def build_image_generator(self, img_z, sen_rep):
with tf.variable_scope('img_generator'):
# now, calculate the size of output during the deconv upsampling
# note that we only use stride 2 during the conv
assert self.config.generator_l1_nchannel % 8 == 0, \
logger.error('[ERROR] Invalid channel size')
l5_h, l5_w, l5_c = 64, 64, 3
l4_h, l4_w, l4_c = 32, 32, 64
l3_h, l3_w, l3_c = 16, 16, 128
l2_h, l2_w, l2_c = 8, 8, 256
l1_h, l1_w, l1_c = 4, 4, 512
# construct the network layer by layer
# layer 0: combines the conditional vec with the noise vec
sen_rep = op.linear(sen_rep, 128, 'conditional_vec')
self.l0 = tf.concat(1, [img_z, op.lrelu(sen_rep)])
# layer 1: the linear projection
self.l1 = op.linear(self.l0, l1_w * l1_h * l1_c, 'l0_lin')
self.l1 = tf.reshape(self.l1, [self.batch_size, l1_h, l1_w, l1_c])
self.l1_bn = op.batch_norm(name='l1_bn0')
self.l1 = tf.nn.relu(self.l1_bn(self.l1, train=self.train))
# layer 2: first conv1
self.l2 = op.deconv2d(self.l1, [self.batch_size, l2_h, l2_w, l2_c],
name='l2')
self.l2_bn = op.batch_norm(name='l2_bn0')
self.l2 = tf.nn.relu(self.l2_bn(self.l2, train=self.train))
# layer 3: conv2
self.l3 = op.deconv2d(self.l2, [self.batch_size, l3_h, l3_w, l3_c],
name='l3')
self.l3_bn = op.batch_norm(name='l3_bn0')
self.l3 = tf.nn.relu(self.l3_bn(self.l3, train=self.train))
# layer 4: conv4
self.l4 = op.deconv2d(self.l3, [self.batch_size, l4_h, l4_w, l4_c],
name='l4')
self.l4_bn = op.batch_norm(name='l4_bn0')
self.l4 = tf.nn.relu(self.l4_bn(self.l4, train=self.train))
# layer 5: conv5 / final
self.l5 = op.deconv2d(self.l4, [self.batch_size, l5_h, l5_w, l5_c],
name='l5')
self.fake_img = tf.nn.tanh(self.l5) # [-1, 1]
img_shape = self.fake_img.get_shape()
# check the size of the image
assert (img_shape[1] == 64) and \
(img_shape[2] == 64) and (img_shape[3] == 3), \
logger.error('Wrong fake image dimension: {}'.format(img_shape))
return
def build_image_generator(self, img_z):
with tf.variable_scope('img_generator'):
nf = 64
# layer 1: a projection after the noise into 4, 4, 512
self.l1 = op.linear(img_z, nf * 8 * 4 * 4) # 'l0_b*[4*4*512]'
self.l1 = tf.reshape(
self.l1, [self.batch_size, 4, 4, nf * 8]) # 'l0_b*4*4*4*512')
self.l1_bn = op.batch_norm(name='l1_bn0')
self.l1 = tf.nn.relu(self.l1_bn(self.l1, train=self.train))
print('name: {}, size: {}'.format(self.l1.name,
self.l1.get_shape()))
# layer 2: first conv1
self.l2 = op.deconv2d(self.l1, [self.batch_size, 8, 8, nf * 4],
name='l2')
self.l2_bn = op.batch_norm(name='l2_bn0')
self.l2 = tf.nn.relu(self.l2_bn(self.l2, train=self.train))
# layer 3: conv2
self.l3 = op.deconv2d(self.l2, [self.batch_size, 16, 16, nf * 2],
name='l3')
self.l3_bn = op.batch_norm(name='l3_bn0')
self.l3 = tf.nn.relu(self.l3_bn(self.l3, train=self.train))
# layer 4: conv4
self.l4 = op.deconv2d(self.l3, [self.batch_size, 32, 32, nf],
name='l4')
self.l4_bn = op.batch_norm(name='l4_bn0')
self.l4 = tf.nn.relu(self.l4_bn(self.l4, train=self.train))
# layer 5: conv5 / final
self.l5 = op.deconv2d(self.l4, [self.batch_size, 64, 64, 3],
name='l5')
# self.l5_bn = op.batch_norm(name='l5_bn0')
# self.l5 = tf.nn.relu(self.l5_bn(self.l5, train=self.train))
self.fake_img = tf.nn.tanh(self.l5)
img_shape = self.fake_img.get_shape()
# check the size of the image
assert (img_shape[1] == 64) and \
(img_shape[2] == 64) and (img_shape[3] == 3), \
logger.error('Wrong fake image dimension: {}'.format(img_shape))
return
def decode_layer(x, output_width, output_filters, layer, enc_layer, dropout=False, do_concat=True):
dec = deconv2d(tf.nn.relu(x), [self.batch_size, output_width,
output_width, output_filters], scope="g_d%d_deconv" % layer)
if layer != 8:
# IMPORTANT: normalization for last layer
# Very important, otherwise GAN is unstable
dec = batch_norm(dec, is_training, scope="g_d%d_bn" % layer)
if dropout:
dec = tf.nn.dropout(dec, 0.5)
if do_concat:
dec = tf.concat([dec, enc_layer], 3)
return dec
def decode_layer(x,
output_width,
output_filters,
layer,
dropout=False):
dec = deconv2d(tf.nn.relu(x), [
self.batch_size, output_width, output_width, output_filters
],
scope="d_d%d_deconv" % layer)
if layer != 8:
dec = batch_norm(dec,
is_training,
scope="d_d%d_bn" % layer)
if dropout:
dec = tf.nn.dropout(dec, 0.5)
return dec
