This project implements a Conditional Generative Adversarial Network (cGAN). Unlike standard GANs, this model doesn't just "paint"βit listens to "prompts" (labels). By training on the TF Flowers dataset, the system learned to bridge the gap between random Gaussian noise and botanical reality.
Evolution from Epoch 1 (Noise) to Epoch 199 (Flowers)
On Day 4, I conducted a deep dive into the architecture's efficiency.
| Metric | Result | Analysis |
| Training Cycles | 200 Epochs | Stable convergence achieved. |
| Peak Memory Usage | ~ 414.48 MB | Efficient convolutional architecture during training. |
| Optimization | Adam Optimizer | Learning rate |
Using Seed Arithmetic, I explored the "manifold" between two random points in the model's imagination.
Moving from Seed A to Seed B within the same class allows us to see how the model interprets features like petal density and light.
[zA * (1 - Ξ±) + zB * Ξ±]
Every great artist has a "Blue Period." Mine was the Pink Sunflower Period.
Leakage Detected: During Epoch 160+, the model began to "leak" colors across classes.
- Symptom: Sunflowers sprouted rose-pink petals.
- Lesson: Localized features are difficult to disentangle when the dataset shares high color-variance.
- Artifact: Noticed the "Chequerboard Pattern" due to Transposed Convolutions.