def build_C_global_layers(x,
name,
n_latents,
start_idx,
scope_idx,
dlatents_withl_in,
n_content,
act,
fused_modconv,
fmaps=128,
**kwargs):
'''
Build continuous latent layers, e.g. C_global layers.
'''
with tf.variable_scope(name + '-' + str(scope_idx)):
if n_content > 0:
with tf.variable_scope('Condition0'):
cond = apply_bias_act(dense_layer(
dlatents_withl_in[:, :n_content], fmaps=128),
act=act)
with tf.variable_scope('Condition1'):
cond = apply_bias_act(dense_layer(cond, fmaps=n_latents),
act='sigmoid')
else:
cond = 1.
C_global_latents = dlatents_withl_in[:, start_idx:start_idx +
n_latents] * cond
x = apply_bias_act(modulated_conv2d_layer(x,
C_global_latents,
fmaps=fmaps,
kernel=3,
up=False,
fused_modconv=fused_modconv),
act=act)
return x
def build_C_global_nocond_layers(x,
name,
n_latents,
start_idx,
scope_idx,
dlatents_withl_in,
act,
fused_modconv,
fmaps=128,
**kwargs):
'''
Build continuous latent layers, e.g. C_global layers.
'''
with tf.variable_scope(name + '-' + str(scope_idx)):
with tf.variable_scope('Conv0'):
C_global_latents = apply_bias_act(dense_layer(
dlatents_withl_in[:, start_idx:start_idx + n_latents], fmaps=128),
act=act)
# C_global_latents = dlatents_withl_in[:, start_idx:start_idx +
# n_latents]
with tf.variable_scope('Modulate'):
x = apply_bias_act(modulated_conv2d_layer(x,
C_global_latents,
fmaps=fmaps,
kernel=3,
up=False,
fused_modconv=fused_modconv),
act=act)
return x
def torgb(res, x): # res = 2..resolution_log2
with tf.variable_scope('ToRGB_lod%d' % (resolution_log2 - res)):
return networks_stylegan2.apply_bias_act(
networks_stylegan2.modulated_conv2d_layer(
x,
dlatents_in[:, res * 2 - 3],
fmaps=num_channels,
kernel=1,
demodulate=False,
fused_modconv=fused_modconv))
def torgb(x, y, res): # res = 2..resolution_log2
with tf.variable_scope('ToRGB'):
t = apply_bias_act(
modulated_conv2d_layer(x,
latents_ready_ls[res * 2 - 3],
fmaps=num_channels,
kernel=1,
demodulate=False,
fused_modconv=fused_modconv))
return t if y is None else y + t
def build_D_layers(x,
name,
n_latents,
start_idx,
scope_idx,
single_const,
dlatents_withl_in,
act,
fused_modconv,
fmaps=128,
**kwargs):
'''
Build discrete latent layers including label and D_global layers.
'''
with tf.variable_scope(name + '-' + str(scope_idx)):
if single_const:
x = apply_bias_act(modulated_conv2d_layer(
x,
dlatents_withl_in[:, start_idx:start_idx + n_latents],
fmaps=fmaps,
kernel=3,
up=False,
fused_modconv=fused_modconv),
act=act)
else:
# soft version
# x_softmax = tf.nn.softmax(
# dlatents_withl_in[:, start_idx:start_idx + n_latents], axis=1)
# x_softmax = tf.reshape(
# x_softmax, [tf.shape(x_softmax)[0], n_latents, 1, 1, 1])
# x = tf.reduce_sum(x * x_softmax, axis=1)
# print('x.shape:', x.shape.as_list())
# hard version
x_indices = tf.argmax(dlatents_withl_in[:, start_idx:start_idx +
n_latents],
axis=1)
print('x_indices.shape:', x_indices.shape.as_list())
print('before gather_nd x.shape:', x.shape.as_list())
x = tf.gather(x, x_indices, axis=0)
print('after gather_nd x.shape:', x.shape.as_list())
return x
def layer(x, layer_idx, fmaps, kernel, up=False):
x = networks_stylegan2.modulated_conv2d_layer(
x,
dlatents_in[:, layer_idx],
fmaps=fmaps,
kernel=kernel,
up=up,
resample_kernel=resample_kernel,
fused_modconv=fused_modconv)
if layer_idx > num_layers * 0.75:
if randomize_noise:
noise = tf.random_normal(
[tf.shape(x)[0], 1, x.shape[2], x.shape[3]], dtype=x.dtype)
else:
noise = tf.cast(noise_inputs[layer_idx], x.dtype)
noise_strength = tf.get_variable(
'noise_strength',
shape=[],
initializer=tf.initializers.zeros())
x += noise * tf.cast(noise_strength, x.dtype)
return networks_stylegan2.apply_bias_act(x, act=act)
def layer(x, layer_idx, fmaps, kernel, up=False):
# start_idx_layer = sum(latent_split_ls_for_std_gen[:layer_idx])
# for i in range(start_idx_layer, start_idx_layer + latent_split_ls_for_std_gen[layer_idx]):
# x = modulated_conv2d_layer(x, latents_ready_spl_ls[i], fmaps=fmaps, kernel=kernel, up=up,
# resample_kernel=resample_kernel, fused_modconv=fused_modconv)
x = modulated_conv2d_layer(x,
latents_ready_ls[layer_idx],
fmaps=fmaps,
kernel=kernel,
up=up,
resample_kernel=resample_kernel,
fused_modconv=fused_modconv)
if randomize_noise:
noise = tf.random_normal(
[tf.shape(x)[0], 1, x.shape[2], x.shape[3]], dtype=x.dtype)
else:
noise = tf.cast(noise_inputs[layer_idx], x.dtype)
noise_strength = tf.get_variable('noise_strength',
shape=[],
initializer=tf.initializers.zeros())
x += noise * tf.cast(noise_strength, x.dtype)
return apply_bias_act(x, act=act)
def build_D_layers(x,
name,
n_latents,
start_idx,
scope_idx,
dlatents_in,
act,
fused_modconv,
fmaps=128,
**kwargs):
'''
Build discrete latent layers including label and D_global layers.
'''
with tf.variable_scope(name + '-' + str(scope_idx)):
x = apply_bias_act(modulated_conv2d_layer(
x,
dlatents_in[:, start_idx:start_idx + n_latents],
fmaps=fmaps,
kernel=3,
up=False,
fused_modconv=fused_modconv),
act=act)
return x
def G_synthesis_sb_general_dsp(
dlatents_withl_in, # Input: Disentangled latents (W) [minibatch, label_size+dlatent_size].
dlatent_size=7, # Disentangled latent (W) dimensionality. Including discrete info, rotation, scaling, xy shearing, and xy translation.
label_size=0, # Label dimensionality, 0 if no labels.
D_global_size=3, # Global D_latents.
C_global_size=0, # Global C_latents.
sb_C_global_size=4, # Global spatial-biased C_latents.
C_local_hfeat_size=0, # Local heatmap*features learned C_latents.
C_local_heat_size=0, # Local heatmap learned C_latents.
num_channels=1, # Number of output color channels.
resolution=64, # Output resolution.
nonlinearity='lrelu', # Activation function: 'relu', 'lrelu', etc.
dtype='float32', # Data type to use for activations and outputs.
resample_kernel=[
1, 3, 3, 1
], # Low-pass filter to apply when resampling activations. None = no filtering.
fused_modconv=True, # Implement modulated_conv2d_layer() as a single fused op?
use_noise=False,
randomize_noise=True, # True = randomize noise inputs every time (non-deterministic), False = read noise inputs from variables.
**_kwargs): # Ignore unrecognized keyword args.
'''
dlatents_withl_in: dims contain: [label, D_global, C_global, sb_C_global,
C_local_hfeat, C_local_feat]
'''
resolution_log2 = int(np.log2(resolution)) # == 6 for resolution 64
assert resolution == 2**resolution_log2 and resolution >= 4
num_layers = resolution_log2 * 2 - 2 # == 10 for resolution 64
act = nonlinearity
images_out = None
# Primary inputs.
assert dlatent_size == D_global_size + C_global_size + sb_C_global_size + \
C_local_hfeat_size + C_local_heat_size
n_cat = label_size + D_global_size
dlatents_withl_in.set_shape([None, label_size + dlatent_size])
dlatents_withl_in = tf.cast(dlatents_withl_in, dtype)
n_content = label_size + D_global_size + C_global_size
# Noise inputs.
noise_inputs = []
for layer_idx in range(num_layers - 3):
res = (layer_idx + 7) // 2 # [3, 4, 4, 5, 5, 6, 6]
shape = [1, 1, 2**res, 2**res]
noise_inputs.append(
tf.get_variable('noise%d' % layer_idx,
shape=shape,
initializer=tf.initializers.random_normal(),
trainable=False))
# Single convolution layer with all the bells and whistles.
def noised_conv_layer(x, layer_idx, fmaps, kernel, up=False):
x = conv2d_layer(x,
fmaps=fmaps,
up=up,
kernel=kernel,
resample_kernel=resample_kernel)
if use_noise:
if randomize_noise:
noise = tf.random_normal(
[tf.shape(x)[0], 1, x.shape[2], x.shape[3]], dtype=x.dtype)
else:
noise = tf.cast(noise_inputs[layer_idx], x.dtype)
noise_strength = tf.get_variable(
'noise_strength',
shape=[],
initializer=tf.initializers.zeros())
x += noise * tf.cast(noise_strength, x.dtype)
return apply_bias_act(x, act=act)
# Early layers consists of 4x4 constant layer,
# label+global discrete latents,
# and global continuous latents.
y = None
with tf.variable_scope('4x4'):
with tf.variable_scope('Const'):
x = tf.get_variable('const',
shape=[1, 128, 4, 4],
initializer=tf.initializers.random_normal())
x = tf.tile(tf.cast(x, dtype),
[tf.shape(dlatents_withl_in)[0], 1, 1, 1])
with tf.variable_scope('Upconv'):
x = apply_bias_act(conv2d_layer(x,
fmaps=128,
kernel=3,
up=True,
resample_kernel=resample_kernel),
act=act)
with tf.variable_scope('8x8'):
with tf.variable_scope('Conv0'):
x = apply_bias_act(conv2d_layer(x, fmaps=128, kernel=3), act=act)
with tf.variable_scope('Label_Dglobal_control'):
x = apply_bias_act(modulated_conv2d_layer(
x,
dlatents_withl_in[:, :n_cat],
fmaps=128,
kernel=3,
up=False,
resample_kernel=resample_kernel,
fused_modconv=fused_modconv),
act=act)
with tf.variable_scope('After_DiscreteGlobal_noised'):
x = noised_conv_layer(x, layer_idx=0, fmaps=128, kernel=3)
with tf.variable_scope('Cglobal_control'):
start_idx = n_cat
x = apply_bias_act(modulated_conv2d_layer(
x,
dlatents_withl_in[:, start_idx:start_idx + C_global_size],
fmaps=128,
kernel=3,
up=False,
resample_kernel=resample_kernel,
fused_modconv=fused_modconv),
act=act)
with tf.variable_scope('After_ContinuousGlobal_noised'):
x = noised_conv_layer(x, layer_idx=1, up=True, fmaps=128, kernel=3)
# Spatial biased layers.
with tf.variable_scope('16x16'):
if C_local_hfeat_size > 0:
with tf.variable_scope('LocalHFeat_C_latents'):
with tf.variable_scope('ConstFeats'):
const_feats = tf.get_variable(
'constfeats',
shape=[1, C_local_hfeat_size, 32, 1, 1],
initializer=tf.initializers.random_normal())
const_feats = tf.tile(
tf.cast(const_feats,
dtype), [tf.shape(const_feats)[0], 1, 1, 1, 1])
with tf.variable_scope('ControlAttHeat'):
hfeat_start_idx = label_size + D_global_size + C_global_size + \
sb_C_global_size
att_heat = get_att_heat(x,
nheat=C_local_hfeat_size,
act=act)
att_heat = tf.reshape(
att_heat,
[tf.shape(att_heat)[0], C_local_hfeat_size, 1] +
att_heat.shape.as_list()[2:4])
# C_local_heat latent [-2, 2] --> [0, 1]
hfeat_modifier = (2 + dlatents_withl_in[:, hfeat_start_idx:hfeat_start_idx + \
C_local_hfeat_size]) / 4.
hfeat_modifier = get_conditional_modifier(
hfeat_modifier,
dlatents_withl_in[:, :n_content],
act=act)
hfeat_modifier = tf.reshape(
hfeat_modifier,
[tf.shape(x)[0], C_local_hfeat_size, 1, 1, 1])
att_heat = att_heat * hfeat_modifier
added_feats = const_feats * att_heat
added_feats = tf.reshape(added_feats, [
tf.shape(att_heat)[0],
C_local_hfeat_size * att_heat.shape.as_list()[2]
] + att_heat.shape.as_list()[3:5])
x = tf.concat([x, added_feats], axis=1)
with tf.variable_scope('SpatialBiased_C_global'):
# Rotation layers.
start_idx = start_idx + C_global_size
with tf.variable_scope('Rotation'):
r_matrix = get_r_matrix(
dlatents_withl_in[:, start_idx:start_idx + 1],
dlatents_withl_in[:, :n_content],
act=act)
x = apply_st(x, r_matrix, up=False, fmaps=128, act=act)
with tf.variable_scope('After_Rotation_noised'):
x = noised_conv_layer(x, layer_idx=2, fmaps=128, kernel=3)
# Scaling layers.
start_idx = start_idx + 1
with tf.variable_scope('Scaling'):
s_matrix = get_s_matrix(
dlatents_withl_in[:, start_idx:start_idx + 1],
dlatents_withl_in[:, :n_content],
act=act)
x = apply_st(x, s_matrix, up=False, fmaps=128, act=act)
with tf.variable_scope('After_Scaling_noised'):
x = noised_conv_layer(x,
layer_idx=3,
up=True,
fmaps=128,
kernel=3)
with tf.variable_scope('32x32'):
with tf.variable_scope('SpatialBiased_C_global'):
# Shearing layers.
with tf.variable_scope('Shearing'):
start_idx = start_idx + 1
sh_matrix = get_sh_matrix(
dlatents_withl_in[:, start_idx:start_idx + 2],
dlatents_withl_in[:, :n_content],
act=act)
x = apply_st(x, sh_matrix, up=False, fmaps=128, act=act)
with tf.variable_scope('After_Shearing_noised'):
x = noised_conv_layer(x, layer_idx=4, fmaps=128, kernel=3)
# Translation layers.
with tf.variable_scope('Translation'):
start_idx = start_idx + 2
t_matrix = get_t_matrix(
dlatents_withl_in[:, start_idx:start_idx + 2],
dlatents_withl_in[:, :n_content],
act=act)
x = apply_st(x, t_matrix, up=False, fmaps=128, act=act)
with tf.variable_scope('After_Translation_noised'):
if resolution_log2 >= 6:
x = noised_conv_layer(x,
layer_idx=5,
up=True,
fmaps=128,
kernel=3)
else:
x = noised_conv_layer(x,
layer_idx=5,
fmaps=128,
kernel=3)
with tf.variable_scope('64x64' if resolution_log2 >= 6 else '32x32'):
with tf.variable_scope('LocalHeat_C_latents'):
with tf.variable_scope('ControlAttHeat'):
heat_start_idx = label_size + D_global_size + C_global_size + \
sb_C_global_size + C_local_hfeat_size
att_heat = get_att_heat(x, nheat=C_local_heat_size, act=act)
# C_local_heat latent [-2, 2] --> [0, 1]
heat_modifier = (2 + dlatents_withl_in[:, heat_start_idx:heat_start_idx + \
C_local_heat_size]) / 4.
heat_modifier = get_conditional_modifier(
heat_modifier, dlatents_withl_in[:, :n_content], act=act)
heat_modifier = tf.reshape(
heat_modifier,
[tf.shape(heat_modifier)[0], C_local_heat_size, 1, 1])
att_heat = att_heat * heat_modifier
x = tf.concat([x, att_heat], axis=1)
with tf.variable_scope('After_LocalHeat_noised'):
x = noised_conv_layer(x, layer_idx=6, fmaps=128, kernel=3)
y = torgb(x, y, num_channels=num_channels)
# # Tail layers.
# for res in range(6, resolution_log2 + 1):
# with tf.variable_scope('%dx%d' % (res * 2, res * 2)):
# x = apply_bias_act(conv2d_layer(x,
# fmaps=128,
# kernel=1,
# up=True,
# resample_kernel=resample_kernel),
# act=act)
# y = torgb(x, y, num_channels=num_channels)
images_out = y
assert images_out.dtype == tf.as_dtype(dtype)
return tf.identity(images_out, name='images_out')
def G_synthesis_spatial_biased_dsp(
dlatents_in, # Input: Disentangled latents (W) [minibatch, dlatent_size].
dlatent_size=7, # Disentangled latent (W) dimensionality. Including discrete info, rotation, scaling, and xy translation.
D_global_size=3, # Discrete latents.
sb_C_global_size=4, # Continuous latents.
label_size=0, # Label dimensionality, 0 if no labels.
num_channels=1, # Number of output color channels.
resolution=64, # Output resolution.
fmap_base=16 <<
10, # Overall multiplier for the number of feature maps.
fmap_decay=1.0, # log2 feature map reduction when doubling the resolution.
fmap_min=1, # Minimum number of feature maps in any layer.
fmap_max=512, # Maximum number of feature maps in any layer.
architecture='skip', # Architecture: 'orig', 'skip', 'resnet'.
nonlinearity='lrelu', # Activation function: 'relu', 'lrelu', etc.
dtype='float32', # Data type to use for activations and outputs.
resample_kernel=[
1, 3, 3, 1
], # Low-pass filter to apply when resampling activations. None = no filtering.
fused_modconv=True, # Implement modulated_conv2d_layer() as a single fused op?
**_kwargs): # Ignore unrecognized keyword args.
resolution_log2 = int(np.log2(resolution))
assert resolution == 2**resolution_log2 and resolution >= 4
def nf(stage):
return np.clip(int(fmap_base / (2.0**(stage * fmap_decay))), fmap_min,
fmap_max)
assert architecture in ['orig', 'skip', 'resnet']
act = nonlinearity
images_out = None
# Primary inputs.
assert dlatent_size == D_global_size + sb_C_global_size
n_cat = label_size + D_global_size
dlatents_in.set_shape([None, label_size + dlatent_size])
dlatents_in = tf.cast(dlatents_in, dtype)
# Return rotation matrix
def get_r_matrix(r_latents, cond_latent):
# r_latents: [-2., 2.] -> [0, 2*pi]
with tf.variable_scope('Condition0'):
cond = apply_bias_act(dense_layer(cond_latent, fmaps=128), act=act)
with tf.variable_scope('Condition1'):
cond = apply_bias_act(dense_layer(cond, fmaps=1), act='sigmoid')
rad = (r_latents + 2) / 4. * 2. * np.pi
rad = rad * cond
tt_00 = tf.math.cos(rad)
tt_01 = -tf.math.sin(rad)
tt_02 = tf.zeros_like(rad)
tt_10 = tf.math.sin(rad)
tt_11 = tf.math.cos(rad)
tt_12 = tf.zeros_like(rad)
theta = tf.concat([tt_00, tt_01, tt_02, tt_10, tt_11, tt_12], axis=1)
return theta
# Return scaling matrix
def get_s_matrix(s_latents, cond_latent):
# s_latents: [-2., 2.] -> [1, 3]
with tf.variable_scope('Condition0'):
cond = apply_bias_act(dense_layer(cond_latent, fmaps=128), act=act)
with tf.variable_scope('Condition1'):
cond = apply_bias_act(dense_layer(cond, fmaps=1), act='sigmoid')
scale = (s_latents / 2. + 2.) * cond
tt_00 = scale
tt_01 = tf.zeros_like(scale)
tt_02 = tf.zeros_like(scale)
tt_10 = tf.zeros_like(scale)
tt_11 = scale
tt_12 = tf.zeros_like(scale)
theta = tf.concat([tt_00, tt_01, tt_02, tt_10, tt_11, tt_12], axis=1)
return theta
# Return shear matrix
def get_sh_matrix(sh_latents, cond_latent):
# sh_latents[:, 0]: [-2., 2.] -> [-1., 1.]
# sh_latents[:, 1]: [-2., 2.] -> [-1., 1.]
with tf.variable_scope('Condition0x'):
cond_x = apply_bias_act(dense_layer(cond_latent, fmaps=128),
act=act)
with tf.variable_scope('Condition1x'):
cond_x = apply_bias_act(dense_layer(cond_x, fmaps=1),
act='sigmoid')
with tf.variable_scope('Condition0y'):
cond_y = apply_bias_act(dense_layer(cond_latent, fmaps=128),
act=act)
with tf.variable_scope('Condition1y'):
cond_y = apply_bias_act(dense_layer(cond_y, fmaps=1),
act='sigmoid')
cond = tf.concat([cond_x, cond_y], axis=1)
xy_shear = sh_latents / 2. * cond
tt_00 = tf.ones_like(xy_shear[:, 0:1])
tt_01 = xy_shear[:, 0:1]
tt_02 = tf.zeros_like(xy_shear[:, 0:1])
tt_10 = xy_shear[:, 1:]
tt_11 = tf.ones_like(xy_shear[:, 1:])
tt_12 = tf.zeros_like(xy_shear[:, 1:])
theta = tf.concat([tt_00, tt_01, tt_02, tt_10, tt_11, tt_12], axis=1)
return theta
# Return translation matrix
def get_t_matrix(t_latents, cond_latent):
# t_latents[:, 0]: [-2., 2.] -> [-0.5, 0.5]
# t_latents[:, 1]: [-2., 2.] -> [-0.5, 0.5]
with tf.variable_scope('Condition0x'):
cond_x = apply_bias_act(dense_layer(cond_latent, fmaps=128),
act=act)
with tf.variable_scope('Condition1x'):
cond_x = apply_bias_act(dense_layer(cond_x, fmaps=1),
act='sigmoid')
with tf.variable_scope('Condition0y'):
cond_y = apply_bias_act(dense_layer(cond_latent, fmaps=128),
act=act)
with tf.variable_scope('Condition1y'):
cond_y = apply_bias_act(dense_layer(cond_y, fmaps=1),
act='sigmoid')
cond = tf.concat([cond_x, cond_y], axis=1)
xy_shift = t_latents / 4. * cond
tt_00 = tf.ones_like(xy_shift[:, 0:1])
tt_01 = tf.zeros_like(xy_shift[:, 0:1])
tt_02 = xy_shift[:, 0:1]
tt_10 = tf.zeros_like(xy_shift[:, 1:])
tt_11 = tf.ones_like(xy_shift[:, 1:])
tt_12 = xy_shift[:, 1:]
theta = tf.concat([tt_00, tt_01, tt_02, tt_10, tt_11, tt_12], axis=1)
return theta
# Apply spatial transform
def apply_st(x, st_matrix, idx, up=True): # idx: 2, 3, 4
with tf.variable_scope('Transform'):
x = tf.transpose(x, [0, 2, 3, 1]) # NCHW -> NHWC
x = transformer(x, st_matrix, out_dims=x.shape.as_list()[1:3])
x = tf.transpose(x, [0, 3, 1, 2]) # NHWC -> NCHW
with tf.variable_scope('Upconv'):
x = apply_bias_act(conv2d_layer(x,
fmaps=nf(idx),
kernel=3,
up=up,
resample_kernel=resample_kernel),
act=act)
with tf.variable_scope('Conv'):
x = apply_bias_act(conv2d_layer(x, fmaps=nf(idx), kernel=3),
act=act)
return x
def upsample(y):
with tf.variable_scope('Upsample'):
return upsample_2d(y, k=resample_kernel)
def torgb(x, y):
with tf.variable_scope('ToRGB'):
t = apply_bias_act(conv2d_layer(x, fmaps=num_channels, kernel=1))
return t if y is None else y + t
# Early layers.
y = None
with tf.variable_scope('4x4'):
with tf.variable_scope('Const'):
x = tf.get_variable('const',
shape=[1, nf(1), 4, 4],
initializer=tf.initializers.random_normal())
x = tf.tile(tf.cast(x, dtype), [tf.shape(dlatents_in)[0], 1, 1, 1])
with tf.variable_scope('Upconv8x8'):
x = apply_bias_act(conv2d_layer(x,
fmaps=nf(1),
kernel=3,
up=True,
resample_kernel=resample_kernel),
act=act)
with tf.variable_scope('Conv0'):
x = apply_bias_act(conv2d_layer(x, fmaps=nf(1), kernel=3), act=act)
with tf.variable_scope('ModulatedConv'):
x = apply_bias_act(modulated_conv2d_layer(
x,
dlatents_in[:, :n_cat],
fmaps=nf(2),
kernel=3,
up=False,
resample_kernel=resample_kernel,
fused_modconv=fused_modconv),
act=act)
with tf.variable_scope('Conv1'):
x = apply_bias_act(conv2d_layer(x, fmaps=nf(2), kernel=3), act=act)
# Rotation layers.
with tf.variable_scope('16x16'):
r_matrix = get_r_matrix(dlatents_in[:, n_cat:n_cat + 1],
dlatents_in[:, :n_cat])
x = apply_st(x, r_matrix, 2)
# Scaling layers.
with tf.variable_scope('32x32'):
s_matrix = get_s_matrix(dlatents_in[:, n_cat + 1:n_cat + 2],
dlatents_in[:, :n_cat])
x = apply_st(x, s_matrix, 3)
# Shearing layers.
with tf.variable_scope('32x32_Shear'):
sh_matrix = get_sh_matrix(dlatents_in[:, n_cat + 2:n_cat + 4],
dlatents_in[:, :n_cat])
x = apply_st(x, sh_matrix, 3, up=False)
# Translation layers.
with tf.variable_scope('64x64'):
t_matrix = get_t_matrix(dlatents_in[:, n_cat + 4:],
dlatents_in[:, :n_cat])
x = apply_st(x, t_matrix, 4)
y = torgb(x, y)
# # Tail layers.
# for res in range(6, resolution_log2 + 1):
# with tf.variable_scope('%dx%d' % (res * 2, res * 2)):
# x = apply_bias_act(conv2d_layer(x,
# fmaps=nf(res),
# kernel=1,
# up=True,
# resample_kernel=resample_kernel),
# act=act)
# if architecture == 'skip':
# y = upsample(y)
# if architecture == 'skip' or res == resolution_log2:
# y = torgb(x, y)
images_out = y
assert images_out.dtype == tf.as_dtype(dtype)
return tf.identity(images_out, name='images_out')
def G_synthesis_simple_dsp(
dlatents_in, # Input: Disentangled latents (W) [minibatch, dlatent_size].
dlatent_size=7, # Disentangled latent (W) dimensionality. Including discrete info, rotation, scaling, and xy translation.
D_global_size=3, # Discrete latents.
sb_C_global_size=4, # Continuous latents.
label_size=0, # Label dimensionality, 0 if no labels.
num_channels=1, # Number of output color channels.
nonlinearity='relu', # Activation function: 'relu', 'lrelu', etc.
dtype='float32', # Data type to use for activations and outputs.
resample_kernel=[
1, 3, 3, 1
], # Low-pass filter to apply when resampling activations. None = no filtering.
fused_modconv=True, # Implement modulated_conv2d_layer() as a single fused op?
**_kwargs): # Ignore unrecognized keyword args.
act = nonlinearity
images_out = None
# Primary inputs.
assert dlatent_size == D_global_size + sb_C_global_size
n_cat = label_size + D_global_size
dlatents_in.set_shape([None, label_size + dlatent_size])
dlatents_in = tf.cast(dlatents_in, dtype)
# Return rotation matrix
def get_r_matrix(r_latents, cond_latent):
# r_latents: [-2., 2.] -> [0, 2*pi]
with tf.variable_scope('Condition0'):
cond = apply_bias_act(dense_layer(cond_latent, fmaps=128), act=act)
with tf.variable_scope('Condition1'):
cond = apply_bias_act(dense_layer(cond, fmaps=1), act='sigmoid')
rad = (r_latents + 2) / 4. * 2. * np.pi
rad = rad * cond
tt_00 = tf.math.cos(rad)
tt_01 = -tf.math.sin(rad)
tt_02 = tf.zeros_like(rad)
tt_10 = tf.math.sin(rad)
tt_11 = tf.math.cos(rad)
tt_12 = tf.zeros_like(rad)
theta = tf.concat([tt_00, tt_01, tt_02, tt_10, tt_11, tt_12], axis=1)
return theta
# Return scaling matrix
def get_s_matrix(s_latents, cond_latent):
# s_latents: [-2., 2.] -> [1, 3]
with tf.variable_scope('Condition0'):
cond = apply_bias_act(dense_layer(cond_latent, fmaps=128), act=act)
with tf.variable_scope('Condition1'):
cond = apply_bias_act(dense_layer(cond, fmaps=1), act='sigmoid')
scale = (s_latents / 2. + 2.) * cond
tt_00 = scale
tt_01 = tf.zeros_like(scale)
tt_02 = tf.zeros_like(scale)
tt_10 = tf.zeros_like(scale)
tt_11 = scale
tt_12 = tf.zeros_like(scale)
theta = tf.concat([tt_00, tt_01, tt_02, tt_10, tt_11, tt_12], axis=1)
return theta
# Return translation matrix
def get_t_matrix(t_latents, cond_latent):
# t_latents[:, 0]: [-2., 2.] -> [-0.5, 0.5]
# t_latents[:, 1]: [-2., 2.] -> [-0.5, 0.5]
with tf.variable_scope('Condition0'):
cond = apply_bias_act(dense_layer(cond_latent, fmaps=128), act=act)
with tf.variable_scope('Condition1'):
cond = apply_bias_act(dense_layer(cond, fmaps=1), act='sigmoid')
xy_shift = t_latents / 4. * cond
tt_00 = tf.ones_like(xy_shift[:, 0:1])
tt_01 = tf.zeros_like(xy_shift[:, 0:1])
tt_02 = xy_shift[:, 0:1]
tt_10 = tf.zeros_like(xy_shift[:, 1:])
tt_11 = tf.ones_like(xy_shift[:, 1:])
tt_12 = xy_shift[:, 1:]
theta = tf.concat([tt_00, tt_01, tt_02, tt_10, tt_11, tt_12], axis=1)
return theta
# Apply spatial transform
def apply_st(x, st_matrix):
with tf.variable_scope('Transform'):
x = tf.transpose(x, [0, 2, 3, 1]) # NCHW -> NHWC
x = transformer(x, st_matrix, out_dims=x.shape.as_list()[1:3])
x = tf.transpose(x, [0, 3, 1, 2]) # NHWC -> NCHW
return x
def torgb(x, y):
with tf.variable_scope('ToRGB'):
t = apply_bias_act(conv2d_layer(x, fmaps=num_channels, kernel=1))
return t if y is None else y + t
# Early layers.
y = None
with tf.variable_scope('4x4'):
with tf.variable_scope('Const'):
x = tf.get_variable('const',
shape=[1, 64, 4, 4],
initializer=tf.initializers.random_normal())
x = tf.tile(tf.cast(x, dtype), [tf.shape(dlatents_in)[0], 1, 1, 1])
with tf.variable_scope('4x4Conv'):
w = get_weight([3, 3, x.shape[1].value, 64])
x = tf.nn.conv2d(x,
tf.cast(w, x.dtype),
data_format='NCHW',
strides=[1, 1, 1, 1],
padding='SAME')
x = apply_bias_act(x, act=act)
with tf.variable_scope('8x8ModulatedConv'):
x = apply_bias_act(modulated_conv2d_layer(
x,
dlatents_in[:, :n_cat],
fmaps=64,
kernel=3,
up=True,
resample_kernel=resample_kernel,
fused_modconv=fused_modconv),
act=act)
with tf.variable_scope('16x16'):
w = get_weight([4, 4, x.shape[1].value, 32])
# Transpose weights.
w = tf.transpose(w, [0, 1, 3, 2])
x = tf.nn.conv2d_transpose(
x,
w,
output_shape=[tf.shape(dlatents_in)[0], 32, 16, 16],
strides=[1, 1, 2, 2],
padding='SAME',
data_format='NCHW')
x = apply_bias_act(x, act=act)
with tf.variable_scope('rotation'):
r_matrix = get_r_matrix(dlatents_in[:, n_cat:n_cat + 1],
dlatents_in[:, :n_cat])
x = apply_st(x, r_matrix)
with tf.variable_scope('scale'):
s_matrix = get_s_matrix(dlatents_in[:, n_cat + 1:n_cat + 2],
dlatents_in[:, :n_cat])
x = apply_st(x, s_matrix)
with tf.variable_scope('translation'):
t_matrix = get_t_matrix(dlatents_in[:, n_cat + 2:],
dlatents_in[:, :n_cat])
x = apply_st(x, t_matrix)
with tf.variable_scope('32x32'):
w = get_weight([4, 4, x.shape[1].value, 32])
# Transpose weights.
w = tf.transpose(w, [0, 1, 3, 2])
x = tf.nn.conv2d_transpose(
x,
w,
output_shape=[tf.shape(dlatents_in)[0], 32, 32, 32],
strides=[1, 1, 2, 2],
padding='SAME',
data_format='NCHW')
x = apply_bias_act(x, act=act)
with tf.variable_scope('64x64'):
w = get_weight([4, 4, x.shape[1].value, 1])
# Transpose weights.
w = tf.transpose(w, [0, 1, 3, 2])
x = tf.nn.conv2d_transpose(
x,
w,
output_shape=[tf.shape(dlatents_in)[0], 1, 64, 64],
strides=[1, 1, 2, 2],
padding='SAME',
data_format='NCHW')
x = apply_bias_act(x)
images_out = x
assert images_out.dtype == tf.as_dtype(dtype)
return tf.identity(images_out, name='images_out')
