disc_real, real_logit = Discriminator(real_data)
disc_fake, fake_logit = Discriminator(fake_data)
gen_params = lib.params_with_name('Generator')
disc_params = lib.params_with_name('Discriminator')
class_loss_real = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(labels=real_label,
logits=real_logit))
class_loss_fake = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(labels=real_label,
logits=fake_logit))
#******************************************
bandwidths = [1.0, 2.0, 5.0, 10.0, 20.0, 40.0, 80.0]
kernel_cost = mmd.mix_rbf_mmd2(disc_real, disc_fake, sigmas=bandwidths)
ind_t = tf.placeholder(tf.int32, [11])
cum = tf.placeholder(tf.float32)
con_kernel_cost = 0
for i in range(10):
find_index = tf.where(tf.equal(real_label, i))
Image_c = tf.gather(disc_real, find_index)
Gimage_c = tf.gather(disc_fake, find_index)
Image_c_s = tf.reshape(Image_c, [-1, 1])
Gimage_c_s = tf.reshape(Gimage_c, [-1, 1])
con_kernel_cost += mmd.mix_rbf_mmd2(Image_c_s,
Gimage_c_s,
sigmas=bandwidths,
id=ind_t[i])
class_fake = tf.argmax(soft_class_fake, 1)
gen_params = lib.params_with_name('Generator')
disc_params = lib.params_with_name('Discriminator')
class_loss_real = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(labels=real_label,
logits=disc_real_logit))
class_loss_fake = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(labels=real_label,
logits=disc_fake_logit))
bandwidths = [1.0, 2.0, 5.0, 10.0, 20.0, 40.0, 80.0]
#bandwidths = [1.0, 2.0, 4.0, 8.0, 16.0]
kernel_cost = mmd.mix_rbf_mmd2(disc_real,
disc_fake,
sigmas=bandwidths,
id=BATCH_SIZE)
kernel_cost = tf.sqrt(kernel_cost)
ind_t = tf.placeholder(tf.int32, [11])
con_kernel_cost = 0
for i in range(10):
find_index = tf.where(tf.equal(real_label, i))
Image_c = tf.gather(disc_real, find_index)
Gimage_c = tf.gather(disc_fake, find_index)
Image_c_s = tf.reshape(Image_c, [-1, 1])
Gimage_c_s = tf.reshape(Gimage_c, [-1, 1])
con_kernel_cost += tf.sqrt(
mmd.mix_rbf_mmd2(Image_c_s, Gimage_c_s, sigmas=bandwidths,
id=ind_t[i]))
alpha = tf.random_uniform(shape=[BATCH_SIZE, 1], minval=0., maxval=1.)
return tf.reshape(output, [BATCH_SIZE, 1])
real_data_int = tf.placeholder(tf.int32, shape=[BATCH_SIZE, OUTPUT_DIM])
real_data = 2 * ((tf.cast(real_data_int, tf.float32) / 255.) - .5)
fake_data = Generator(BATCH_SIZE)
disc_real = Discriminator(real_data)
disc_fake = Discriminator(fake_data)
gen_params = lib.params_with_name('Generator')
disc_params = lib.params_with_name('Discriminator')
#******************************************
bandwidths = [2.0, 5.0, 10.0, 20.0, 40.0, 80.0]
kernel_cost = mmd.mix_rbf_mmd2(disc_real, disc_fake, sigmas=bandwidths)
gen_cost = kernel_cost
disc_cost = -1 * kernel_cost
# gen_cost = -tf.reduce_mean(disc_fake)
# disc_cost = tf.reduce_mean(disc_fake) - tf.reduce_mean(disc_real)
# Gradient penalty
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
