def __call__(self, s): # sは入力文章の単語id配列
accum_loss = None
# 入力単語をword2vecで縮約
v, k = self.embed.W.data.shape
h = Variable(np.zeros((1, k), dtype=np.float32)) # 履歴入力
c = Variable(np.zeros((1, k), dtype=np.float32)) # 入力
# 注目単語の次の単語を訓練データとして学習
for i in range(len(s)):
next_w_id = eos_is if (i == len(s) - 1) else s[i + 1]
# 訓練データ
tx = Variable(np.array([next_w_id], dtype=np.int32))
# 入力文のうちの単語[i]
x_k = self.embed(Variable(np.array([s[i]], dtype=np.int32)))
# RNN層の計算
z0 = self.Wz(x_k) + self.Rz(h)
z1 = F.tan(z0)
# Input層の計算
i0 = self.Wi(x_k) + self.Ri(F.dropout(h))
i1 = F.sigmoid(i0)
# Forget層の計算
f0 = self.Wf(x_k) + slef.Rf(F.dropout(h))
f1 = F.sigmoid(f0)
# MemoryCellに渡す値の計算
c = i1 * z1 + f1 * c
# LSTM出力層の計算
o0 = slef.Wo(x, k) + self.Ro(h)
o1 = F.sigmoid(o0)
# 次の時刻に渡す履歴計算
h = o1 * F.tanh(c)
loss = F.softmax_cross_entropy(self.W(h), tx)
accum_loss = loss if accum_loss is None else accum_loss + loss
return accum_loss
def perspective(vertices, angle=None):
assert (vertices.ndim == 3)
if angle is None:
xp = chainer.cuda.get_array_module(vertices)
angle = xp.array([xp.radians(60)] * vertices.shape[0], 'float32')
width = cf.tan(angle / 2)
width = cf.broadcast_to(width[:, None], vertices.shape[:2])
z = vertices[:, :, 2]
x = vertices[:, :, 0] / z / width
y = vertices[:, :, 1] / z / width
vertices = cf.concat((x[:, :, None], y[:, :, None], z[:, :, None]), axis=2)
return vertices
def get_sample(self, mu, sigma, number_samples):
"""
One line description
Parameters
----------
Returns
-------
"""
mu, sigma = broadcast_and_squeeze(mu, sigma)
sample = mu + sigma*F.tan(np.pi*np.random.uniform(0,1,size=mu.shape).astype(np.float32))
return sample
def perspective(vertices, angle=30.):
assert (vertices.ndim == 3)
xp = chainer.cuda.get_array_module(vertices)
if isinstance(angle, float) or isinstance(angle, int):
angle = chainer.Variable(xp.array(angle, 'float32'))
angle = angle / 180. * 3.1416
angle = cf.broadcast_to(angle[None], (vertices.shape[0],))
width = cf.tan(angle)
width = cf.broadcast_to(width[:, None], vertices.shape[:2])
z = vertices[:, :, 2]
x = vertices[:, :, 0] / z / width
y = vertices[:, :, 1] / z / width
vertices = cf.concat((x[:, :, None], y[:, :, None], z[:, :, None]), axis=2)
return vertices
def perspective(vertices, angle=30.):
assert (vertices.ndim == 3)
xp = chainer.cuda.get_array_module(vertices)
if isinstance(angle, float) or isinstance(angle, int):
angle = chainer.Variable(xp.array(angle, 'float32'))
angle = angle / 180. * 3.1416
angle = cf.broadcast_to(angle[None], (vertices.shape[0], ))
width = cf.tan(angle)
width = cf.broadcast_to(width[:, None], vertices.shape[:2])
z = vertices[:, :, 2]
x = vertices[:, :, 0] / z / width
y = vertices[:, :, 1] / z / width
vertices = cf.concat((x[:, :, None], y[:, :, None], z[:, :, None]), axis=2)
return vertices
def __call__(self, x):
xTan = self.dilatedConvTan(x)
xTan = self.batchNormTan(x)
xTan = F.tan(xTan)
xSig = self.dilatedConvSig(x)
xSig = self.batchNormSig(x)
xSig = F.sigmoid(xSig)
h = xTan * xSig
skip = self.conv(h)
skip = self.batchNormCnn(h)
s = x.shape
xp = chainer.cuda.cupy if self.useGPU else np
zero = xp.zeros((s[0], s[1], s[2], s[3] - skip.shape[3]), dtype='f')
skip = F.concat((skip, zero), axis=3)
next = skip + x
return next, skip
def shoot(self):
W = self.width
H = self.height
origin = self.origin
xaxis = self.xaxis
zaxis = self.zaxis
yaxis = self.yaxis
angle = self.fov
t0 = self.t0
t1 = self.t1
xp = chainer.backend.get_array_module(origin)
#zaxis = zaxis.reshape((1, 3, 1, 1))
#yaxis = yaxis.reshape((1, 3, 1, 1))
#xaxis = vnorm(vcross(yaxis, zaxis))
#yaxis = vnorm(vcross(zaxis, xaxis))
xaxis = vnorm(xaxis.reshape((1, 3, 1, 1))).reshape((1, 1, 3))
yaxis = vnorm(yaxis.reshape((1, 3, 1, 1))).reshape((1, 1, 3))
zaxis = vnorm(zaxis.reshape((1, 3, 1, 1))).reshape((1, 1, 3))
#print(angle.shape)
angle = (angle / 2) * math.pi / 180.0
HH = F.tan(angle)
ro = origin
ro = F.tile(ro, (H, W, 1))
ro = F.transpose(ro, (2, 0, 1))
yy = xp.tile(xp.arange(H, dtype=np.float32).reshape((H, 1, 1)), (1, W, 1)) #(H, W, 1)
xx = xp.tile(xp.arange(W, dtype=np.float32).reshape((1, W, 1)), (H, 1, 1)) #(H, W, 1)
yy = (1 - 2 * (yy + 0.5) / H) * HH
xx = (2 * (xx + 0.5) / W - 1) * HH
rd = xx * xaxis + yy * yaxis + F.broadcast_to(zaxis, (H, W, 3))
rd = F.transpose(rd, (2, 0, 1))
rd = rd.reshape((1, 3, H, W))
rd = vnorm(rd)
rd = rd.reshape((3, H, W))
t0 = F.broadcast_to(t0.reshape((1, 1, 1)), (1, H, W))
t1 = F.broadcast_to(t1.reshape((1, 1, 1)), (1, H, W))
#print(ro.shape, ro.dtype)
#print(rd.shape, rd.dtype)
return ro, rd, t0, t1
def tan(self, x):
return F.tan(x)
def forward(self, x):
y1 = F.tan(x)
return y1