def __init__(self, x_shape, z_dim, batch_norm=True):
super().__init__()
ch = x_shape[-1]
self.net = ts.Sequential(
dense(128),
ts.ReLU(),
dense(4 * 4 * 64),
ts.ReLU(),
ts.Reshape((-1, 4, 4, 64)),
deconv(64, 4, 2, "same"),
ts.LeakyReLU(),
deconv(32, 4, 2, "same"),
ts.LeakyReLU(),
deconv(32, 4, 2, "same"),
ts.LeakyReLU(),
deconv(ch, 4, 2, "same"),
ts.Sigmoid(),
)
# Add batchnorm post-activation (attach to activation out_hook)
if batch_norm:
self.net.apply(add_bn, targets=(ts.ReLU, ts.LeakyReLU))
ut.log("Building generator...")
self.build((1, z_dim))
self.apply(ut.reset_parameters)
def __init__(self, x_shape, y_dim, width=1, share_dense=False,
uncond_bias=False):
super().__init__()
self.y_dim = y_dim
self.body = ts.Sequential(
conv(32 * width, 4, 2, "same"), ts.LeakyReLU(),
conv(32 * width, 4, 2, "same"), ts.LeakyReLU(),
conv(64 * width, 4, 2, "same"), ts.LeakyReLU(),
conv(64 * width, 4, 2, "same"), ts.LeakyReLU(),
ts.Flatten(),
)
self.aux = ts.Sequential(
dense(128 * width), ts.LeakyReLU(),
)
if share_dense:
self.body.append(dense(128 * width), ts.LeakyReLU())
self.aux.append(dense(128 * width), ts.LeakyReLU())
self.head = ts.Sequential(
dense(128 * width), ts.LeakyReLU(),
dense(128 * width), ts.LeakyReLU(),
dense(1, bias=uncond_bias)
)
for m in (self.body, self.aux, self.head):
m.apply(ts.SpectralNorm.add, targets=ts.Affine)
ut.log("Building label discriminator...")
x_shape, y_shape = [1] + x_shape, (1, y_dim)
self.build(x_shape, y_shape)
self.apply(ut.reset_parameters)
def __init__(self,
x_shape,
y_dim,
width=1,
share_dense=False,
uncond_bias=False,
cond_bias=False,
mask_type="match"):
super().__init__()
self.y_dim = y_dim
self.mask_type = mask_type
self.body = ts.Sequential(
conv(32 * width, 4, 2, "same"),
ts.LeakyReLU(),
conv(32 * width, 4, 2, "same"),
ts.LeakyReLU(),
conv(64 * width, 4, 2, "same"),
ts.LeakyReLU(),
conv(64 * width, 4, 2, "same"),
ts.LeakyReLU(),
ts.Flatten(),
)
if share_dense:
self.body.append(dense(128 * width), ts.LeakyReLU())
if mask_type == "match":
self.neck = ts.Sequential(
dense(128 * width),
ts.LeakyReLU(),
dense(128 * width),
ts.LeakyReLU(),
)
self.head_uncond = dense(1, bias=uncond_bias)
self.head_cond = dense(128 * width, bias=cond_bias)
for m in (self.body, self.neck, self.head_uncond):
m.apply(ts.SpectralNorm.add, targets=ts.Affine)
add_wn(self.head_cond)
x_shape, y_shape = [1] + x_shape, ((1, ), tf.int32)
elif mask_type == "rank":
self.body.append(dense(128 * width), ts.LeakyReLU(),
dense(128 * width), ts.LeakyReLU(),
dense(1 + y_dim, bias=uncond_bias))
self.body.apply(ts.SpectralNorm.add, targets=ts.Affine)
x_shape, y_shape = [1] + x_shape, (1, y_dim)
ut.log("Building {} discriminator...".format(mask_type))
self.build(x_shape, x_shape, y_shape)
self.apply(ut.reset_parameters)
def __init__(self, x_shape, z_dim, width=1, spectral_norm=True):
super().__init__()
self.net = ts.Sequential(
conv(32 * width, 4, 2, "same"), ts.LeakyReLU(),
conv(32 * width, 4, 2, "same"), ts.LeakyReLU(),
conv(64 * width, 4, 2, "same"), ts.LeakyReLU(),
conv(64 * width, 4, 2, "same"), ts.LeakyReLU(),
ts.Flatten(),
dense(128 * width), ts.LeakyReLU(),
dense(2 * z_dim)
)
if spectral_norm:
self.net.apply(ts.SpectralNorm.add, targets=ts.Affine)
ut.log("Building encoder...")
self.build([1] + x_shape)
self.apply(ut.reset_parameters)