class Twitch(object):
RING_BUFFER_SIZE_KEY = 'ringbuffer-size'
OAUTH_TOKEN_KEY = 'oauth_token'
LIVESTREAMER_PLUGIN_TWITCH = 'twitch'
def __init__(self, buffer_size, resolution, oauth, channel):
self.oauth = oauth
self.resolution = resolution
self.channel = channel
self.buffer_size = buffer_size
self.buffer = RingBuffer(buffer_size=buffer_size)
self.initialized = False
self.stream = None
def __del__(self):
if self.initialized:
self.stream.close()
def initialize(self):
self.buffer.clear()
stream = self._init_stream(self.oauth, self.channel)
if stream:
self.initialized = True
self.stream = stream.open()
def get_stream_data(self):
if not self.initialized:
print('Read: Try to initialize')
self.initialize()
raise StreamBufferIsEmptyException
return self.buffer.read_all()
def update_stream_data(self):
if self.initialized:
data = self.stream.read(self.buffer_size)
print('Update: {length}'.format(length=len(data)))
if len(data) != 0:
self.buffer.write(data)
else:
print('Update: Try to initialize')
self.initialize()
else:
print('Update: Try to initialize')
self.initialize()
def stream_initialized(self):
return self.stream is not None
def _init_stream(self, oauth, channel):
session = Livestreamer()
session.set_plugin_option(self.LIVESTREAMER_PLUGIN_TWITCH,
self.OAUTH_TOKEN_KEY, oauth)
session.set_option(self.RING_BUFFER_SIZE_KEY, self.buffer_size)
streams = session.streams(self._generate_stream_url(channel))
return streams.get(self.resolution)
@staticmethod
def _generate_stream_url(channel):
return 'https://www.twitch.tv/{channel}'.format(channel=channel)
def main(argv):
del argv
config = FLAGS
tf.set_random_seed(config.seed)
np_state = np.random.RandomState(config.seed)
global_step = tf.train.get_or_create_global_step()
global_step_update = tf.assign(global_step, global_step + 1)
real_ds = tf.data.TFRecordDataset(config.input_path)
real_ds = real_ds.map(lambda x: _parse_record(x, config.image_size))
real_ds = real_ds.shuffle(buffer_size=1000)
real_ds = real_ds.batch(config.batch_size // 2) # Half will be generated
real_ds = real_ds.repeat()
real_ds_iterator = real_ds.make_one_shot_iterator()
real_ds_example = real_ds_iterator.get_next()
discriminator = Discriminator('discriminator')
generator = Generator('generator')
z = tf.placeholder(dtype=tf.float32, shape=[None, 100])
G_sample = generator.create_main_graph(z)
D_logit_real = discriminator.create_main_graph(real_ds_example)
D_logit_fake = discriminator.create_main_graph(G_sample)
D_expected_real = tf.zeros_like(D_logit_real)
D_expected_fake = tf.ones_like(D_logit_fake)
D_loss_real = tf.losses.sigmoid_cross_entropy(D_expected_real,
D_logit_real,
label_smoothing=0.2)
D_loss_fake = tf.losses.sigmoid_cross_entropy(D_expected_fake,
D_logit_fake,
label_smoothing=0.00)
D_loss = 0.5 * (D_loss_real + D_loss_fake)
G_loss = tf.losses.sigmoid_cross_entropy(tf.zeros_like(D_logit_fake),
D_logit_fake,
label_smoothing=0.00)
with tf.variable_scope('metrics'):
D_prediction_real = tf.round(tf.nn.sigmoid(D_logit_real))
D_prediction_fake = tf.round(tf.nn.sigmoid(D_logit_fake))
D_accuracy_real = accuracy(D_prediction_real, D_expected_real)
D_accuracy_fake = accuracy(D_prediction_fake, D_expected_fake)
real_size = tf.to_float(tf.shape(D_prediction_real)[0])
fake_size = tf.to_float(tf.shape(D_prediction_fake)[0])
D_accuracy = (real_size * D_accuracy_real +
fake_size * D_accuracy_fake) / (real_size + fake_size)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS,
scope='discriminator')
with tf.control_dependencies(update_ops):
D_optimizer = tf.train.AdamOptimizer(
config.discriminator_learning_rate).minimize(
D_loss, var_list=discriminator.get_variables())
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope='generator')
with tf.control_dependencies(update_ops):
G_optimizer = tf.train.AdamOptimizer(
config.generator_learning_rate).minimize(
G_loss, var_list=generator.get_variables())
with tf.variable_scope('summaries'):
D_loss_summary = tf.summary.scalar('loss',
D_loss,
family='discriminator')
D_accuracy_real_summary = tf.summary.scalar('real_accuracy',
D_accuracy_real,
family='discriminator')
D_accuracy_fake_summary = tf.summary.scalar('fake_accuracy',
D_accuracy_fake,
family='discriminator')
D_accuracy_summary = tf.summary.scalar('accuracy',
D_accuracy,
family='discriminator')
G_loss_summary = tf.summary.scalar('loss', G_loss, family='generator')
G_image_summary = tf.summary.image('generation',
G_sample,
max_outputs=1,
family='generator')
Real_image_summary = tf.summary.image('real',
real_ds_example,
max_outputs=1)
summary_op = tf.summary.merge_all()
# Session
hooks = []
hooks.append(tf.train.StopAtStepHook(num_steps=config.iterations))
if (config.save_checkpoints):
hooks.append(
tf.train.CheckpointSaverHook(
checkpoint_dir=config.checkpoint_directory,
save_secs=config.checkpoint_save_secs,
save_steps=config.checkpoint_save_steps))
if (config.save_summaries):
hooks.append(
tf.train.SummarySaverHook(output_dir=config.summary_directory,
save_secs=config.summary_save_secs,
save_steps=config.summary_save_steps,
summary_op=summary_op))
if config.restore:
sess = tf.train.MonitoredTrainingSession(
checkpoint_dir=config.checkpoint_directory,
save_checkpoint_steps=None,
save_checkpoint_secs=None,
save_summaries_steps=None,
save_summaries_secs=None,
log_step_count_steps=None,
hooks=hooks)
else:
sess = tf.train.MonitoredTrainingSession(save_checkpoint_steps=None,
save_checkpoint_secs=None,
save_summaries_steps=None,
save_summaries_secs=None,
log_step_count_steps=None,
hooks=hooks)
def step_generator(step_context, accuracy_buffer):
np_global_step = step_context.session.run(global_step)
step_context.session.run(global_step_update)
random_noise = np_state.normal(size=[config.batch_size, 100])
_, np_loss, np_accuracy = step_context.run_with_hooks(
[G_optimizer, G_loss, D_accuracy], feed_dict={z: random_noise})
accuracy_buffer.add(np_accuracy)
if np_global_step % config.log_step == 0:
logging.debug(
'Training Generator: Step: {} Loss: {:.3e} Accuracy: {:.2f}'
.format(np_global_step, np_loss,
accuracy_buffer.mean() * 100))
def step_discriminator(step_context, accuracy_buffer):
np_global_step = step_context.session.run(global_step)
step_context.session.run(global_step_update)
random_noise = np_state.normal(size=[config.batch_size // 2, 100])
_, np_loss, np_accuracy = step_context.run_with_hooks(
[D_optimizer, D_loss, D_accuracy], feed_dict={z: random_noise})
accuracy_buffer.add(np_accuracy)
if np_global_step % config.log_step == 0:
logging.debug(
'Training Discriminator: Step: {} Loss Mean: {:.3e} Accuracy: {:.2f}'
.format(np_global_step, np_loss,
accuracy_buffer.mean() * 100))
accuracy_buffer = RingBuffer(config.buffer_size)
accuracy_buffer.clear()
while not sess.should_stop():
for _ in xrange(config.D_steps):
sess.run_step_fn(lambda step_context: step_discriminator(
step_context, accuracy_buffer))
for _ in xrange(config.G_steps):
sess.run_step_fn(lambda step_context: step_generator(
step_context, accuracy_buffer))
sess.close()
