class Visualization(object):
"""Helper class to visualize the progress of the GAN training procedure.
"""
def __init__(self, file_name, model_name, fps=15):
"""Initialize the helper class.
:param fps: The number of frames per second when saving the gif animation.
"""
self.fps = fps
self.figure, (self.ax2) = plt.subplots(1, 1, figsize=(5, 5))
self.figure.suptitle("{}".format(model_name))
sns.set(color_codes=True, style='white', palette='colorblind')
sns.despine(self.figure)
plt.show(block=False)
self.real_data = Dataset()
self.step = 0
self.writer = ImageMagickWriter(fps=self.fps)
self.writer.setup(self.figure, file_name, dpi=100)
def plot_progress(self, gen_net):
"""Plot the progress of the training procedure. This can be called back from the GAN fit method.
:param gan: The GAN we are fitting.
:param session: The current session of the GAN.
:param data: The data object from which we are sampling the input data.
"""
real = self.real_data.next_batch(batch_size=10000)
r1, r2 = self.ax2.get_xlim()
x = np.linspace(r1, r2, 10000)[:, np.newaxis]
g = gen_net(torch.FloatTensor(np.random.randn(10000, 1)))
self.ax2.clear()
self.ax2.set_ylim([0, 1])
self.ax2.set_xlim([0, 8])
sns.kdeplot(real.numpy().flatten(),
shade=True,
ax=self.ax2,
label='Real data')
sns.kdeplot(g.detach().numpy().flatten(),
shade=True,
ax=self.ax2,
label='Generated data')
self.ax2.set_title('Distributions')
self.ax2.set_title('{} iterations'.format(self.step * 50))
self.ax2.set_xlabel('Input domain')
self.ax2.set_ylabel('Probability density')
self.ax2.legend(loc='upper left', frameon=True)
self.figure.canvas.draw()
self.figure.canvas.flush_events()
self.writer.grab_frame()
self.step += 1
class Visualization(object):
"""Helper class to visualize the progress of the GAN training procedure.
"""
def __init__(self, save_animation=False, fps=30):
"""Initialize the helper class.
:param save_animation: Whether the animation should be saved as a gif. Requires the ImageMagick library.
:param fps: The number of frames per second when saving the gif animation.
"""
self.save_animation = save_animation
self.fps = fps
self.figure, (self.ax1, self.ax2) = plt.subplots(1, 2, figsize=(8, 4))
self.figure.suptitle("1D GAAN")
sns.set(color_codes=True, style='white', palette='colorblind')
sns.despine(self.figure)
plt.show(block=False)
if self.save_animation:
self.writer = ImageMagickWriter(fps=self.fps)
self.writer.setup(self.figure, 'Toy_1D_GAAN.gif', dpi=100)
def plot_progress(self, gan, session, data):
"""Plot the progress of the training procedure. This can be called back from the GAN fit method.
:param gan: The GAN we are fitting.
:param session: The current session of the GAN.
:param data: The data object from which we are sampling the input data.
"""
# Plot the training curve.
steps = gan.log_interval * np.arange(len(gan.loss_d_curve))
self.ax1.clear()
self.ax1.plot(steps, gan.loss_d_curve, label='D loss')
self.ax1.plot(steps, gan.loss_g_curve, label='G loss')
self.ax1.set_title('Learning curve')
self.ax1.set_xlabel('Iteration')
if gan.model == 'gan' or gan.model == 'rgan' or gan.model == 'mdgan' or gan.model == 'vaegan' or gan.model == 'gaan':
title = 'Loss'
elif gan.model == 'wgangp':
title = 'Negative critic loss'
self.ax1.set_ylabel(title)
# Plot the generated and the input data distributions.
g = gan.sample(session)
r1, r2 = self.ax2.get_xlim()
x = np.linspace(r1, r2, gan.n_sample)[:, np.newaxis]
critic = gan.dreal(session, x)
if gan.model == 'wgangp':
# Normalize the critic to be in [0, 1] to make visualization easier.
critic = (critic - critic.min()) / (critic.max() - critic.min())
d, _ = data.next_batch(gan.n_sample)
if gan.model == 'gaan' or gan.model == 'rgan' or gan.model == 'mdgan':
r = gan.reconstruct(session, d)
e = gan.encode(session, d)
self.ax2.clear()
self.ax2.set_ylim([0, 1])
self.ax2.set_xlim([-8, 8])
if gan.model == 'gan' or gan.model == 'rgan' or gan.model == 'mdgan' or gan.model == 'gaan' or gan.model == 'vaegan':
self.ax2.plot(x, critic, label='Decision boundary')
elif gan.model == 'wgangp':
self.ax2.plot(x, critic, label='Critic (normalized)')
sns.kdeplot(d.flatten(), shade=True, ax=self.ax2, label='Real data')
sns.kdeplot(g.flatten(),
shade=True,
ax=self.ax2,
label='Generated data')
self.ax2.set_title('Distributions')
self.ax2.set_xlabel('Input domain')
self.ax2.set_ylabel('Probability density')
self.ax2.legend(loc='upper left', frameon=True)
if len(steps) - 1 == gan.n_step // gan.log_interval:
if self.save_animation:
wait_seconds = 3
[
self.writer.grab_frame()
for _ in range(wait_seconds * self.fps)
]
self.writer.finish()
plt.show()
else:
self.figure.canvas.draw()
self.figure.canvas.flush_events()
if self.save_animation:
self.writer.grab_frame()
class Visualization(object):
"""Helper class to visualize the progress of the GAN training procedure.
"""
def __init__(self, save_animation=False, fps=30):
"""Initialize the helper class.
:param save_animation: Whether the animation should be saved as a gif. Requires the ImageMagick library.
:param fps: The number of frames per second when saving the gif animation.
"""
self.save_animation = save_animation
self.fps = fps
self.figure, (self.ax1, self.ax2) = plt.subplots(1, 2, figsize=(8, 4))
self.figure.suptitle("1D GAAN")
sns.set(color_codes=True, style='white', palette='colorblind')
sns.despine(self.figure)
plt.show(block=False)
if self.save_animation:
self.writer = ImageMagickWriter(fps=self.fps)
self.writer.setup(self.figure, 'Toy_1D_GAAN.gif', dpi=100)
def plot_progress(self, gan, session, data):
"""Plot the progress of the training procedure. This can be called back from the GAN fit method.
:param gan: The GAN we are fitting.
:param session: The current session of the GAN.
:param data: The data object from which we are sampling the input data.
"""
# Plot the training curve.
steps = gan.log_interval * np.arange(len(gan.loss_d_curve))
self.ax1.clear()
self.ax1.plot(steps, gan.loss_d_curve, label='D loss')
self.ax1.plot(steps, gan.loss_g_curve, label='G loss')
self.ax1.set_title('Learning curve')
self.ax1.set_xlabel('Iteration')
if gan.model == 'gan' or gan.model == 'rgan' or gan.model == 'mdgan' or gan.model == 'vaegan' or gan.model == 'gaan':
title = 'Loss'
elif gan.model == 'wgangp':
title = 'Negative critic loss'
self.ax1.set_ylabel(title)
# Plot the generated and the input data distributions.
g = gan.sample(session)
r1,r2 = self.ax2.get_xlim()
x = np.linspace(r1, r2, gan.n_sample)[:, np.newaxis]
critic = gan.dreal(session, x)
if gan.model == 'wgangp':
# Normalize the critic to be in [0, 1] to make visualization easier.
critic = (critic - critic.min()) / (critic.max() - critic.min())
d, _ = data.next_batch(gan.n_sample)
if gan.model == 'gaan' or gan.model == 'rgan' or gan.model == 'mdgan':
r = gan.reconstruct(session, d)
e = gan.encode(session, d)
self.ax2.clear()
self.ax2.set_ylim([0, 1])
self.ax2.set_xlim([-8, 8])
if gan.model == 'gan' or gan.model == 'rgan' or gan.model == 'mdgan' or gan.model == 'gaan' or gan.model == 'vaegan':
self.ax2.plot(x, critic, label='Decision boundary')
elif gan.model == 'wgangp':
self.ax2.plot(x, critic, label='Critic (normalized)')
sns.kdeplot(d.flatten(), shade=True, ax=self.ax2, label='Real data')
sns.kdeplot(g.flatten(), shade=True, ax=self.ax2, label='Generated data')
self.ax2.set_title('Distributions')
self.ax2.set_xlabel('Input domain')
self.ax2.set_ylabel('Probability density')
self.ax2.legend(loc='upper left', frameon=True)
if len(steps) - 1 == gan.n_step // gan.log_interval:
if self.save_animation:
wait_seconds = 3
[self.writer.grab_frame() for _ in range(wait_seconds * self.fps)]
self.writer.finish()
plt.show()
else:
self.figure.canvas.draw()
self.figure.canvas.flush_events()
if self.save_animation:
self.writer.grab_frame()
# Load the model and an example
model = LevelSetMachineLearning.load('./LSML-model.pkl')
example = model.testing_data[13]
# Set up plotting for the movie frames
fig, ax = plt.subplots(1, 1, figsize=(2, 2))
ax.axis('off')
ax.imshow(example.img, cmap=plt.cm.gray, interpolation='bilinear')
lines = []
# Set up the movie writer and grab the frame at initialization
writer = ImageMagickWriter(fps=5)
writer.setup(fig, 'evolution.gif', 100)
writer.grab_frame()
# Define the callback function to be used during segmentation evolution
def update_movie(i, u):
if i % 10 != 0:
return
for line in lines:
line.remove()
lines.clear()
lines.extend(
plot_iso_contours(ax, u, value=0, c='b')