- TensorFlow implementation of Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks (ICLR 2016).
- Both generator and discriminator are multiple layers CNNs. Generator takes a random vector as input. Discriminator takes real and fake images as input. Feature maps are up-sampled and down-sampled through convolutional layers with stride = 2, so there is no pooling layer.
- DCGAN model is defined in
/src/nets/dcgan.py. An example to show how to train and test the model is defined inexamples/gans.py. - Random input vector are uniformly sampled within [-1, 1].
- As mentioned in the paper, batch normalization is used for all layers of generator and discriminator except the first and output layer of discriminator and the last layer of generator.
- ReLU is used in the generator except for the output, which uses Tanh. LeakyReLU with Leaky = 0.2 is used in the discriminators.
- All weights are initialized from a zero-centered Normal distribution with standard deviation 0.02. Learning rate is set to be 2e-4 and Adam optimizer with beta1 = 0.5 is used for optimization.
- When applied on MNIST, Dropout with 0.5 is used both for training and testing phase after each convolutional layer of the generator except the output layer. Because I found this reduced the noise on the generated images.
- see here.
-
vector length = 100, images size 28 x 28, generator dropout = 0.5 for both training and testing.
-
If dropout is removed, you can see several noise pixels around digits as shown below.
- The generated images get much better after adding dropout.
| Epoch 1 | Epoch 7 | Epoch 14 | Epoch 21 |
|---|---|---|---|
 |
 |
 |
 |
- Interpolation between two digits
- Images generated by uniformly sampling along x = [-1, 1] and y = [-1, 1] when input vector length is 2.
- vector length = 100, images are rescaled to 64 x 64, no dropout
| Epoch 1 | Epoch 10 | Epoch 20 | Epoch 25 |
|---|---|---|---|
 |
 |
 |
 |
- More result at epoch 25
- Interpolation between two faces
