def run(self):
if self.method == 'train':
self.train()
elif self.method == 'build':
self.gan = hg.GAN(config=self.gan_config, inputs=self.create_input(blank=True))
if not self.gan.load(self.save_file):
raise ValidationException("Could not load model: "+ self.save_file)
self.build()
elif self.method == 'new':
self.new()
elif self.method == 'sample':
self.gan = hg.GAN(config=self.gan_config, inputs=self.create_input(blank=False))
if not self.gan.load(self.save_file):
print("Initializing new model")
self.sample_forever()
def setup_gan(config, inputs, args):
if "encode" in config:
print("CHARGAN")
gan = CharGAN(config, inputs=inputs)
else:
gan = hg.GAN(config, inputs=inputs)
gan.load(save_file)
return gan
def train(config, args):
title = "[hypergan] 2d-test " + config_filename
GlobalViewer.set_options(enabled=args.viewer, title=title, viewer_size=1)
print("ARGS", args)
gan = hg.GAN(config,
inputs=Custom2DInputDistribution(
{"batch_size": args.batch_size}))
gan.name = config_filename
if gan.config.use_latent:
accuracy_x_to_g = lambda: distribution_accuracy(
gan.inputs.next(1), gan.generator(gan.latent.next()))
accuracy_g_to_x = lambda: distribution_accuracy(
gan.generator(gan.latent.next()), gan.inputs.next(1))
else:
accuracy_x_to_g = lambda: distribution_accuracy(
gan.inputs.next(1), gan.generator(gan.inputs.next()))
accuracy_g_to_x = lambda: distribution_accuracy(
gan.generator(gan.inputs.next()), gan.inputs.next(1))
sampler = Custom2DSampler(gan)
gan.selected_sampler = sampler
samples = 0
steps = args.steps
sample_file = "samples/" + config_filename + "/000000.png"
os.makedirs(os.path.expanduser(os.path.dirname(sample_file)),
exist_ok=True)
sampler.sample(sample_file, args.save_samples)
metrics = [accuracy_x_to_g, accuracy_g_to_x]
sum_metrics = [0 for metric in metrics]
broken = False
for i in range(steps):
if broken:
break
gan.step()
if args.viewer and i % args.sample_every == 0:
samples += 1
print("Sampling " + str(samples))
sample_file = "samples/" + config_filename + "/%06d.png" % (
samples)
sampler.sample(sample_file, args.save_samples)
if i % 100 == 0:
for k, metric in enumerate(metrics):
_metric = metric().cpu().detach().numpy()
sum_metrics[k] += _metric
if not np.isfinite(_metric):
broken = True
break
return sum_metrics
def test_config(self):
with self.test_session():
config = {
'd_learn_rate': 1e-3,
'g_learn_rate': 1e-3,
'd_trainer': 'rmsprop',
'g_trainer': 'adam'
}
gan = hg.GAN()
trainer = AlternatingTrainer(gan, config)
self.assertEqual(trainer.config.d_learn_rate, 1e-3)
def train(config, args):
save_file = "save/chargan/model.ckpt"
with tf.device(args.device):
text_input = TextInput(config, args.batch_size, one_hot=one_hot)
gan = hg.GAN(config, inputs=text_input)
gan.create()
if(args.action != 'search' and os.path.isfile(save_file+".meta")):
gan.load(save_file)
with gan.session.as_default():
text_input.table.init.run()
tf.train.start_queue_runners(sess=gan.session)
s = [int(g) for g in gan.generator.sample.get_shape()]
x_0 = gan.session.run(gan.inputs.x)
z_0 = gan.session.run(gan.encoder.z)
ax_sum = 0
ag_sum = 0
diversity = 0.00001
dlog = 0
last_i = 0
samples = 0
vocabulary = text_input.get_vocabulary()
for i in range(args.steps):
gan.step()
if args.action == 'train' and i % args.save_every == 0 and i > 0:
print("saving " + save_file)
gan.save(save_file)
if i % args.sample_every == 0:
g, x_val = gan.session.run([gan.generator.sample, gan.inputs.x], {gan.encoder.z: z_0})
bs = np.shape(x_val)[0]
samples+=1
print("X: "+text_input.sample_output(x_val[0]))
print("G:")
for j, g0 in enumerate(g):
if j > 4:
break
print(text_input.sample_output(g0))
if args.config is None:
with open("sequence-results-10k.csv", "a") as myfile:
myfile.write(config_name+","+str(ax_sum)+","+str(ag_sum)+","+ str(ax_sum+ag_sum)+","+str(ax_sum*ag_sum)+","+str(dlog)+","+str(diversity)+","+str(ax_sum*ag_sum*(1/diversity))+","+str(last_i)+"\n")
tf.reset_default_graph()
gan.session.close()
def setup_gan(config, inputs, args):
gan = hg.GAN(config, inputs=inputs, name=args.config)
if (os.path.isfile(save_file + ".meta")):
gan.load(save_file)
tf.train.start_queue_runners(sess=gan.session)
config_name = args.config
GlobalViewer.title = "[hypergan] colorizer " + config_name
GlobalViewer.enabled = args.viewer
return gan
def train(config, args):
title = "[hypergan] 2d-test " + args.config
GlobalViewer.title = title
GlobalViewer.enabled = args.viewer
with tf.device(args.device):
config.generator['end_features'] = 2
config.generator["class"] = "class:__main__.Custom2DGenerator" # TODO
config.discriminator[
"class"] = "class:__main__.Custom2DDiscriminator" # TODO
gan = hg.GAN(config, inputs=Custom2DInputDistribution(args))
gan.name = config_name
accuracy_x_to_g = distribution_accuracy(gan.inputs.x,
gan.generator.sample)
accuracy_g_to_x = distribution_accuracy(gan.generator.sample,
gan.inputs.x)
sampler = Custom2DSampler(gan)
tf.train.start_queue_runners(sess=gan.session)
samples = 0
steps = args.steps
sampler.sample("samples/000000.png", args.save_samples)
metrics = [accuracy_x_to_g, accuracy_g_to_x]
sum_metrics = [0 for metric in metrics]
for i in range(steps):
gan.step()
if args.viewer and i % args.sample_every == 0:
samples += 1
print("Sampling " + str(samples), args.save_samples)
sample_file = "samples/%06d.png" % (samples)
sampler.sample(sample_file, args.save_samples)
if i > steps * 9.0 / 10:
for k, metric in enumerate(gan.session.run(metrics)):
sum_metrics[k] += metric
if i % 300 == 0:
for k, metric in enumerate(gan.metrics().keys()):
metric_value = gan.session.run(gan.metrics()[metric])
print("--", metric, metric_value)
if math.isnan(metric_value) or math.isinf(metric_value):
print("Breaking due to invalid metric")
return None
tf.reset_default_graph()
gan.session.close()
return sum_metrics
def setup_gan(config, inputs, args):
gan = hg.GAN(config, inputs=inputs)
gan.create()
if (args.action != 'search' and os.path.isfile(save_file + ".meta")):
gan.load(save_file)
with tf.device(args.device):
with gan.session.as_default():
inputs.table.init.run()
tf.train.start_queue_runners(sess=gan.session)
return gan
def setup_gan(config, inputs, args):
gan = hg.GAN(config=config, inputs=inputs)
gan.create()
if (args.action != 'search' and os.path.isfile(save_file + ".meta")):
gan.load(save_file)
tf.train.start_queue_runners(sess=gan.session)
config_name = args.config
title = "[hypergan] autoencode " + config_name
GlobalViewer.title = title
GlobalViewer.enabled = args.viewer
return gan
def test_create(self):
with self.test_session():
remove_d_config = hg.Configuration.default()
remove_d_config['discriminator'] = None
remove_d_config['loss'] = None
remove_d_config['trainer'] = None
gan = hg.GAN(config=remove_d_config, inputs=MockInput())
discriminator = AutoencoderDiscriminator(gan, config)
gan.encoder = gan.create_component(gan.config.encoder)
gan.encoder.create()
gan.generator = gan.create_component(gan.config.generator)
gan.generator.create()
net = discriminator.create()
gan.create()
self.assertEqual(int(net.get_shape()[1]), 32)
def mock_gan(batch_size=1, y=1, config=None):
mock_config = config or hc.Config({
"latent": {
"class":
"function:hypergan.distributions.uniform_distribution.UniformDistribution",
"max":
1,
"min":
-1,
"projections":
["function:hypergan.distributions.uniform_distribution.identity"],
"z":
128
},
"generator": {
"class":
"class:hypergan.discriminators.configurable_discriminator.ConfigurableDiscriminator",
"defaults": {
"activation": "tanh",
"initializer": "he_normal"
},
"layers": ["linear 32*32*1 activation=null"]
},
"discriminator": {
"class":
"class:hypergan.discriminators.configurable_discriminator.ConfigurableDiscriminator",
"defaults": {
"activation": "tanh",
"initializer": "he_normal"
},
"layers": ["linear 1 activation=null"]
},
"loss": {
"class": "function:hypergan.losses.ragan_loss.RaganLoss",
"reduce": "reduce_mean"
},
"trainer": {
"class":
"function:hypergan.trainers.alternating_trainer.AlternatingTrainer",
"optimizer": {
"class":
"function:tensorflow.python.training.adam.AdamOptimizer",
"learn_rate": 1e-4
}
}
})
return hg.GAN(config=mock_config,
inputs=MockInput(batch_size=batch_size, y=y))
def setup_gan(config, inputs, args):
gan = hg.GAN(config, inputs=inputs)
gan.create()
if (args.action != 'search' and os.path.isfile(save_file + ".meta")):
gan.load(save_file)
tf.train.start_queue_runners(sess=gan.session)
GlobalViewer.enable()
config_name = args.config
title = "[hypergan] colorizer " + config_name
GlobalViewer.window.set_title(title)
return gan
def train(config, args):
if (args.viewer):
title = "[hypergan] 2d-test " + args.config
GlobalViewer.title = title
GlobalViewer.enabled = args.viewer
with tf.device(args.device):
config.generator['end_features'] = 2
gan = hg.GAN(config, inputs=Custom2DInputDistribution(args))
gan.discriminator = Custom2DDiscriminator(gan, config.discriminator)
gan.generator = Custom2DGenerator(gan, config.generator)
gan.encoder = gan.create_component(gan.config.encoder)
gan.encoder.create()
gan.generator.create()
gan.discriminator.create()
gan.create()
accuracy_x_to_g = batch_accuracy(gan.inputs.x, gan.generator.sample)
accuracy_g_to_x = batch_accuracy(gan.generator.sample, gan.inputs.x)
sampler = Custom2DSampler(gan)
tf.train.start_queue_runners(sess=gan.session)
samples = 0
steps = args.steps
sampler.sample("samples/000000.png", args.save_samples)
metrics = [accuracy_x_to_g, accuracy_g_to_x]
sum_metrics = [0 for metric in metrics]
for i in range(steps):
gan.step()
if args.viewer and i % args.sample_every == 0:
samples += 1
print("Sampling " + str(samples), args.save_samples)
sample_file = "samples/%06d.png" % (samples)
sampler.sample(sample_file, args.save_samples)
if i > steps * 9.0 / 10:
for k, metric in enumerate(gan.session.run(metrics)):
sum_metrics[k] += metric
tf.reset_default_graph()
gan.session.close()
return sum_metrics
def train(self):
i=0
if(self.args.ipython):
import fcntl
fd = sys.stdin.fileno()
fl = fcntl.fcntl(fd, fcntl.F_GETFL)
fcntl.fcntl(fd, fcntl.F_SETFL, fl | os.O_NONBLOCK)
self.gan = hg.GAN(config=self.gan_config, inputs=self.create_input(), device=self.args.parameter_server_device)
self.gan.cli = self #TODO remove this link
self.gan.inputs.next()
self.lazy_create()
self.trainable_gan = hg.TrainableGAN(self.gan, save_file = self.save_file, devices = self.devices, backend_name = self.args.backend)
if self.trainable_gan.load():
print("Model loaded")
else:
print("Initializing new model")
self.trainable_gan.sample(self.sampler, self.sample_path)
while((self.steps < self.total_steps or self.total_steps == -1) and not self.gan.destroy):
self.step()
if self.should_sample:
self.should_sample = False
self.sample(False)
if (self.args.save_every != None and
self.args.save_every != -1 and
self.args.save_every > 0 and
self.steps % self.args.save_every == 0):
print(" |= Saving network")
self.trainable_gan.save()
self.create_path(self.advSavePath+'advSave.txt')
if os.path.isfile(self.advSavePath+'advSave.txt'):
with open(self.advSavePath+'advSave.txt', 'w') as the_file:
the_file.write(str(self.samples)+"\n")
if self.args.ipython:
self.check_stdin()
print("Done training model. Saving")
self.trainable_gan.save()
print("============================")
print("HyperGAN model trained")
print("============================")
def train(config, args):
title = "[hypergan] 2d-test " + args.config
GlobalViewer.title = title
GlobalViewer.enabled = args.viewer
with tf.device(args.device):
config.generator["class"]="class:__main__.Sequence2DGenerator"
config.discriminator["class"]="class:__main__.Sequence2DDiscriminator"
gan = hg.GAN(config, inputs = Sequence2DInputDistribution(args))
sampler = Sequence2DSampler(gan)
tf.train.start_queue_runners(sess=gan.session)
samples = 0
steps = args.steps
sampler.sample("samples/000000.png", args.save_samples)
#metrics = [accuracy_x_to_g, accuracy_g_to_x]
#sum_metrics = [0 for metric in metrics]
for i in range(steps):
gan.step({gan.inputs.current_step: i})
if args.viewer and i % args.sample_every == 0:
samples += 1
print("Sampling "+str(samples), args.save_samples)
sample_file="samples/%06d.png" % (samples)
sampler.sample(sample_file, args.save_samples)
#if i > steps * 9.0/10:
# for k, metric in enumerate(gan.session.run(metrics)):
# sum_metrics[k] += metric
#if i % 300 == 0:
# for k, metric in enumerate(gan.metrics.keys()):
# metric_value = gan.session.run(gan.metrics[metric])
# print("--", metric, metric_value)
# if math.isnan(metric_value) or math.isinf(metric_value):
# print("Breaking due to invalid metric")
# return None
tf.reset_default_graph()
gan.session.close()
return {}#sum_metrics
def optimize_g(g, d, config, initial_graph):
config['generator'] = g
config['discriminators'] = [d]
x, y, f, num_labels, examples_per_epoch = image_loader.labelled_image_tensors_from_directory(
args.directory,
config['batch_size'],
channels=channels,
format=args.format,
crop=args.crop,
width=width,
height=height)
initial_graph = {
'x': x,
'y': y,
'f': f,
'num_labels': num_labels,
'examples_per_epoch': examples_per_epoch
}
gan = hg.GAN(config, initial_graph)
return gan
def test_validate(self):
with self.assertRaises(ValidationException):
PyramidDiscriminator(hg.GAN(), {})
def test_config(self):
with self.test_session():
loss = SoftmaxLoss(hg.GAN(), loss_config)
self.assertTrue(loss.config.test)
def setup_gan(config, inputs, args):
gan = hg.GAN(config, inputs=inputs)
gan.load(save_file)
return gan
def test_config(self):
with self.test_session():
loss = LeastSquaresLoss(hg.GAN(), loss_config)
self.assertTrue(loss.config.test)
def train():
selector = hg.config.selector(args)
config_name = "2d-measure-accuracy-" + str(uuid.uuid4())
config = selector.random_config()
config_filename = os.path.expanduser('~/.hypergan/configs/' + config_name +
'.json')
trainers = []
rms_opts = {
'g_momentum': [0, 0.1, 0.01, 1e-6, 1e-5, 1e-1, 0.9, 0.999, 0.5],
'd_momentum': [0, 0.1, 0.01, 1e-6, 1e-5, 1e-1, 0.9, 0.999, 0.5],
'd_decay': [0.8, 0.9, 0.99, 0.999, 0.995, 0.9999, 1],
'g_decay': [0.8, 0.9, 0.99, 0.999, 0.995, 0.9999, 1],
'clipped_gradients': [False, 1e-2],
'clipped_d_weights': [False, 1e-2],
'd_learn_rate': [1e-3, 1e-4, 5e-4, 1e-6, 4e-4, 5e-5],
'g_learn_rate': [1e-3, 1e-4, 5e-4, 1e-6, 4e-4, 5e-5]
}
stable_rms_opts = {
"clipped_d_weights": 0.01,
"clipped_gradients": False,
"d_decay": 0.995,
"d_momentum": 1e-05,
"d_learn_rate": 0.001,
"g_decay": 0.995,
"g_momentum": 1e-06,
"g_learn_rate": 0.0005,
}
trainers.append(hg.trainers.rmsprop_trainer.config(**rms_opts))
adam_opts = {}
adam_opts = {
'd_learn_rate': [1e-3, 1e-4, 5e-4, 1e-2, 1e-6],
'g_learn_rate': [1e-3, 1e-4, 5e-4, 1e-2, 1e-6],
'd_beta1': [0.9, 0.99, 0.999, 0.1, 0.01, 0.2, 1e-8],
'd_beta2': [0.9, 0.99, 0.999, 0.1, 0.01, 0.2, 1e-8],
'g_beta1': [0.9, 0.99, 0.999, 0.1, 0.01, 0.2, 1e-8],
'g_beta2': [0.9, 0.99, 0.999, 0.1, 0.01, 0.2, 1e-8],
'd_epsilon': [1e-8, 1, 0.1, 0.5],
'g_epsilon': [1e-8, 1, 0.1, 0.5],
'd_clipped_weights': [False, 0.01],
'clipped_gradients': [False, 0.01]
}
trainers.append(hg.trainers.adam_trainer.config(**adam_opts))
sgd_opts = {
'd_learn_rate': [1e-3, 1e-4, 5e-4, 1e-2, 1e-6],
'g_learn_rate': [1e-3, 1e-4, 5e-4, 1e-2, 1e-6],
'd_clipped_weights': [False, 0.01],
'clipped_gradients': [False, 0.01]
}
trainers.append(hg.trainers.sgd_trainer.config(**sgd_opts))
encoders = []
projections = []
projections.append([
hg.encoders.uniform_encoder.modal, hg.encoders.uniform_encoder.identity
])
projections.append([
hg.encoders.uniform_encoder.modal, hg.encoders.uniform_encoder.sphere,
hg.encoders.uniform_encoder.identity
])
projections.append([
hg.encoders.uniform_encoder.binary, hg.encoders.uniform_encoder.sphere
])
projections.append([
hg.encoders.uniform_encoder.sphere,
hg.encoders.uniform_encoder.identity
])
projections.append([
hg.encoders.uniform_encoder.modal, hg.encoders.uniform_encoder.sphere
])
projections.append([
hg.encoders.uniform_encoder.sphere,
hg.encoders.uniform_encoder.identity,
hg.encoders.uniform_encoder.gaussian
])
encoder_opts = {
'z': [16],
'modes': [2, 4, 8, 16],
'projections': projections
}
stable_encoder_opts = {
"max":
1,
"min":
-1,
"modes":
8,
"projections": [[
"function:hypergan.encoders.uniform_encoder.modal",
"function:hypergan.encoders.uniform_encoder.sphere",
"function:hypergan.encoders.uniform_encoder.identity"
]],
"z":
16
}
losses = []
lamb_loss_opts = {
'reverse': [True, False],
'reduce': [
tf.reduce_mean, hg.losses.wgan_loss.linear_projection,
tf.reduce_sum, tf.reduce_logsumexp
],
'labels': [[-1, 1, 0], [0, 1, 1], [0, -1, -1], [1, -1, 0], [0, -1, 1],
[0, 1, -1], [0, 0.5, -0.5], [0.5, -0.5, 0], [0.5, 0, -0.5]],
'alpha': [
0, 1e-3, 1e-2, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.99,
0.999
],
'beta': [
0, 1e-3, 1e-2, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.99,
0.999
]
}
lsgan_loss_opts = {
'reduce': [
tf.reduce_mean, hg.losses.wgan_loss.linear_projection,
tf.reduce_sum, tf.reduce_logsumexp
],
'labels': [[-1, 1, 0], [0, 1, 1], [0, -1, -1], [1, -1, 0], [0, -1, 1],
[0, 1, -1], [0, 0.5, -0.5], [0.5, -0.5, 0], [0.5, 0, -0.5]]
}
stable_loss_opts = {
"alpha": 0.5,
"beta": [0.5, 0.8],
"discriminator": None,
"label_smooth": 0.26111111111111107,
"labels": [[0, -1, -1]],
"reduce": "function:tensorflow.python.ops.math_ops.reduce_mean",
"reverse": True
}
#losses.append([hg.losses.wgan_loss.config(**loss_opts)])
losses.append([hg.losses.lamb_gan_loss.config(**lamb_loss_opts)])
#losses.append([hg.losses.lamb_gan_loss.config(**stable_loss_opts)])
#losses.append([hg.losses.lamb_gan_loss.config(**stable_loss_opts)])
losses.append([hg.losses.lsgan_loss.config(**lsgan_loss_opts)])
#encoders.append([hg.encoders.uniform_encoder.config(**encoder_opts)])
encoders.append(
[hg.encoders.uniform_encoder.config(**stable_encoder_opts)])
custom_config = {
'model': args.config,
'batch_size': args.batch_size,
'trainer': trainers,
'generator': custom_generator_config(),
'discriminators': [[custom_discriminator_config()]],
'losses': losses,
'encoders': encoders
}
custom_config_selector = hc.Selector()
for key, value in custom_config.items():
custom_config_selector.set(key, value)
print("Set ", key, value)
custom_config_selection = custom_config_selector.random_config()
for key, value in custom_config_selection.items():
config[key] = value
config['dtype'] = tf.float32
config = hg.config.lookup_functions(config)
def circle(x):
spherenet = tf.square(x)
spherenet = tf.reduce_sum(spherenet, 1)
lam = tf.sqrt(spherenet)
return x / tf.reshape(lam, [int(lam.get_shape()[0]), 1])
def modes(x):
return tf.round(x * 2) / 2.0
if args.distribution == 'circle':
x = tf.random_normal([args.batch_size, 2])
x = circle(x)
elif args.distribution == 'modes':
x = tf.random_uniform([args.batch_size, 2], -1, 1)
x = modes(x)
elif args.distribution == 'sin':
x = tf.random_uniform((1, args.batch_size), -10.5, 10.5)
x = tf.transpose(x)
r_data = tf.random_normal((args.batch_size, 1), mean=0, stddev=0.1)
xy = tf.sin(0.75 * x) * 7.0 + x * 0.5 + r_data * 1.0
x = tf.concat([xy, x], 1) / 16.0
elif args.distribution == 'arch':
offset1 = tf.random_uniform((1, args.batch_size), -10, 10)
xa = tf.random_uniform((1, 1), 1, 4)
xb = tf.random_uniform((1, 1), 1, 4)
x1 = tf.random_uniform((1, args.batch_size), -1, 1)
xcos = tf.cos(x1 * np.pi + offset1) * xa
xsin = tf.sin(x1 * np.pi + offset1) * xb
x = tf.transpose(tf.concat([xcos, xsin], 0)) / 16.0
initial_graph = {
'x': x,
'num_labels': 1,
}
print("Starting training for: " + config_filename)
selector.save(config_filename, config)
with tf.device(args.device):
gan = hg.GAN(config, initial_graph)
accuracy_x_to_g = batch_accuracy(gan.graph.x, gan.graph.g[0])
accuracy_g_to_x = batch_accuracy(gan.graph.g[0], gan.graph.x)
s = [int(g) for g in gan.graph.g[0].get_shape()]
slice1 = tf.slice(gan.graph.g[0], [0, 0], [s[0] // 2, -1])
slice2 = tf.slice(gan.graph.g[0], [s[0] // 2, 0], [s[0] // 2, -1])
accuracy_g_to_g = batch_accuracy(slice1, slice2)
x_0 = gan.sess.run(gan.graph.x)
z_0 = gan.sess.run(gan.graph.z[0])
gan.initialize_graph()
ax_sum = 0
ag_sum = 0
diversity = 0.00001
dlog = 0
last_i = 0
tf.train.start_queue_runners(sess=gan.sess)
for i in range(500000):
d_loss, g_loss = gan.train()
if (np.abs(d_loss) > 100 or np.abs(g_loss) > 100):
ax_sum = ag_sum = 100000.00
break
if i % 1000 == 0 and i != 0:
ax, ag, agg, dl = gan.sess.run([
accuracy_x_to_g, accuracy_g_to_x, accuracy_g_to_g,
gan.graph.d_log
], {
gan.graph.x: x_0,
gan.graph.z[0]: z_0
})
print("ERROR", ax, ag)
if np.abs(ax) > 50.0 or np.abs(ag) > 50.0:
ax_sum = ag_sum = 100000.00
break
#if(i % 10000 == 0 and i != 0):
# g_vars = [var for var in tf.trainable_variables() if 'g_' in var.name]
# init = tf.initialize_variables(g_vars)
# gan.sess.run(init)
if (i > 490000):
ax, ag, agg, dl = gan.sess.run([
accuracy_x_to_g, accuracy_g_to_x, accuracy_g_to_g,
gan.graph.d_log
], {
gan.graph.x: x_0,
gan.graph.z[0]: z_0
})
diversity += agg
ax_sum += ax
ag_sum += ag
dlog = dl
with open("results.csv", "a") as myfile:
myfile.write(config_name + "," + str(ax_sum) + "," + str(ag_sum) +
"," + str(ax_sum + ag_sum) + "," +
str(ax_sum * ag_sum) + "," + str(dlog) + "," +
str(diversity) + "," + str(ax_sum * ag_sum *
(1 / diversity)) + "," +
str(last_i) + "\n")
tf.reset_default_graph()
gan.sess.close()
def setup_gan(config, inputs, args):
gan = hg.GAN(config, inputs=inputs, batch_size=args.batch_size)
gan.inputs.gradient_penalty_label = gan.inputs.feed_y # TODO: Our X dimensions dont always match the G. This causes gradient_penalty to fail.
gan.create()
return gan
def setup_gan(config, inputs, args):
gan = hg.GAN(config, inputs=inputs, batch_size=args.batch_size)
return gan
config['x_dims'] = [height, width]
config['channels'] = channels
config['model'] = 'colorizer'
config = hg.config.lookup_functions(config)
initial_graph = {
'x': x,
'y': y,
'f': f,
'num_labels': num_labels,
'examples_per_epoch': examples_per_epoch
}
with tf.device(args.device):
gan = hg.GAN(config,
initial_graph,
graph_type='generator',
device=args.device)
save_file = os.path.expanduser("~/.hypergan/saves/colorizer.ckpt")
gan.load_or_initialize_graph(save_file)
tf.train.start_queue_runners(sess=gan.sess)
build_file = os.path.expanduser(
"~/.hypergan/builds/colorizer/generator.ckpt")
saver = tf.train.Saver()
saver.save(gan.sess, build_file)
print("Saved generator to ", build_file)
tf.reset_default_graph()
self.sess.close()
def test_cli(self):
with self.test_session():
gan = hg.GAN()
args = {}
cli = hg.CLI(gan, args)
self.assertEqual(cli.gan, gan)
def mock_gan(batch_size=1, y=1):
return hg.GAN(inputs=MockInput(batch_size=batch_size, y=y))
height=height)
config['y_dims']=num_labels
config['x_dims']=[height,width]
config['channels']=channels
config = hg.config.lookup_functions(config)
initial_graph = {
'x':x,
'y':y,
'f':f,
'num_labels':num_labels,
'examples_per_epoch':examples_per_epoch
}
gan = hg.GAN(config, initial_graph)
save_file = os.path.expanduser("~/.hypergan/saves/super-resolution.ckpt")
gan.load_or_initialize_graph(save_file)
tf.train.start_queue_runners(sess=gan.sess)
for i in range(1000000):
d_loss, g_loss = gan.train()
if i % args.save_every == 0 and i > 0:
print("Saving " + save_file)
gan.save(save_file)
if i % args.sample_every == 0 and i > 0:
print("Sampling "+str(i))
sample_file = "samples/"+str(i)+".png"
import tensorflow as tf
import hyperchamber as hc
import numpy as np
from hypergan.gan_component import ValidationException
from hypergan.ops import TensorflowOps
from hypergan.gan_component import GANComponent
from hypergan.multi_component import MultiComponent
import hypergan as hg
from unittest.mock import MagicMock
gan = hg.GAN()
component = GANComponent(gan=gan, config={'test': True})
class GanComponentTest(tf.test.TestCase):
def test_config(self):
with self.test_session():
self.assertEqual(component.config.test, True)
def test_validate(self):
with self.test_session():
self.assertEqual(component.validate(), [])
def test_gan(self):
with self.test_session():
self.assertEqual(component.gan, gan)
def test_ops(self):
with self.test_session():
self.assertEqual(type(component.ops), TensorflowOps)
