This repository contains code to evaluate performance of GANs on several metrics namely:
- Inception Score (Jupyter Notebook: src/StandardMetricsEngine.ipynb)
- FID Score (Jupyter Notebook: src/StandardMetricsEngine.ipynb)
- GAN-Train (Jupyter Notebook: src/GAN-train.ipynb)
- GAN-Test (Jupyter Notebook: src/GAN-test.ipynb)
These scores were computed on a fake dataset based off of CIFAR10 generated by either a DCGAN network or AGAN network. The DCGAN_CIFAR10 produces 10K fake images (Jupyter Notebook: src/DCGAN_CIFAR10.ipynb)
The following table shows these metrics calculated for GAN models with varying hyperparameters
| Exp | Discriminator LR | Generator LR | Batch Size | Iterations | Discriminator-Generator Ratio | Inception Score | FID Score | GAN train Score | GAN test score |
|---|---|---|---|---|---|---|---|---|---|
| 1 | 1.00E-02 | 1.00E-02 | 128 | 10,200 | 1 | 1.54 | 15.646 | 0.14120 | 0.11070 |
| 2 | 1.00E-04 | 1.00E-06 | 128 | 5,100 | 3 | 1.55 | 22.413 | 0.13250 | 0.10610 |
| 3 | 1.00E-03 | 1.00E-04 | 128 | 12,200 | 3 | 2.277 | 11.03 | 0.11150 | 0.12580 |
| 4 | 1.00E-04 | 1.00E-04 | 128 | 12,000 | 5 | 2.461 | 10.58 | 0.14700 | 0.13600 |
| 5 | 1.00E-04 | 1.00E-04 | 256 | 6,000 | 5 -> 2(last 3000 iters) | 1.988 | 12.046 | 0.11000 | 0.10260 |
| 6 | 1.00E-04 | 1.00E-04 | 64 | 30000 | 5 -> 2 (last 10000 iters) | 3.12 | 9.732 | 0.12000 | 0.10000 |
| 7 | 1.00E-04 | 1.00E-04 | 128 | 100000 | 5 -> 4 (5000 iters) -> 3 (10000 iters) -> 2 (20000 iters)" | 3.84 | 7.772 | 0.15000 | 0.10000 |
| 8 (ACGAN) | 1.00E-04 | 2.00E-04 | 100 | 500 epochs | 1 | 2.26 | 11.593 | 0.21000 | 0.14000 |