def call_gan_training(generator, adversary):
gan_model = VanillaGAN(
generator=generator,
adversary=adversary,
z_dim=z_dim,
x_dim=im_dim,
folder="./veganModels/GAN",
)
# gan_model.summary(save=True)
gan_model.fit(
X_train=X_train,
X_test=X_test,
batch_size=batch_size,
epochs=epochs,
steps={"Adversary": 5},
print_every="0.5e",
save_model_every=None,
save_images_every="0.5e",
save_losses_every=1,
enable_tensorboard=True,
)
samples, losses = gan_model.get_training_results(by_epoch=False)
utils.plot_images(images=samples.reshape(-1, *samples.shape[2:]),
show=False)
utils.plot_losses(losses=losses, show=True)
def _save_losses_plot(self):
""" Creates the `losses.png` plot in the `self.folder` path.
"""
if hasattr(self, "logged_losses"):
fig, axs = plot_losses(self.logged_losses, show=False, share=False)
plt.savefig(os.path.join(self.folder, "losses.png"))
plt.close()
gan_model.summary(save=True)
gan_model.fit(X_train=X_train,
y_train=y_train,
X_test=X_test,
y_test=y_test,
batch_size=batch_size,
epochs=epochs,
steps=None,
print_every="0.2e",
save_model_every=None,
save_images_every="0.2e",
save_losses_every=10,
enable_tensorboard=False)
samples, losses = gan_model.get_training_results(by_epoch=False)
training_time = np.round(gan_model.total_training_time / 60, 2)
title = "Epochs: {}, z_dim: {}, Time trained: {} minutes\nParams: {}\n\n".format(
epochs, z_dim, training_time, gan_model.get_number_params())
fixed_labels = np.argmax(gan_model.get_fixed_labels(), axis=1)
fig, axs = utils.plot_images(images=samples.reshape(-1, 32, 32),
labels=fixed_labels,
show=False)
fig.suptitle(title, fontsize=12)
fig.tight_layout()
plt.savefig(gan_model.folder + "/generated_images.png")
fig, axs = utils.plot_losses(losses=losses, show=False)
fig.suptitle(title, fontsize=12)
fig.tight_layout()
plt.savefig(gan_model.folder + "/losses.png")
# gan_model.save()
discriminator = loader.load_adversary(x_dim=x_dim, y_dim=None)
gan = VanillaGAN(generator=generator,
adversary=discriminator,
z_dim=z_dim,
x_dim=x_dim,
folder=None)
gan.summary() # optional, shows architecture
# Training
gan.fit(X_train[:300], enable_tensorboard=False)
# Vizualise results
images, losses = gan.get_training_results()
utils.plot_images(images)
utils.plot_losses(losses)
# Sample new images, you can also pass a specific noise vector
samples = gan.generate(n=36)
utils.plot_images(samples)
elif mode == "supervised":
import torch
import numpy as np
import vegans.utils as utils
import vegans.utils.loading as loading
from vegans.GAN import ConditionalVanillaGAN
# Data preparation (Load your own data or example MNIST)
loader = loading.MNISTLoader()
X_train, y_train, X_test, y_test = loader.load()