan

MNIST Conditional DCGAN Generator

• Model Type: Conditional Generative Adversarial Network (cGAN) - Generator Only
• Task: Conditional Image Generation
• Dataset: MNIST (Grayscale handwritten digits, 0-9)
• Image Size: 32x32 pixels (upscaled from original 28x28)
• Framework: PyTorch

Architecture

The model is a Deep Convolutional Generator (DCGAN) modified for conditional generation.

• Inputs: A 100-dimensional random latent vector ($z$) and a class label (integer 0-9).
• Conditioning: The class label is passed through an nn.Embedding layer (10 classes $ ightarrow$ 10 dimensions) and concatenated with the latent noise vector.
• Blocks: 4 layers of nn.ConvTranspose2d with BatchNorm2d and ReLU activations.
• Output: A Tanh activation producing a 1x32x32 tensor with values normalized between [-1, 1].

Limitations & Biases

• Resolution: The output is strictly limited to 32x32 resolution.
• Domain: The model can only generate MNIST-style digits. It cannot generalize to alphabets, symbols, or real-world photographs.
• Artifacts: Standard Transposed Convolution artifacts (checkerboard patterns) may occasionally be visible upon close inspection.