def enc_up(x,
c,
init=False,
dropout_p=0.5,
n_scales=1,
n_residual_blocks=2,
activation="elu",
n_filters=64,
max_filters=128):
with model_arg_scope(init=init, dropout_p=dropout_p,
activation=activation):
# outputs
hs = []
# prepare input
# 这一行也很奇怪, 为什么要把x和c连起来呢?
# xc = tf.concat([x,c], axis = -1)
xc = x
h = nn.nin(xc, n_filters)
for l in range(n_scales):
# level module
for i in range(n_residual_blocks):
h = nn.residual_block(h)
hs.append(h)
# prepare input to next level
if l + 1 < n_scales:
# 似乎它这个channel一直都是128, 没有增长过.
n_filters = min(2 * n_filters, max_filters)
h = nn.downsample(h, n_filters)
return hs
def enc_up(
x,
c,
init=False,
dropout_p=0.5,
n_scales=1,
n_residual_blocks=2,
activation="elu",
n_filters=64,
max_filters=128,
):
with model_arg_scope(init=init, dropout_p=dropout_p,
activation=activation):
"""c is actually not used"""
# outputs
hs = []
# prepare input
# xc = tf.concat([x,c], axis = -1)
xc = x
h = nn.nin(xc, n_filters)
for l in range(n_scales):
# level module
for i in range(n_residual_blocks):
h = nn.residual_block(h)
hs.append(h)
# prepare input to next level
if l + 1 < n_scales:
n_filters = min(2 * n_filters, max_filters)
h = nn.downsample(h, n_filters)
return hs
def dec_up(
c,
init=False,
dropout_p=0.5,
n_scales=1,
n_residual_blocks=2,
activation="elu",
n_filters=64,
max_filters=128,
):
with model_arg_scope(init=init, dropout_p=dropout_p,
activation=activation):
# outputs
hs = []
# prepare input
h = nn.nin(c, n_filters)
for l in range(n_scales):
# level module
for i in range(n_residual_blocks):
h = nn.residual_block(h)
hs.append(h)
# prepare input to next level
if l + 1 < n_scales:
n_filters = min(2 * n_filters, max_filters)
h = nn.downsample(h, n_filters)
return hs
def dec_up(c,
init=False,
dropout_p=0.5,
n_scales=1,
n_residual_blocks=2,
activation="elu",
n_filters=64,
max_filters=128):
with model_arg_scope(init=init, dropout_p=dropout_p,
activation=activation):
hs = []
h = nn.nin(c, n_filters)
for l in range(n_scales):
for i in range(n_residual_blocks):
h = nn.residual_block(h)
hs.append(h)
if l + 1 < n_scales:
n_filters = min(2 * n_filters, max_filters)
h = nn.downsample(h, n_filters)
return hs
def cfn(
x, init = False, dropout_p = 0.5,
n_scales = 1, n_residual_blocks = 2, activation = "elu", n_filters = 64, max_filters = 128):
with model_arg_scope(
init = init, dropout_p = dropout_p, activation = activation):
# outputs
hs = []
# prepare input
xc = x
h = nn.nin(xc, n_filters)
for l in range(n_scales):
# level module
for i in range(n_residual_blocks):
h = nn.residual_block(h)
hs.append(h)
# prepare input to next level
if l + 1 < n_scales:
n_filters = min(2*n_filters, max_filters)
h = nn.downsample(h, n_filters)
h_shape = h.shape.as_list()
h = tf.reshape(h, [h_shape[0],1,1,h_shape[1]*h_shape[2]*h_shape[3]])
h = nn.nin(h, 2*max_filters)
hs.append(h)
return hs
