def _generator(self, z, dims, train_phase, activation=tf.nn.relu, scope_name="generator"):
N = len(dims)
image_size = self.resized_image_size // (2 ** (N - 1))
with tf.variable_scope(scope_name) as scope:
W_z = utils.weight_variable([self.z_dim, dims[0] * image_size * image_size], name="W_z")
b_z = utils.bias_variable([dims[0] * image_size * image_size], name="b_z")
h_z = tf.matmul(z, W_z) + b_z
h_z = tf.reshape(h_z, [-1, image_size, image_size, dims[0]])
h_bnz = utils.batch_norm(h_z, dims[0], train_phase, scope="gen_bnz")
h = activation(h_bnz, name='h_z')
utils.add_activation_summary(h)
for index in range(N - 2):
image_size *= 2
W = utils.weight_variable([5, 5, dims[index + 1], dims[index]], name="W_%d" % index)
b = utils.bias_variable([dims[index + 1]], name="b_%d" % index)
deconv_shape = tf.stack([tf.shape(h)[0], image_size, image_size, dims[index + 1]])
h_conv_t = utils.conv2d_transpose_strided(h, W, b, output_shape=deconv_shape)
h_bn = utils.batch_norm(h_conv_t, dims[index + 1], train_phase, scope="gen_bn%d" % index)
h = activation(h_bn, name='h_%d' % index)
utils.add_activation_summary(h)
image_size *= 2
W_pred = utils.weight_variable([5, 5, dims[-1], dims[-2]], name="W_pred")
b_pred = utils.bias_variable([dims[-1]], name="b_pred")
deconv_shape = tf.stack([tf.shape(h)[0], image_size, image_size, dims[-1]])
h_conv_t = utils.conv2d_transpose_strided(h, W_pred, b_pred, output_shape=deconv_shape)
pred_image = tf.nn.tanh(h_conv_t, name='pred_image')
utils.add_activation_summary(pred_image)
return pred_image
def generator_net(x, training, opts, name='Generator'):
with tf.variable_scope(name, reuse=tf.AUTO_REUSE):
sz = opts.img_size // 16
# x is input. 1 x 1 x nz
# state size. sz x sz x (ngf*8)
y = deconv(x, opts.ngf*8, sz, 1, 'valid', 'deconv1')
y = batch_norm(y, training)
y = relu(y)
# state size. (sz*2) x (sz*2) x (ngf*4)
y = deconv(y, opts.ngf*4, 4, 2, 'same', 'deconv2')
y = batch_norm(y, training)
y = relu(y)
# state size. (sz*4) x (sz*4) x (ngf*2)
y = deconv(y, opts.ngf*2, 4, 2, 'same', 'deconv3')
y = batch_norm(y, training)
y = relu(y)
# state size. (sz*8) x (sz*8) x ngf
y = deconv(y, opts.ngf, 4, 2, 'same', 'deconv4')
y = batch_norm(y, training)
y = relu(y)
# state size. (sz*16) x (sz*16) x nc
y = deconv(y, opts.nc, 4, 2, 'same', 'deconv5')
y = tf.nn.tanh(y)
tf.logging.info("Generator output: {}".format(y))
return y
def f(input, params, stats, mode, prefix=''):
x = F.conv2d(input, params[prefix + 'conv0'], padding=1)
g0 = group(x, params, stats, prefix + 'group0', mode, 1)
g1 = group(g0, params, stats, prefix + 'group1', mode, 2)
g2 = group(g1, params, stats, prefix + 'group2', mode, 2)
o = F.relu(batch_norm(g2, params, stats, prefix + 'bn', mode))
o = F.avg_pool2d(o, 8, 1, 0)
o = o.view(o.size(0), -1)
o = F.linear(o, params[prefix + 'fc.weight'],
params[prefix + 'fc.bias'])
#x_s = F.conv2d(input, params[prefix+'conv0_s'], padding=1)
if stu_depth != 0:
if opt.param_share:
gs0 = group_student(x, params, stats, prefix + 'group0', mode,
1, n_s)
gs1 = group_student(gs0, params, stats, prefix + 'group1',
mode, 2, n_s)
gs2 = group_student(gs1, params, stats, prefix + 'group2',
mode, 2, n_s)
else:
gs0 = group_student(x, params, stats, prefix + 'groups0', mode,
1, n_s)
gs1 = group_student(gs0, params, stats, prefix + 'groups1',
mode, 2, n_s)
gs2 = group_student(gs1, params, stats, prefix + 'groups2',
mode, 2, n_s)
os = F.relu(batch_norm(gs2, params, stats, prefix + 'bns', mode))
os = F.avg_pool2d(os, 8, 1, 0)
os = os.view(os.size(0), -1)
os = F.linear(os, params[prefix + 'fcs.weight'],
params[prefix + 'fcs.bias'])
return os, o, [g0, g1, g2, gs0, gs1, gs2]
else:
return o, [g0, g1, g2]
def generative_res(self, images, reuse=False):
img_size = self.run_flags.scale * self.run_flags.input_length
with tf.variable_scope('generator') as scope:
if reuse:
scope.reuse_variables()
if self.run_flags.run == 'train':
is_training = True
else:
is_training = False
if self.run_flags.run == 'gan':
leak = 0.01
else:
leak = 0.2
n = conv2d(images,
self.run_flags.input_length,
kernel=3,
stride=1,
name='g_res_conv1')
n = tf.nn.relu(n)
tmp = n
for i in range(16):
nn = conv2d(n,
self.run_flags.input_length,
3,
1,
name='g_resblock/%s/c1' % i)
nn = batch_norm(nn, name='g_resblock/%s/b1' % i)
nn = tf.nn.relu(nn)
nn = conv2d(nn,
self.run_flags.input_length,
3,
1,
name='g_resblock/%s/c2' % i)
nn = batch_norm(nn, name='g_resblock/%s/b2' % i)
nn = tf.add(nn, n, name='g_resblock/%s/add' % i)
n = nn
n = conv2d(n,
self.run_flags.input_length,
3,
1,
name='g_res_conv2')
n = batch_norm(n, name='g_res_bn2')
n = tf.add(n, tmp, name='g_res_add2')
n = conv2d(n,
2 * self.run_flags.input_length,
3,
1,
name='g_res_conv3')
n = tf.image.resize_images(n, size=[img_size, img_size])
n = conv2d(n, 3, 1, 1, name='g_res_conv4')
n = tf.nn.tanh(n)
return n
def discriminative(self, images, reuse=False):
with tf.variable_scope('discriminative') as scope:
if reuse:
scope.reuse_variables()
if self.run_flags.run == 'train':
is_training = True
else:
is_training = False
conv1 = lrelu(batch_norm(conv2d(images, output_dim=64, kernel=7, stride=1, name='d_conv1'), \
is_training=is_training, name='g_conv1_bn')) # 128 x 128 x 64
conv2 = lrelu(batch_norm(conv2d(conv1, output_dim=64, kernel=7, stride=2, name='d_conv2'), \
is_training=is_training, name='g_conv2_bn')) # 64 x 64 x 64
conv3 = lrelu(batch_norm(conv2d(conv2, output_dim=32, kernel=3, stride=2, name='d_conv3'), \
is_training=is_training, name='g_conv3_bn')) # 32 x 32 x 32
conv4 = lrelu(batch_norm(conv2d(conv3, output_dim=1, kernel=3, stride=2, name='d_conv4'), \
is_training=is_training, name='g_conv4_bn')) # 16 x 16 x 1
# conv1 = lrelu(conv2d(images, output_dim=64, kernel=7, stride=1, name='d_conv1')) # 128 x 128 x 64
#
# conv2 = lrelu(conv2d(conv1, output_dim=64, kernel=7, stride=2, name='d_conv2')) # 64 x 64 x 64
#
# conv3 = lrelu(conv2d(conv2, output_dim=32, kernel=3, stride=2, name='d_conv3')) # 32 x 32 x 32
#
# conv4 = lrelu(conv2d(conv3, output_dim=1, kernel=3, stride=2, name='d_conv4')) # 16 x 16 x 1
fc = tf.reshape(conv4, [-1, 16 * 16 * 1])
fc = ful_connect(fc, output_size=1, name='d_fc')
return fc
def discriminative_gan(self, images, reuse=False):
'''Discriminate 128 x 128 x 3 images fake or real within the range [fake, real] = [0, 1].'''
with tf.variable_scope('discriminator') as scope:
if reuse:
scope.reuse_variables()
if self.run_flags.run == 'train':
is_training = True
else:
is_training = False
conv1 = conv2d(images,
output_dim=64,
kernel=7,
stride=1,
name='d_conv1')
conv1 = batch_norm(conv1,
is_training=is_training,
name='d_conv1_bn')
conv1 = lrelu(conv1, 0.01)
# 128 x 128 x 64
conv2 = conv2d(conv1,
output_dim=64,
kernel=7,
stride=2,
name='d_conv2')
conv2 = batch_norm(conv2,
is_training=is_training,
name='d_conv2_bn')
conv2 = lrelu(conv2, 0.01)
# 64 x 64 x 64
conv3 = conv2d(conv2,
output_dim=32,
kernel=3,
stride=2,
name='d_conv3')
conv3 = batch_norm(conv3,
is_training=is_training,
name='d_conv3_bn')
conv3 = lrelu(conv3, 0.01)
# 32 x 32 x 32
conv4 = conv2d(conv3,
output_dim=1,
kernel=3,
stride=2,
name='d_conv4')
conv4 = batch_norm(conv4,
is_training=is_training,
name='d_conv4_bn')
conv4 = lrelu(conv4, 0.01)
# 16 x 16 x 1
fc = tf.reshape(conv4, [-1, 16 * 16 * 1])
fc = ful_connect(fc, output_size=1, name='d_fc')
return fc
def residual(x, in_channel, out_channel, is_training, norm):
"""residual unit with 2 layers
convolution:
width filter: 1
height filter: 3
"""
orig_x = x
with tf.variable_scope('conv1'):
conv1 = utils.conv(x, [1, 3, in_channel, out_channel],
[out_channel],
padding='SAME')
if norm:
conv1 = utils.batch_norm(conv1, is_training)
relu1 = utils.activation(conv1)
with tf.variable_scope('conv2'):
conv2 = utils.conv(relu1, [1, 3, out_channel, out_channel],
[out_channel],
padding='SAME')
if norm:
conv2 = utils.batch_norm(conv2, is_training)
with tf.variable_scope('add'):
if in_channel != out_channel:
orig_x = utils.conv(x, [1, 1, in_channel, out_channel],
[out_channel],
padding='SAME')
return utils.activation(conv2 + orig_x)
def discriminator_net(x, training, opts, name='Discriminator'):
with tf.variable_scope(name, reuse=tf.AUTO_REUSE):
sz = opts.img_size // 16
# input is (sz*16) x (sz*16) x nc
# state size. (sz*8) x (sz*8) x ndf
y = conv(x, opts.ndf, 4, 2, 'same', 'conv1')
y = leaky_relu(y, 0.2)
# state size. (sz*4) x (sz*4) x (ndf*2)
y = conv(y, opts.ndf * 2, 4, 2, 'same', 'conv2')
y = batch_norm(y, training)
y = leaky_relu(y, 0.2)
# state size. (sz*2) x (sz*2) x (ndf*4)
y = conv(y, opts.ndf * 4, 4, 2, 'same', 'conv3')
y = batch_norm(y, training)
y = leaky_relu(y, 0.2)
# state size. sz x sz x (ndf*8)
y = conv(y, opts.ndf * 8, 4, 2, 'same', 'conv4')
y = batch_norm(y, training)
y = leaky_relu(y, 0.2)
# output
y = conv(y, 1, sz, 1, 'valid', 'conv5')
logits = tf.reshape(y, (-1, 1))
return logits
def block(x, params, base, mode, stride):
y = F.conv2d(x, params[base+'.conv0'], stride=stride, padding=1)
o1 = F.relu(utils.batch_norm(y, params, base+'.bn0', mode), inplace=True)
z = F.conv2d(o1, params[base+'.conv1'], stride=1, padding=1)
o2 = utils.batch_norm(z, params, base+'.bn1', mode)
if base + '.convdim' in params:
return F.relu(o2 + F.conv2d(x, params[base+'.convdim'], stride=stride), inplace=True)
else:
return F.relu(o2 + x, inplace=True)
def block(x, params, base, mode, stride):
o1 = F.relu(utils.batch_norm(x, params, base + '.bn0', mode), inplace=True)
y = F.conv2d(o1, params[base + '.conv0'], stride=stride, padding=1)
o2 = F.relu(utils.batch_norm(y, params, base + '.bn1', mode), inplace=True)
z = F.conv2d(o2, params[base + '.conv1'], stride=1, padding=1)
if base + '.convdim' in params:
return z + F.conv2d(o1, params[base + '.convdim'], stride=stride)
else:
return z + x
def block(x, params, stats, base, mode, stride):
o1 = F.relu(batch_norm(x, params, stats, base + '.bn0', mode), inplace=True)
y = F.conv2d(o1, params[base + '.conv0'], stride=stride, padding=1)
o2 = F.relu(batch_norm(y, params, stats, base + '.bn1', mode), inplace=True)
z = F.conv2d(o2, params[base + '.conv1'], stride=1, padding=1)
if base + '.convdim' in params:
return z + F.conv2d(o1, params[base + '.convdim'], stride=stride)
else:
return z + x
def build_dcgan_discriminator(input_op,
is_training=True,
scope='UDiscriminator',
reuse=False,
block_num=4,
min_filters=64,
kernel_size=3,
activation=tf.nn.leaky_relu):
"""build_dcgan_discriminator
The network structure follows that of DCGAN except that
1) one more conv before each stride=2 down-sampling conv layer;
2) change kernel size from 5 to 3 since double 3x3 kernels make similar effect with 5x5
"""
with tf.variable_scope(scope, reuse=reuse):
net = input_op
end_points = dict()
for idx in range(block_num):
net = tf.layers.conv2d(net,
min_filters * (idx + 1),
kernel_size=kernel_size,
padding='same',
name='conv_' + str(2 * idx),
use_bias=False,
activation=None,
kernel_initializer=tf.contrib.layers.
xavier_initializer_conv2d())
net = activation(
batch_norm(net,
is_training=is_training,
name='bn_' + str(2 * idx)))
net = tf.layers.conv2d(net,
min_filters * (idx + 1),
kernel_size=kernel_size,
strides=2,
padding='same',
name='conv_' + str(2 * idx + 1),
activation=None,
use_bias=False,
kernel_initializer=tf.contrib.layers.
xavier_initializer_conv2d())
net = activation(
batch_norm(net,
is_training=is_training,
name='bn_' + str(2 * idx + 1)))
# end_points should be returned to calculate the
end_points['pool_' + str(idx)] = net
batch_size = net.get_shape().as_list()[0]
net = tf.reshape(net, [batch_size, -1])
logits = tf.layers.dense(
net,
1,
use_bias=True,
kernel_initializer=tf.contrib.layers.xavier_initializer())
return logits, end_points
def block(x, params, base, mode, stride):
o1 = F.relu(utils.batch_norm(x, params, base + '.bn0', mode), inplace=True)
y = F.conv2d(o1, params[base + '.conv0'], stride=stride, padding=1)
o2 = F.relu(utils.batch_norm(y, params, base + '.bn1', mode), inplace=True)
if dropout > 0:
o2 = F.dropout(o2, p=dropout, training=mode, inplace=False)
z = F.conv2d(o2, params[base + '.conv1'], stride=1, padding=1)
if base + '.convdim' in params:
return z + F.conv2d(o1, params[base + '.convdim'], stride=stride)
else:
return z + x
def discriminator(self, img, cond, reuse):
dim = len(img.get_shape())
with tf.variable_scope("disc", reuse=reuse):
image = tf.concat([img, cond], dim - 1)
feature = conf.conv_channel_base
h0 = lrelu(conv2d(image, feature, name="h0"))
h1 = lrelu(batch_norm(conv2d(h0, feature * 2, name="h1"), "h1"))
h2 = lrelu(batch_norm(conv2d(h1, feature * 4, name="h2"), "h2"))
h3 = lrelu(batch_norm(conv2d(h2, feature * 8, name="h3"), "h3"))
h4 = linear(tf.reshape(h3, [1, -1]), 1, "linear")
return h4
def block(x, params, base, mode, stride):
o1 = F.relu(utils.batch_norm(x, params, base + '.bn0', mode),
inplace=True)
y = F.conv2d(o1, params[base + '.conv0'], stride=stride, padding=1)
o2 = F.relu(utils.batch_norm(y, params, base + '.bn1', mode),
inplace=True)
z = F.conv2d(o2, params[base + '.conv1'], stride=1, padding=1)
#加入的是bottlneck?1*1的卷积层
if base + '.convdim' in params:
return z + F.conv2d(o1, params[base + '.convdim'], stride=stride)
else:
return z + x
def deconv_res_block(self, name, x, h, w, cout):
with tf.variable_scope(name):
bn1 = utils.batch_norm(name=name + 'bn1')
x = self.deconv2d(x, [params.bs, h, w, cout],
k_h=3,
k_w=3,
name=name + '.1')
x = tf.nn.relu(bn1(x))
bn2 = utils.batch_norm(name=name + 'bn2')
x = self.conv2d(x, cout, d_h=1, d_w=1, name='conv') + x
x = tf.nn.relu(bn2(x))
return x
def build_vgg_upsample_block(input_op, concat, filters, name):
with tf.variable_scope(name):
upsampled = tf.layers.conv2d_transpose(input_op, filters,
**upsample_params)
upsampled = tf.nn.relu(
batch_norm(upsampled, is_training=is_training, name='upbn_1'))
net = tf.concat([upsampled, concat], axis=-1)
net = tf.layers.conv2d(net, filters, **conv_params)
net = tf.nn.relu(
batch_norm(net, is_training=is_training, name='upbn_2'))
conv_params['use_bias'] = True
return tf.layers.conv2d(net, filters, **conv_params)
def block(x, params, base, mode, stride, out_dict):
o1 = F.relu(utils.batch_norm(x, params, base + '.bn0', mode),
inplace=True)
y = F.conv2d(o1, params[base + '.conv0'], stride=stride, padding=1)
o2 = F.relu(utils.batch_norm(y, params, base + '.bn1', mode),
inplace=True)
z = F.conv2d(o2, params[base + '.conv1'], stride=1, padding=1)
if base + '.convdim' in params:
o = z + F.conv2d(o1, params[base + '.convdim'], stride=stride)
else:
o = z + x
out_dict[base + '.relu0'] = o1
out_dict[base + '.relu1'] = o2
return o, out_dict
def _generator(self,
z,
dims,
train_phase,
activation=tf.nn.relu,
scope_name="generator"):
N = len(dims)
image_size = self.resized_image_size // (2**(N - 1))
with tf.variable_scope(scope_name) as scope:
W_z = utils.weight_variable(
[self.z_dim, dims[0] * image_size * image_size], name="W_z")
h_z = tf.matmul(z, W_z)
h_z = tf.reshape(h_z, [-1, image_size, image_size, dims[0]])
# h_bnz = tf.contrib.layers.batch_norm(inputs=h_z, decay=0.9, epsilon=1e-5, is_training=train_phase,
# scope="gen_bnz")
# h_bnz = utils.batch_norm(h_z, dims[0], train_phase, scope="gen_bnz")
h_bnz = utils.batch_norm('gen_bnz', h_z, True, 'NHWC', train_phase)
h = activation(h_bnz, name='h_z')
utils.add_activation_summary(h)
for index in range(N - 2):
image_size *= 2
W = utils.weight_variable([4, 4, dims[index + 1], dims[index]],
name="W_%d" % index)
b = tf.zeros([dims[index + 1]])
deconv_shape = tf.stack(
[tf.shape(h)[0], image_size, image_size, dims[index + 1]])
h_conv_t = utils.conv2d_transpose_strided(
h, W, b, output_shape=deconv_shape)
# h_bn = tf.contrib.layers.batch_norm(inputs=h_conv_t, decay=0.9, epsilon=1e-5, is_training=train_phase,
# scope="gen_bn%d" % index)
# h_bn = utils.batch_norm(h_conv_t, dims[index + 1], train_phase, scope="gen_bn%d" % index)
h_bn = utils.batch_norm("gen_bn%d" % index, h_conv_t, True,
'NHWC', train_phase)
h = activation(h_bn, name='h_%d' % index)
utils.add_activation_summary(h)
image_size *= 2
W_pred = utils.weight_variable([4, 4, dims[-1], dims[-2]],
name="W_pred")
b = tf.zeros([dims[-1]])
deconv_shape = tf.stack(
[tf.shape(h)[0], image_size, image_size, dims[-1]])
h_conv_t = utils.conv2d_transpose_strided(
h, W_pred, b, output_shape=deconv_shape)
pred_image = tf.nn.tanh(h_conv_t, name='pred_image')
utils.add_activation_summary(pred_image)
return pred_image
def _discriminator(self,
input_images,
dims,
train_phase,
activation=tf.nn.relu,
scope_name="discriminator",
scope_reuse=False):
N = len(dims)
with tf.variable_scope(scope_name) as scope:
if scope_reuse:
scope.reuse_variables()
h = input_images
skip_bn = True # First layer of discriminator skips batch norm
for index in range(N - 2):
W = utils.weight_variable([4, 4, dims[index], dims[index + 1]],
name="W_%d" % index)
b = tf.zeros([dims[index + 1]])
h_conv = utils.conv2d_strided(h, W, b)
if skip_bn:
h_bn = h_conv
skip_bn = False
else:
h_bn = utils.batch_norm(h_conv,
dims[index + 1],
train_phase,
scope="disc_bn%d" % index)
h = activation(h_bn, name="h_%d" % index)
utils.add_activation_summary(h)
W_pred = utils.weight_variable([4, 4, dims[-2], dims[-1]],
name="W_pred")
b = tf.zeros([dims[-1]])
h_pred = utils.conv2d_strided(h, W_pred, b)
return None, h_pred, None # Return the last convolution output. None values are returned to maintatin disc from other GAN
def _discriminator(self, input_images, dims, train_phase, activation=tf.nn.relu, scope_name="discriminator",
scope_reuse=False):
N = len(dims)
with tf.variable_scope(scope_name) as scope:
if scope_reuse:
scope.reuse_variables()
h = input_images
skip_bn = True # First layer of discriminator skips batch norm
for index in range(N - 2):
W = utils.weight_variable([5, 5, dims[index], dims[index + 1]], name="W_%d" % index)
b = utils.bias_variable([dims[index + 1]], name="b_%d" % index)
h_conv = utils.conv2d_strided(h, W, b)
if skip_bn:
h_bn = h_conv
skip_bn = False
else:
h_bn = utils.batch_norm(h_conv, dims[index + 1], train_phase, scope="disc_bn%d" % index)
h = activation(h_bn, name="h_%d" % index)
utils.add_activation_summary(h)
shape = h.get_shape().as_list()
image_size = self.resized_image_size // (2 ** (N - 2)) # dims has input dim and output dim
h_reshaped = tf.reshape(h, [self.batch_size, image_size * image_size * shape[3]])
W_pred = utils.weight_variable([image_size * image_size * shape[3], dims[-1]], name="W_pred")
b_pred = utils.bias_variable([dims[-1]], name="b_pred")
h_pred = tf.matmul(h_reshaped, W_pred) + b_pred
return tf.nn.sigmoid(h_pred), h_pred, h
def generative(self, x, reuse=False):
with tf.variable_scope('generator') as scope:
if reuse:
scope.reuse_variables()
if self.run_flags.run == 'train':
is_training = True
else:
is_training = False
conv1 = lrelu(batch_norm(conv2d(x, output_dim=32, stride=1, name='g_conv1'), \
is_training=is_training, name='g_conv1_bn')) # 64 x 64 x 32
conv2 = lrelu(batch_norm(conv2d(conv1, output_dim=128, stride=1, name='g_conv2'), \
is_training=is_training, name='g_conv2_bn')) # 64 x 64 x 128
conv3 = lrelu(batch_norm(conv2d(conv2, output_dim=128, stride=1, name='g_conv3'), \
is_training=is_training, name='g_conv3_bn')) # 64 x 64 x 128
conv3_up = tf.image.resize_images(conv3, size=[128, 128])
conv4 = lrelu(batch_norm(conv2d(conv3_up, output_dim=128, stride=1, name='g_conv4'), \
is_training=is_training, name='g_conv4_bn')) # 128 x 128 x 128
conv5 = lrelu(batch_norm(conv2d(conv4, output_dim=64, stride=1, name='g_conv5'), \
is_training=is_training, name='g_conv5_bn')) # 128 x 128 x 64
conv6 = tf.nn.sigmoid(
conv2d(conv5, output_dim=3, stride=1,
name='g_conv6')) #128 x 128 x 3
# conv1 = lrelu(conv2d(x, output_dim=32, stride=1, name='g_conv1')) # 64 x 64 x 32
#
# conv2 = lrelu(conv2d(conv1, output_dim=128, stride=1, name='g_conv2')) # 64 x 64 x 128
#
# conv3 = lrelu(conv2d(conv2, output_dim=128, stride=1, name='g_conv3')) # 64 x 64 x 128
#
# conv3_up = tf.image.resize_images(conv3, size=[128, 128])
#
# conv4 = lrelu(conv2d(conv3_up, output_dim=128, stride=1, name='g_conv4')) # 128 x 128 x 128
#
# conv5 = lrelu(conv2d(conv4, output_dim=64, stride=1, name='g_conv5')) # 128 x 128 x 64
#
# conv6 = tf.nn.sigmoid(conv2d(conv5, output_dim=3, stride=1, name='g_conv6')) #128 x 128 x 3
return conv6
def graph(self, input, is_training):
with tf.name_scope('model'):
net = ut.conv_layer(input, 64, 7, 2, name='conv1')
net = ut.bottleneck(net,
128,
stride=1,
training=is_training,
name='res1')
net = ut.max_pool(net, 2, 2, 'max_pool')
net = ut.bottleneck(net,
int(self.nFeats / 2),
stride=1,
training=is_training,
name='res2')
net = ut.bottleneck(net,
self.nFeats,
stride=1,
training=is_training,
name='res3')
with tf.name_scope('stacks'):
stack_out = []
with tf.name_scope('stage_0'):
hg = ut.hourglass(net, self.nLow, self.nFeats, 'hourglass')
drop = ut.dropout(hg, self.dropout_rate, is_training,
'dropout')
ll = ut.conv_layer_bn(drop, self.nFeats, 1, 1, is_training)
out = ut.conv_layer(ll, self.num_points, 1, 1, name='out')
out_ = ut.conv_layer(out, self.nFeats, 1, 1, name='out_')
sum_ = tf.add(net, out_, name='merge')
stack_out.append(out)
for i in range(1, self.nStacks):
with tf.name_scope('stage_' + str(i)):
hg = ut.hourglass(sum_, self.nLow, self.nFeats,
'hourglass')
drop = ut.dropout(hg, self.dropout_rate, is_training,
'dropout')
ll = ut.conv_layer_bn(drop, self.nFeats, 1, 1,
is_training)
out = ut.conv_layer(ll,
self.num_points,
1,
1,
name='out')
out_ = ut.conv_layer(ll,
self.nFeats,
1,
1,
name='out_')
sum_ = tf.add(sum_, out_, name='merge')
stack_out.append(out)
with tf.name_scope('upsampling'):
net = ut.batch_norm(sum_, is_training)
net = ut.conv_layer_bn(net, self.nFeats, 3, 1, is_training)
up1 = ut.deconv_layer(net, self.num_points, 1, 2, name='up_1')
net = ut.conv_layer_bn(up1, self.nFeats, 3, 1, is_training)
up2 = ut.deconv_layer(net, self.num_points, 1, 2, name='up_2')
return tf.stack(stack_out, axis=1, name='stack_out'), up1, up2
def build_lenet(input_op, is_training=True):
params = {
'kernel_size': 5,
'padding': 'same',
'use_bias': False,
'activation': None,
'strides': 2,
'kernel_initializer': tf.contrib.layers.xavier_initializer_conv2d()
}
conv_1 = coord_conv(input_op, 32, name='conv_1', **params)
conv_1 = tf.nn.leaky_relu(
batch_norm(conv_1, is_training=is_training, name='bn_1'))
conv_2 = tf.layers.conv2d(conv_1, 64, name='conv_2', **params)
conv_2 = tf.nn.leaky_relu(
batch_norm(conv_2, is_training=is_training, name='bn_2'))
conv_3 = tf.layers.conv2d(conv_2, 128, name='conv_3', **params)
conv_3 = tf.nn.leaky_relu(
batch_norm(conv_3, is_training=is_training, name='bn_3'))
params = {
'kernel_size': 5,
'strides': 2,
'padding': 'same',
'activation': None,
'use_bias': False,
'kernel_initializer': tf.contrib.layers.xavier_initializer_conv2d()
}
deconv_1 = tf.layers.conv2d_transpose(conv_3,
64,
name='deconv_1',
**params)
deconv_1 = tf.nn.leaky_relu(batch_norm(deconv_1, is_training=is_training))
deconv_2 = tf.layers.conv2d_transpose(tf.concat([deconv_1, conv_2],
axis=-1),
32,
name='deconv_2',
**params)
deconv_2 = tf.nn.leaky_relu(
batch_norm(deconv_2, is_training=is_training, name='bn_4'))
params['activation'] = None
params['use_bias'] = True
return tf.layers.conv2d_transpose(tf.concat([deconv_2, conv_1], axis=-1),
1,
name='deconv_3',
**params)
def __init__(self, sess, image_size=64, is_crop=False, batch_size=64, sample_size=64, lowres=8, z_dim=100,
gf_dim=64, df_dim=64, gfc_dim=1024, dfc_dim=1024, c_dim=3, checkpoint_dir=None, lam=0.1):
"""
:param sess: TensorFlow session
:param batch_size: the size of batch. Should be specified before training
:param lowres: (optional) Low resolution image/mask shrink factor.
:param gf_dim: (optional)Dimension of generator filters in first conv layer
:param df_dim: (optional)Dimension of discriminator filters in first conv layer
:param gfc_dim: (optional)Dimension of gen units for fully connected layer
:param dfc_dim: (optional)Dimension of discrim units for fully connected layer
:param c_dim: (optional) Dimension of image color. For grayscale input, set to 1.
"""
self.sess = sess
self.is_crop = is_crop
self.batch_size = batch_size
self.image_size = image_size
self.sample_size = sample_size
self.image_shape = [image_size, image_size, c_dim]
self.lowres = lowres
self.lowres_size = image_size // lowres
self.lowres_shape = [self.lowres_size, self.lowres_size, c_dim]
self.z_dim = z_dim
self.gf_dim = gf_dim
self.df_dim = df_dim
self.gfc_dim = gfc_dim
self.dfc = dfc_dim
self.lam = lam
self.c_dim = c_dim
self.d_bns = [batch_norm(name='d_bn{}'.format(format(i, )) for i in range(4))]
log_size = int(math.log(image_size) / math.log(2))
self.g_bns = [batch_norm(name='g_bn{}'.format(format(i, )) for i in range(log_size))]
self.checkpoint_dir = checkpoint_dir
self.build_model()
self.model_name = 'DCGAN.model'
def f(input, params, stats, mode, prefix=''):
x = F.conv2d(input, params[prefix+'conv0'], padding=1)
g0 = group(x, params, stats, prefix+'group0', mode, 1)
g1 = group(g0, params, stats, prefix+'group1', mode, 2)
g2 = group(g1, params, stats, prefix+'group2', mode, 2)
o = F.relu(batch_norm(g2, params, stats, prefix+'bn', mode))
o = F.avg_pool2d(o, 8, 1, 0)
o = o.view(o.size(0), -1)
o = F.linear(o, params[prefix+'fc.weight'], params[prefix+'fc.bias'])
return o, [g0, g1, g2]
def harmonic_block(x, params, base, mode, stride=1, padding=1):
y = F.conv2d(x,
params['dct0'],
stride=stride,
padding=padding,
groups=x.size(1))
if base + '.bn.running_mean' in params:
y = utils.batch_norm(y, params, base + '.bn', mode, affine=False)
z = F.conv2d(y, params[base + '.conv'], padding=0)
return z
def f(input, params, mode):
x = F.conv2d(input, params['conv0'], padding=1)
g0 = group(x, params, 'group0', mode, 1)
g1 = group(g0, params, 'group1', mode, 2)
g2 = group(g1, params, 'group2', mode, 2)
o = F.relu(utils.batch_norm(g2, params, 'bn', mode))
o = F.avg_pool2d(o, 8, 1, 0)
o = o.view(o.size(0), -1)
o = F.linear(o, params['fc.weight'], params['fc.bias'])
return o
def f(input, params, stats, mode):
x = F.conv2d(input, params['conv0'], padding=1)
g0 = group(x, params, stats, 'group0', mode, 1)
g1 = group(g0, params, stats, 'group1', mode, 2)
g2 = group(g1, params, stats, 'group2', mode, 2)
o = F.relu(batch_norm(g2, params, stats, 'bn', mode, 1.))
o = F.conv2d(o, params['conv1'])
o = F.avg_pool2d(o, 8, 1, 0)
o = o.view(o.size(0), -1)
return o
def VGG19(x, n_classes, is_pretrain=True):
with tf.name_scope('VGG16'):
x = utils.conv('conv1_1', x, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv1_2', x, 64, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
with tf.name_scope('pool1'):
x = utils.pool('pool1', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = utils.conv('conv2_1', x, 128, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv2_2', x, 128, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
with tf.name_scope('pool2'):
x = utils.pool('pool2', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = utils.conv('conv3_1', x, 256, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv3_2', x, 256, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv3_3', x, 256, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv3_4', x, 256, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
with tf.name_scope('pool3'):
x = utils.pool('pool3', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = utils.conv('conv4_1', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv4_2', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv4_3', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv4_4', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
with tf.name_scope('pool4'):
x = utils.pool('pool4', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = utils.conv('conv5_1', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv5_2', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv5_3', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
x = utils.conv('conv5_4', x, 512, kernel_size=[3, 3], stride=[1, 1, 1, 1], is_pretrain=is_pretrain)
with tf.name_scope('pool5'):
x = utils.pool('pool5', x, kernel=[1, 2, 2, 1], stride=[1, 2, 2, 1], is_max_pool=True)
x = utils.FC_layer('fc6', x, out_nodes=4096)
with tf.name_scope('batch_norm1'):
x = utils.batch_norm(x)
x = utils.FC_layer('fc7', x, out_nodes=4096)
with tf.name_scope('batch_norm2'):
x = utils.batch_norm(x)
x = utils.FC_layer('fc8', x, out_nodes=n_classes)
return x
def net(x, is_training=False, verbose=False):
x = x.as_in_context(w1.context)
h1_conv = nd.Convolution(data=x,
weight=w1,
bias=b1,
kernel=w1.shape[2:],
num_filter=c1)
h1_bn = utils.batch_norm(h1_conv, gamma1, beta1, is_training, moving_mean1,
moving_variance1)
h1_activation = nd.relu(h1_conv)
h1 = nd.Pooling(data=h1_activation,
pool_type='max',
kernel=(2, 2),
stride=(2, 2))
h2_conv = nd.Convolution(data=h1,
weight=w2,
bias=b2,
kernel=w2.shape[2:],
num_filter=c2)
h2_bn = utils.batch_norm(h2_conv, gamma2, beta2, is_training, moving_mean2,
moving_variance2)
h2_activation = nd.relu(h2_conv)
h2 = nd.Pooling(data=h2_activation,
pool_type='max',
kernel=(2, 2),
stride=(2, 2))
h2 = nd.flatten(h2)
h3_linear = nd.dot(h2, w3) + b3
h3 = nd.relu(h3_linear)
h4_linear = nd.dot(h3, w4) + b4
if verbose:
print('h1 conv block: ', h1.shape)
print('h2 conv block: ', h2.shape)
print('h3 conv block: ', h3.shape)
print('h4 conv block: ', h4_linear.shape)
print('output: ', h4_linear)
return h4_linear.as_in_context(ctx)
