def network(self, image, batch_size, update_collection):
from core.resnet import block, ops
if self.format == 'NHWC':
image = tf.transpose(image, [0, 3, 1, 2]) # NHWC to NCHW
h0 = lrelu(
ops.conv2d.Conv2D(self.prefix + 'h0_conv', 3, self.dim, 3, image))
h1 = block.ResidualBlock(self.prefix + 'res1',
self.dim,
2 * self.dim,
3,
h0,
resample='down')
h2 = block.ResidualBlock(self.prefix + 'res2',
2 * self.dim,
4 * self.dim,
3,
h1,
resample='down')
h3 = block.ResidualBlock(self.prefix + 'res3',
4 * self.dim,
8 * self.dim,
3,
h2,
resample='down')
h4 = block.ResidualBlock(self.prefix + 'res4',
8 * self.dim,
8 * self.dim,
3,
h3,
resample='down')
h4 = tf.reshape(h4, [-1, 4 * 4 * 8 * self.dim])
hF = linear(h4, self.o_dim, self.prefix + 'h5_lin')
return {'h0': h0, 'h1': h1, 'h2': h2, 'h3': h3, 'h4': h4, 'hF': hF}
def network(self, seed, batch_size, update_collection):
from core.resnet import block, ops
s1, s2, s4, s8, s16, s32 = conv_sizes(self.output_size,
layers=5,
stride=2)
# project `z` and reshape
z_ = linear(seed, self.dim * 16 * s32 * s32, self.prefix + 'h0_lin')
h0 = tf.reshape(z_, [-1, self.dim * 16, s32, s32]) # NCHW format
h1 = block.ResidualBlock(self.prefix + 'res1',
16 * self.dim,
8 * self.dim,
3,
h0,
resample='up')
h2 = block.ResidualBlock(self.prefix + 'res2',
8 * self.dim,
4 * self.dim,
3,
h1,
resample='up')
h3 = block.ResidualBlock(self.prefix + 'res3',
4 * self.dim,
2 * self.dim,
3,
h2,
resample='up')
h4 = block.ResidualBlock(self.prefix + 'res4',
2 * self.dim,
self.dim,
3,
h3,
resample='up')
h4 = ops.batchnorm.Batchnorm('g_h4', [0, 2, 3], h4)
h4 = tf.nn.relu(h4)
# h5 = lib.ops.conv2d.Conv2D('g_h5', dim, 3, 3, h4)
if self.format == 'NHWC':
h4 = tf.transpose(h4, [0, 2, 3, 1]) # NCHW to NHWC
h5 = deconv2d(h4,
self.data_format(batch_size, s1, s1, self.c_dim),
name=self.prefix + 'g_h5')
return tf.nn.sigmoid(h5)
def network(self, seed, y, batch_size, update_collection):
from core.resnet import block, ops
s1, s2, s4, s8, s16, s32 = conv_sizes(self.output_size,
layers=5,
stride=2)
# project `z` and reshape
if self.output_size == 64:
s32 = 4
z_ = linear(seed, self.dim * 16 * s32 * s32, self.prefix + 'h0_lin')
h0 = tf.reshape(z_, [-1, self.dim * 16, s32, s32]) # NCHW format
if self.output_size == 64:
h0_bis = h0
else:
h0_bis = block.ResidualBlock(self.prefix + 'res0_bis',
16 * self.dim,
16 * self.dim,
3,
h0,
y=y,
num_classes=self.num_classes,
resample='up',
mode='cond_batchnorm')
h1 = block.ResidualBlock(self.prefix + 'res1',
16 * self.dim,
8 * self.dim,
3,
h0_bis,
y=y,
num_classes=self.num_classes,
resample='up',
mode='cond_batchnorm')
h2 = block.ResidualBlock(self.prefix + 'res2',
8 * self.dim,
4 * self.dim,
3,
h1,
y=y,
num_classes=self.num_classes,
resample='up',
mode='cond_batchnorm')
h3 = block.ResidualBlock(self.prefix + 'res3',
4 * self.dim,
2 * self.dim,
3,
h2,
y=y,
num_classes=self.num_classes,
resample='up',
mode='cond_batchnorm')
h4 = block.ResidualBlock(self.prefix + 'res4',
2 * self.dim,
self.dim,
3,
h3,
y=y,
num_classes=self.num_classes,
resample='up',
mode='cond_batchnorm')
h4 = ops.batchnorm.Batchnorm('g_h4', [0, 2, 3], h4)
h4 = tf.nn.relu(h4)
if self.format == 'NHWC':
h4 = tf.transpose(h4, [0, 2, 3, 1]) # NCHW to NHWC
h5 = deconv2d(h4,
self.data_format(batch_size, s1, s1, self.c_dim),
k_h=3,
k_w=3,
d_h=1,
d_w=1,
name=self.prefix + 'g_h5')
return tf.nn.sigmoid(h5)
def network(self, image, batch_size, update_collection, y):
from core.resnet import block, ops
if self.format == 'NHWC':
image = tf.transpose(image, [0, 3, 1, 2]) # NHWC to NCHW
h0 = lrelu(
ops.conv2d.Conv2D(
self.prefix + 'h0_conv',
3,
self.dim,
3,
image,
update_collection=update_collection,
with_sn=self.with_sn,
with_learnable_sn_scale=self.with_learnable_sn_scale))
h1 = block.ResidualBlock(
self.prefix + 'res1',
self.dim,
2 * self.dim,
3,
h0,
resample='down',
update_collection=update_collection,
with_sn=self.with_sn,
with_learnable_sn_scale=self.with_learnable_sn_scale)
h2 = block.ResidualBlock(
self.prefix + 'res2',
2 * self.dim,
4 * self.dim,
3,
h1,
resample='down',
update_collection=update_collection,
with_sn=self.with_sn,
with_learnable_sn_scale=self.with_learnable_sn_scale)
h3 = block.ResidualBlock(
self.prefix + 'res3',
4 * self.dim,
8 * self.dim,
3,
h2,
resample='down',
update_collection=update_collection,
with_sn=self.with_sn,
with_learnable_sn_scale=self.with_learnable_sn_scale)
h4 = block.ResidualBlock(
self.prefix + 'res4',
8 * self.dim,
16 * self.dim,
3,
h3,
resample='down',
update_collection=update_collection,
with_sn=self.with_sn,
with_learnable_sn_scale=self.with_learnable_sn_scale)
if image.get_shape().as_list()[2] == 64:
h4_bis = h4
else:
h4_bis = block.ResidualBlock(
self.prefix + 'res4_bis',
16 * self.dim,
16 * self.dim,
3,
h4,
resample=None,
update_collection=update_collection,
with_sn=self.with_sn,
with_learnable_sn_scale=self.with_learnable_sn_scale)
h4_bis = lrelu(h4_bis)
h4_bis = tf.reduce_sum(h4_bis, axis=[2, 3])
hF = linear(h4_bis,
self.o_dim,
self.prefix + 'h5_lin',
update_collection=update_collection,
with_sn=self.with_sn,
with_learnable_sn_scale=self.with_learnable_sn_scale)
if not y is None:
w_y = linear_one_hot(
y,
self.o_dim,
self.num_classes,
name=self.prefix + "Linear_one_hot",
update_collection=update_collection,
with_sn=self.with_sn,
with_learnable_sn_scale=self.with_learnable_sn_scale)
hF += tf.reduce_sum(w_y * hF, axis=1, keepdims=True)
return {'h0': h0, 'h1': h1, 'h2': h2, 'h3': h3, 'h4': h4, 'hF': hF}