def __call__(self, w, x=None, add_noise=False):
h = x
batch_size, _ = w.shape
if self.upsample:
assert h is not None
if self.blur_k is None:
k = np.asarray([1, 2, 1]).astype('f')
k = k[:, None] * k[None, :]
k = k / np.sum(k)
self.blur_k = self.xp.asarray(k)[None, None, :]
if self.enable_blur:
h = blur(upscale2x(h), self.blur_k)
else:
h = upscale2x(h)
h = self.c0(h)
else:
h = F.broadcast_to(self.W, (batch_size, self.ch_in, 4, 4))
# h should be (batch, ch, size, size)
if add_noise:
h = self.n0(h)
h = F.leaky_relu(self.b0(h))
h = self.s0(w, h)
h = self.c1(h)
if add_noise:
h = self.n1(h)
h = F.leaky_relu(self.b1(h))
h = self.s1(w, h)
return h
def forward(self, x, add_noise=False):
h = x
if self.blur_k is None:
k = np.asarray([1, 2, 1]).astype('f')
k = k[:, None] * k[None, :]
k = k / np.sum(k)
self.blur_k = self.xp.asarray(k)[None, None, :]
if self.enable_blur:
h = blur(upscale2x(h), self.blur_k)
else:
h = upscale2x(h)
h = self.c0(h)
# h should be (batch, ch, size, size)
if add_noise:
h = self.n0(h)
h = F.leaky_relu(self.b0(h))
h = F.normalize(h)
h = self.c1(h)
if add_noise:
h = self.n1(h)
h = F.leaky_relu(self.b1(h))
h = F.normalize(h)
return h
def __call__(self, h, w, stage):
h1 = F.leaky_relu(self.c0(h))
h1 = self.s0(w, h1)
h2 = F.leaky_relu(self.c1(h1))
h2 = self.s1(w, h2)
h3 = F.leaky_relu(self.c4(h2))
h3 = self.s4(w, h3)
h3 = upscale2x(h3)
h3 = F.leaky_relu(self.c5(h3))
h3 = F.concat([self.s5(w, h3), h1])
h3 = upscale2x(h3)
h3 = F.leaky_relu(self.c6(h3))
h3 = F.concat([self.s6(w, h3), h])
h = self.c7(h3)
return h
def __call__(self, w, stage, add_noise=True, w2=None):
'''
for alpha in [0, 1), and 2*k+2 + alpha < self.max_stage (-1 <= k <= ...):
stage 0 + alpha : z -> block[0] -> out[0] * 1
stage 2*k+1 + alpha : z -> ... -> block[k] -> (up -> out[k]) * (1 - alpha)
.................... -> (block[k+1] -> out[k+1]) * (alpha)
stage 2*k+2 + alpha : z -> ............... -> (block[k+1] -> out[k+1]) * 1
over flow stages continues.
'''
stage = min(stage, self.max_stage - 1e-8)
alpha = stage - math.floor(stage)
stage = math.floor(stage)
h = None
if stage % 2 == 0:
k = (stage - 2) // 2
# Enable Style Mixing:
if w2 is not None and k >= 0:
lim = np.random.randint(1, k+2)
else:
lim = k+2
for i in range(0, (k + 1) + 1): # 0 .. k+1
if i == lim:
w = w2
h = self.blocks[i](w, x=h, add_noise=add_noise)
h = self.outs[k + 1](h)
else:
k = (stage - 1) // 2
if w2 is not None and k >= 1:
lim = np.random.randint(1, k+1)
else:
lim = k+1
for i in range(0, k + 1): # 0 .. k
if i == lim:
w = w2
h = self.blocks[i](w, x=h, add_noise=add_noise)
h_0 = self.outs[k](upscale2x(h))
h_1 = self.outs[k + 1](self.blocks[k + 1](w, x=h, add_noise=add_noise))
assert 0. <= alpha < 1.
h = (1.0 - alpha) * h_0 + alpha * h_1
if chainer.configuration.config.train:
return h
else:
min_sample_image_size = 64
if h.data.shape[2] < min_sample_image_size: # too small
scale = int(min_sample_image_size // h.data.shape[2])
return F.unpooling_2d(h, scale, scale, 0, outsize=(min_sample_image_size, min_sample_image_size))
else:
return h
def forward(self,
z,
stage,
theta=None,
style_mixing_rate=None,
add_noise=True,
return_feature=False):
'''
for alpha in [0, 1), and 2*k+2 + alpha < self.max_stage (-1 <= k <= ...):
stage 0 + alpha : z -> block[0] -> out[0] * 1
stage 2*k+1 + alpha : z -> ... -> block[k] -> (up -> out[k]) * (1 - alpha)
.................... -> (block[k+1] -> out[k+1]) * (alpha)
stage 2*k+2 + alpha : z -> ............... -> (block[k+1] -> out[k+1]) * 1
over flow stages continues.
'''
# theta: (batchsize, )
# if self.rgbd:
add_noise = False
stage = min(stage, self.max_stage - 1e-8)
alpha = stage - math.floor(stage)
stage = math.floor(stage)
if self.rgbd and theta is None:
assert False, "theta is None"
if self.rgbd:
h = F.concat([z, theta * 10])
else:
h = z
h = self.linear(h).reshape(z.shape[0], self.ch, 4, 4)
if stage % 2 == 0:
k = (stage - 2) // 2
for i in range(0, (k + 1)): # 0 .. k+1
h = self.blocks[i](x=h, add_noise=add_noise)
if return_feature and i == 2:
feat = h
h = self.outs[k](h)
else:
k = (stage - 1) // 2
for i in range(0, k): # 0 .. k
h = self.blocks[i](x=h, add_noise=add_noise)
if return_feature and i == 2:
feat = h
h_0 = upscale2x(self.outs[k - 1](h))
h_1 = self.outs[k](self.blocks[k](x=h, add_noise=add_noise))
assert 0. <= alpha < 1.
h = (1.0 - alpha) * h_0 + alpha * h_1
if self.rgbd:
# inverse depth
depth = 1 / (F.softplus(h[:, -1:]) + 1e-4)
# print(depth.array.mean(), depth.array.std())
h = h[:, :3]
h = F.concat([h, depth])
if chainer.configuration.config.train:
if return_feature:
return h, feat
else:
return h
else:
min_sample_image_size = 64
if h.data.shape[2] < min_sample_image_size: # too small
scale = int(min_sample_image_size // h.data.shape[2])
return F.unpooling_2d(h,
scale,
scale,
0,
outsize=(min_sample_image_size,
min_sample_image_size))
else:
return h
def forward(self,
w,
w2,
stage,
theta=None,
add_noise=True,
return_feature=False):
'''
for alpha in [0, 1), and 2*k+2 + alpha < self.max_stage (-1 <= k <= ...):
stage 0 + alpha : z -> block[0] -> out[0] * 1
stage 2*k+1 + alpha : z -> ... -> block[k] -> (up -> out[k]) * (1 - alpha)
.................... -> (block[k+1] -> out[k+1]) * (alpha)
stage 2*k+2 + alpha : z -> ............... -> (block[k+1] -> out[k+1]) * 1
over flow stages continues.
'''
# theta: (batchsize, )
# if self.rgbd:
add_noise = False
stage = min(stage, self.max_stage - 1e-8)
alpha = stage - math.floor(stage)
stage = math.floor(stage)
if self.rgbd and theta is None:
assert False, "theta is None"
h = None
if stage % 2 == 0:
k = (stage - 2) // 2
for i in range(0, (k + 1) + 1): # 0 .. k+1
if i == 3: # resolution 32~
w = w2
if self.rgbd and i < 2:
if self.rotate_conv_input:
_w = self.w_from_theta(theta)
else:
_w = self.rotate_w(w, theta)
else:
_w = w
h = self.blocks[i](_w, x=h, add_noise=add_noise)
if return_feature and i == 3:
feat = h
h = self.outs[k + 1](h)
else:
k = (stage - 1) // 2
for i in range(0, k + 1): # 0 .. k
if i == 3:
w = w2
if self.rgbd and i < 2:
if self.rotate_conv_input:
_w = self.w_from_theta(theta)
else:
_w = self.rotate_w(w, theta)
else:
_w = w
h = self.blocks[i](_w, x=h, add_noise=add_noise)
if return_feature and i == 3:
feat = h
h_0 = upscale2x(self.outs[k](h))
h_1 = self.outs[k + 1](self.blocks[k + 1](w,
x=h,
add_noise=add_noise))
assert 0. <= alpha < 1.
h = (1.0 - alpha) * h_0 + alpha * h_1
if self.rgbd:
# inverse depth
depth = 1 / (F.softplus(h[:, -1:]) + 1e-4)
# print(depth.array.mean(), depth.array.std())
h = h[:, :3]
h = F.concat([h, depth])
if chainer.configuration.config.train:
if return_feature:
return h, feat
else:
return h
else:
min_sample_image_size = 64
if h.data.shape[2] < min_sample_image_size: # too small
scale = int(min_sample_image_size // h.data.shape[2])
return F.unpooling_2d(h,
scale,
scale,
0,
outsize=(min_sample_image_size,
min_sample_image_size))
else:
return h
