https://github.com/m-rishab/monotype-assignment
This project implements a Generative Adversarial Network (GAN) to generate images of cats. The GAN architecture consists of a Generator and a Discriminator, both implemented using PyTorch. The model is trained on a dataset of cat images, and the performance is monitored through the convergence of the training and validation loss.
https://github.com/m-rishab/monotype-assignment
Last synced: 8 months ago
JSON representation
This project implements a Generative Adversarial Network (GAN) to generate images of cats. The GAN architecture consists of a Generator and a Discriminator, both implemented using PyTorch. The model is trained on a dataset of cat images, and the performance is monitored through the convergence of the training and validation loss.
- Host: GitHub
- URL: https://github.com/m-rishab/monotype-assignment
- Owner: m-rishab
- Created: 2024-08-09T10:50:08.000Z (almost 2 years ago)
- Default Branch: main
- Last Pushed: 2024-08-09T11:00:06.000Z (almost 2 years ago)
- Last Synced: 2025-01-22T04:21:46.659Z (over 1 year ago)
- Language: Jupyter Notebook
- Size: 23.7 MB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Monotype-Assignment
This project implements a Generative Adversarial Network (GAN) to generate images of cats. The GAN architecture consists of a Generator and a Discriminator, both implemented using PyTorch. The model is trained on a dataset of cat images, and the performance is monitored through the convergence of the training and validation loss.
## Dataset
The model is trained on a dataset of cat images. You can use the [Cat Dataset from Kaggle](https://www.kaggle.com/datasets/spandan2/cats-faces-64x64-for-generative-models/data), or any other dataset of your choice.
The dataset should be structured in a way that the images are stored in folders, which is suitable for the `ImageFolder` class in PyTorch.
## Model Architecture
### Generator
The Generator is designed to take random noise as input and generate a 64x64 RGB image. The architecture includes several transposed convolutional layers, batch normalization, and ReLU activation.
### Discriminator
The Discriminator is a binary classifier that takes an image as input and outputs a probability that the image is real or fake. The architecture includes several convolutional layers, batch normalization, and LeakyReLU activation.
### Enhancements
- **Spectral Normalization**: Applied to the Discriminator to stabilize training.
- **Dropout**: Added to the Generator to prevent overfitting and improve the diversity of generated images.
- **Label Smoothing**: Used for the real labels to make the Discriminator less confident, helping the Generator learn more effectively.
## Training
The model is trained for a maximum of 10 epochs. The training process involves alternating updates to the Discriminator and Generator. Losses for both models are monitored and plotted to evaluate the performance.
## Generated Images
```python
# Key Training Parameters
batch_size = 128
image_size = 64
nz = 100 # Size of the latent z vector (input to generator)
ngf = 64 # Size of feature maps in generator
ndf = 64 # Size of feature maps in discriminator
num_epochs = 11
lr = 0.0002
beta1 = 0.5