Developing a DCGAN Model in Tensorflow 2.
0Mouhamed NdoyeBlockedUnblockFollowFollowingMar 15IntroductionIn early March 2019, TensorFlow 2.
0 was released and we decided to create an image generator based on Taehoon Kim’s implementation of DCGAN.
Here’s a tutorial on how to develop a DCGAN model in TensorFlow 2.
0.
“To avoid the fast convergence of D (discriminator) network, G (generator) network is updated twice for each D network update, which differs from original paper.
”— Taehoon KimPrerequisitesJupyter NotebookTensorFlow 2.
0Access to high-performing GPUDCGAN ArchitectureThe image below illustrates the generator referenced in the DCGAN paper.
Essentially, this network takes in a 100×1 noise vector, labeled 100z, and maps it into the G(Z) output which is 64x64x3.
100×1 → 1024x4x4 → 512x8x8 → 256x16x16 → 128x32x32 → 64x64x3The first layer expands the random noise by projecting and reshaping at each stepSource: carpedm20StridesThe stride specifies the ‘steps’ of the convolution along the height and width.
Here’s an animated example:An example of stridesInput Noisegenerator = make_generator_model()noise = tf.
random.
normal([1,100]) # shape is 1, 100generated_image = generator(noise, training = False)plt.
imshow(generated_image[0, :, :, 0], cmap =”gist_rainbow”)Dense LayerIn Keras, you can create layers to develop models.
A model is usually a network of layers, in which, the most common type is a stack of layersAdding a densely-connected layer to the model will take as input arrays of shape (*, 100).
The shape of the data will be (*, 4*4*1024) after the first layer.
In this case, you won’t need to specify the size of the input moving forward because of automatic shape inferenceBatch normalization functions similarly to preprocessing at every layer of the network.
ReLU is linear for all positive values and set to zero for all negative values.
Leaky ReLU has a smaller slope for negative values, instead of altogether zero.
For example, leaky ReLU may have y = 0.
01x when x < 0def make_generator_model(): model = tf.
keras.
Sequential() model.
add(layers.
Dense(4*4*1024, use_bias = False, input_shape = (100,))) model.
add(layers.
BatchNormalization()) model.
add(layers.
LeakyReLU())First LayerThe generator uses a transposed convolutional layer (upsampling) to produce an image from seed (random noise).
512 is the dimensionality of the output space(5,5) specifies the height and width of the 2D convolution windowStrides = (2,2)model.
add(layers.
Conv2DTranspose(512, (5, 5), strides = (2,2), padding = “same”, use_bias = False)) assert model.
output_shape == (None, 8, 8, 512)model.
add(layers.
BatchNormalization())model.
add(layers.
LeakyReLU())Second LayerThe generator uses a transposed convolutional layer (upsampling) to produce an image from the previous layer.
256 is the dimensionality of the output space(5,5) specifies the height and width of the 2D convolution windowStrides = (2,2)model.
add(layers.
Conv2DTranspose(256, (5,5), strides = (2,2), padding = “same”, use_bias = False))assert model.
output_shape == (None, 16, 16, 256)model.
add(layers.
BatchNormalization())model.
add(layers.
LeakyReLU())Third LayerThe generator uses a transposed convolutional layer (upsampling) to produce an image from the previous layer.
128 is the dimensionality of the output space(5,5) specifies the height and width of the 2D convolution windowStrides = (2,2)model.
add(layers.
Conv2DTranspose(128, (5,5), strides = (2,2), padding = “same”, use_bias = False))assert model.
output_shape == (None, 32, 32, 128)model.
add(layers.
BatchNormalization())model.
add(layers.
LeakyReLU())Last LayerThe generator uses a transposed convolutional layer (upsampling) to produce an image from the previous layer.
64 is the dimensionality of the output space(5,5) specifies the height and width of the 2D convolution windowStrides = (2,2) specifies the strides of the convolution along the height and widthmodel.
add(layers.
Conv2DTranspose(3, (5,5), strides = (2,2), padding = “same”, use_bias = False, activation = “tanh”))assert model.
output_shape == (None, 64, 64, 3) return modelCost FunctionThe generator’s loss quantifies how well it was able to trick the discriminator.
Intuitively, if the generator is performing well, the discriminator will classify the fake images as real (or 1).
Here, we will compare the discriminators decisions on the generated images to an array of 1s.
def generator_loss(fake_output): return cross_entropy(tf.
ones_like(fake_output), fake_output)OptimizerThe discriminator and the generator optimizers are different since we will train two networks, separately.
The Adam optimization algorithm is an extension of stochastic gradient descent.
Stochastic gradient descent maintains a single learning rate (termed alpha) for all weight updates while the learning rate does not change during training.
A learning rate is maintained for each network weight (parameter) and adapts as learning unfolds.
generator_optimizer = tf.
keras.
optimizers.
Adam(1e-4)discriminator_optimizer = tf.
keras.
optimizers.
Adam(1e-4)Result — MNISTTo start off, we trained our DCGAN model with 3 Deconvolutional layers on the MNIST dataset (28 x 28 grayscale images) which resulted in clearer renderings which can be seen below:Result — CelebAUsing a DCGAN model with 4 DC layers trained on a subset of the CelebA dataset (25,600 / 202,599 images), we were able to generate images resembling faces with a run of 100 epochs.
Tools/ResourcesPaperspaceTensorFlow 1.
5 — GPU — Py3 — P5000TensorFlow 2.
0 — GPU — Py3 — P5000TensorFlow 2.
0 — GPU — Py3 — Tesla V100MNIST Dataset60,000 images128 x 128CelebA Dataset202,599 imagesTransformed 218 x 178 to 64 x 64ChallengesUnderstanding GAN model created using TensorflowDeveloping DCGAN model using Tensorflow 2.
0Resizing/cropping the dataset images from 218 x 178 to 64 x 64Memory leakage in Jupyter Notebook while training the new modelIncompatibilities with TensorFlow 1.
3 and 2.
0Paperspace auto-shutdown after 12 hrsFuture WorkMigrate the model to an environment which can compute 100 epochs using the CelebA dataset in its entiretyTune the model architecture to generate better imagesDevelop a solution to resize images for more efficientlyskywall34/BabyGanImage Generator with DCGAN.
Contribute to skywall34/BabyGan development by creating an account on GitHub.
github.
comReferencesWe couldn’t have completed this project without the great people of the Internet.
Please check out the following resources for your next data science project:Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks — https://arxiv.
org/abs/1511.
06434Compressed Sensing using Generative Models — https://arxiv.
org/pdf/1703.
03208.
pdfImage Completion with Deep Learning in TensorFlow — http://bamos.
github.
io/2016/08/09/deep-completion/.