Generative models

Generative models

Using the Generative Adversarial Network

First generate the TFRecord files of the images to sample from. The n_samples parameter choses n_samples frames from the original dataset:

import gan

src = [["<some_path_to_your_data>/gfp.tif", "<some_path_to_your_data>/rfp.tif"]]
filename = "train_GAN.tfrecord"
gan.create_GAN_tfrecord(src, filename, n_samples=256)

Then, to train the network:

# set up a GAN configuration
config = gan.GAN2DConfiguration()
config.batch_size = 32
config.training_data = "train_GAN.tfrecord"

# set up the GAN
mode = tf.estimator.ModeKeys.TRAIN
GAN = gan.GenerativeAdverserialNetwork(config.to_params(), mode)
GAN.output_dir = "<some_path_to_your_output>/GAN/"

# train it
GAN.train()

There are lots of other parameters to play with. By default this creates the discriminator and generator networks on different GPUs. You may need to change that. Once training, the stdout looks like this:

Building new graph...
Building <function generator at 0x7fa6bd55cc08> on dev:/gpu:1
Building <function discriminator at 0x7fa6bd55ccf8> on dev:/gpu:0
Building <function discriminator at 0x7fa6bd55ccf8> on dev:/gpu:0
tensorflow/core/kernels/data/shuffle_dataset_op.cc:98] Filling up shuffle buffer (this may take a while): 339 of 1536
tensorflow/core/kernels/data/shuffle_dataset_op.cc:98] Filling up shuffle buffer (this may take a while): 728 of 1536
tensorflow/core/kernels/data/shuffle_dataset_op.cc:98] Filling up shuffle buffer (this may take a while): 1112 of 1536
tensorflow/core/kernels/data/shuffle_dataset_op.cc:98] Filling up shuffle buffer (this may take a while): 1505 of 1536
tensorflow/core/kernels/data/shuffle_dataset_op.cc:136] Shuffle buffer filled.
Alpha: 0.00, Exp: 0, Sz: (4, 4), Step: 0/ 6720, D_loss: 0.69923, G_loss: 0.91889, Model size: 14
Alpha: 0.02, Exp: 0, Sz: (4, 4), Step: 10/ 6720, D_loss: 0.50238, G_loss: 1.52228, Model size: 14
Alpha: 0.04, Exp: 0, Sz: (4, 4), Step: 20/ 6720, D_loss: 0.51882, G_loss: 0.94966, Model size: 14
Alpha: 0.06, Exp: 0, Sz: (4, 4), Step: 30/ 6720, D_loss: 0.38884, G_loss: 2.19388, Model size: 14
Alpha: 0.08, Exp: 0, Sz: (4, 4), Step: 40/ 6720, D_loss: 0.47636, G_loss: 2.56428, Model size: 14
Alpha: 0.10, Exp: 0, Sz: (4, 4), Step: 50/ 6720, D_loss: 0.47226, G_loss: 2.78979, Model size: 14
Alpha: 0.13, Exp: 0, Sz: (4, 4), Step: 60/ 6720, D_loss: 0.24605, G_loss: 2.47885, Model size: 14
...