def train(self):
tf.reset_default_graph()
is_training = tf.placeholder(dtype=tf.bool, name='is_training')
with tf.variable_scope('generator'):
z = tf.placeholder(dtype=tf.float32, shape=[None, self.z_dim])
#gen_data = archs.generator_conv(z, is_training)
gen_data = archs.generator_vanilla(z, self.z_dim, self.im_dim)
with tf.variable_scope('discriminator'):
#x = tf.placeholder(dtype=tf.float32, shape=[None, self.im_dim, self.im_dim], name='x')
x = tf.placeholder(\
tf.float32, shape=[None, self.im_dim*self.im_dim], name='inputs')
#d_real = archs.discriminator_conv(x, self.im_dim, reuse=False)
#d_fake = archs.discriminator_conv(gen_data, self.im_dim, reuse=True)
d_real = archs.discriminator_vanilla(x, self.im_dim, reuse=False)
d_fake = archs.discriminator_vanilla(gen_data, self.im_dim, reuse=True)
vars = tf.trainable_variables()
d_params = [v for v in vars if v.name.startswith('discriminator/')]
g_params = [v for v in vars if v.name.startswith('generator/')]
loss_d = tf.reduce_mean(d_real) - tf.reduce_mean(d_fake)
loss_g = - tf.reduce_mean(d_fake)
alpha = tf.random_uniform(
shape=[BATCH_SIZE,1],
minval=0.,
maxval=1.
)
differences = fake_data - real_data
interpolates = real_data + (alpha*differences)
gradients = tf.gradients(Discriminator(interpolates), [interpolates])[0]
slopes = tf.sqrt(tf.reduce_sum(tf.square(gradients), reduction_indices=[1]))
gradient_penalty = tf.reduce_mean((slopes-1.)**2)
disc_cost += LAMBDA*gradient_penalty
gen_train_op = tf.train.AdamOptimizer(
learning_rate=1e-4,
beta1=0.5,
beta2=0.9
).minimize(gen_cost, var_list=gen_params)
disc_train_op = tf.train.AdamOptimizer(
learning_rate=1e-4,
beta1=0.5,
beta2=0.9
).minimize(disc_cost, var_list=disc_params)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
d_optimizer = (tf.train.RMSPropOptimizer(learning_rate=5e-5)
.minimize(-loss_d, var_list=d_params))
g_optimizer = (tf.train.RMSPropOptimizer(learning_rate=5e-5)
.minimize(loss_g, var_list=g_params))
clip_d = [p.assign(tf.clip_by_value(p, -0.01, 0.01)) for p in d_params]
saver = tf.train.Saver()
with tf.Session() as sess:
try:
saver.restore(sess, "tmp/" + self.model_name + ".ckpt")
except:
print("Failed to restore")
init = tf.global_variables_initializer()
sess.run(init)
for i in range(self.num_iter):
#update discriminator multiple times
for n_critic in range(5):
train_data, _ = self.dataset.next()
#train_data = np.reshape(train_data, (-1, 28, 28))
loss_disc, _, _ = sess.run([loss_d, d_optimizer, clip_d], {
x: train_data,
z: sample_z(self.batch_size, self.z_dim),
is_training:True
})
#update generator
loss_gen, _ = sess.run([loss_g, g_optimizer], {
z: sample_z(self.batch_size, self.z_dim),
is_training:True
})
if i % self.steps["save"] == 0:
saver.save(sess, "tmp/" + self.model_name + ".ckpt")
if i % self.steps["sample"] == 0:
fake_data = sess.run(gen_data, {
z: sample_z(16, self.z_dim),
is_training: False
})
fig = plot_im(fake_data)
plt.savefig('out/' + self.model_name + '/{}.png'.format(str(i).zfill(3)), bbox_inches='tight')
plt.close(fig)
print(loss_disc, loss_gen)
def train(self):
tf.reset_default_graph()
is_training = tf.placeholder(dtype=tf.bool, name='is_training')
with tf.variable_scope('generator'):
z = tf.placeholder(dtype=tf.float32, shape=[None, self.z_dim])
gen_data = archs.generator_conv(z, is_training)
with tf.variable_scope('generator'):
x = tf.placeholder(dtype=tf.float32, shape=[None, 28, 28], name='x')
d_real = archs.discriminator_conv(x, self.im_dim, reuse=False)
d_fake = archs.discriminator_conv(gen_data, self.im_dim, reuse=True)
vars = tf.trainable_variables()
d_params = [v for v in vars if v.name.startswith('discriminator/')]
g_params = [v for v in vars if v.name.startswith('generator/')]
loss_d = - tf.reduce_mean(tf.log(d_real) + tf.log(1.0 - d_fake))
loss_g = - tf.reduce_mean(tf.log(d_fake))
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
d_optimizer = tf.train.RMSPropOptimizer(\
learning_rate=0.00015).minimize(loss_d, var_list=d_params)
g_optimizer = tf.train.RMSPropOptimizer(\
learning_rate=0.00015).minimize(loss_g, var_list=g_params)
saver = tf.train.Saver()
with tf.Session() as sess:
try:
saver.restore(sess, "tmp/" + self.model_name + ".ckpt")
except:
print("Failed to restore")
init = tf.global_variables_initializer()
sess.run(init)
for i in range(self.num_iter):
train_data, _ = self.dataset.next()
train_data = np.reshape(train_data, (-1, 28, 28))
#update discriminator
loss_disc, _ = sess.run([loss_d, d_optimizer], {
x: train_data,
z: sample_z(self.batch_size, self.z_dim),
is_training:True
})
#update generator
loss_gen, _ = sess.run([loss_g, g_optimizer], {
z: sample_z(self.batch_size, self.z_dim),
is_training:True
})
loss_gen, _ = sess.run([loss_g, g_optimizer], {
z: sample_z(self.batch_size, self.z_dim),
is_training:True
})
if i % self.steps["save"] == 0:
saver.save(sess, "tmp/" + self.model_name + ".ckpt")
if i % self.steps["sample"] == 0:
fake_data = sess.run(gen_data, {
z: sample_z(16, self.z_dim),
is_training: False
})
fig = plot_im(fake_data)
plt.savefig('out/' + self.model_name + '/{}.png'.format(str(i).zfill(3)), bbox_inches='tight')
plt.close(fig)
print(loss_disc, loss_gen)
def train(self):
tf.reset_default_graph() #clear graph
with tf.variable_scope('generator'):
z = tf.placeholder(tf.float32, shape=[None, self.z_dim], name='z')
gen_data = archs.generator_vanilla(z, self.z_dim, self.im_dim)
#gen_data = self.generator(z)
with tf.variable_scope('discriminator'):
x = tf.placeholder(\
tf.float32, shape=[None, self.im_dim*self.im_dim], name='inputs')
d_real = archs.discriminator_vanilla(x, self.im_dim, reuse=False)
d_fake = archs.discriminator_vanilla(gen_data, self.im_dim, reuse=True)
vars = tf.trainable_variables()
d_params = [v for v in vars if v.name.startswith('discriminator/')]
g_params = [v for v in vars if v.name.startswith('generator/')]
loss_d = - tf.reduce_mean(tf.log(d_real) + tf.log(1.0 - d_fake))
loss_g = - tf.reduce_mean(tf.log(d_fake))
d_optimizer = tf.train.AdamOptimizer().minimize(loss_d, var_list=d_params)
g_optimizer = tf.train.AdamOptimizer().minimize(loss_g, var_list=g_params)
saver = tf.train.Saver()
with tf.Session() as sess:
try:
saver.restore(sess, "tmp/" + self.model_name + ".ckpt")
except:
print("Failed to restore")
init = tf.global_variables_initializer()
sess.run(init)
for i in range(self.num_iter):
train_data, _ = self.dataset.next()
#update discriminator
loss_disc, _ = sess.run([loss_d, d_optimizer], {
x: train_data,
z: sample_z(self.batch_size, self.z_dim)
})
#update generator
loss_gen, _ = sess.run([loss_g, g_optimizer], {
z: sample_z(self.batch_size, self.z_dim)
})
loss_gen, _ = sess.run([loss_g, g_optimizer], {
z: sample_z(self.batch_size, self.z_dim)
})
if i % self.steps["save"] == 0:
saver.save(sess, "tmp/" + self.model_name + ".ckpt")
if i % self.steps["sample"] == 0:
fake_data = sess.run(gen_data, {
z: sample_z(16, self.z_dim)
})
fig = plot_im(fake_data)
plt.savefig('out/' + self.model_name + '/{}.png'.format(str(i).zfill(3)), bbox_inches='tight')
plt.close(fig)
print(loss_disc, loss_gen)