Deep Convolutional Generative Adversarial Networks (DCGAN)

Implemented a Deep Convolutional Generative Adversarial Network (DCGAN) using Python, TensorFlow, and Keras in a team of 4 for realistic image generation from the Fashion MNIST dataset. Achieved visually similar fashion item images after 50 epochs of training, with a comprehensive overview of the architecture, training process, and outcomes. Demonstrated applications of GANs in computer vision, serving as a foundational project for further exploration. Technologies include Python, TensorFlow, Keras, and Numpy.

---

Training

• Training Parameters:
  • Trained for 50 epochs.
  • Batch size of 128.
  • Learning rate set to 2e-4.
  • Generator input: Random noise vectors of 100 numbers, sampled from a uniform distribution between -1 to 1.
  • Generator uses deconvolution layers for upsampling.
• Evaluation Metrics
  • Metric: Accuracy.
  • Use: Binary classification accuracy for generated images.
  • Note: Both classes (real and generated) are equally important.
  • Loss Function: Binary Cross-Entropy (Log Loss).
  • Game of Minimax:

---

Outcomes

Graph of the DCGAN loss vs. epoch