def test_answer_gpu_cpu(self):
x = cuda.to_gpu(self.x)
y = F.batch_det(chainer.Variable(x))
gpu = cuda.to_cpu(y.data)
if self.dtype == numpy.float16:
cpu = numpy.linalg.det(self.x.astype(numpy.float32)).astype(
numpy.float16)
testing.assert_allclose(gpu, cpu, atol=5e-3, rtol=5e-3)
else:
cpu = numpy.linalg.det(self.x)
testing.assert_allclose(gpu, cpu)
def test_answer_gpu_cpu(self):
x = cuda.to_gpu(self.x)
y = F.batch_det(chainer.Variable(x))
gpu = cuda.to_cpu(y.data)
if self.dtype == numpy.float16:
cpu = numpy.linalg.det(
self.x.astype(numpy.float32)).astype(numpy.float16)
testing.assert_allclose(gpu, cpu, atol=5e-3, rtol=5e-3)
else:
cpu = numpy.linalg.det(self.x)
testing.assert_allclose(gpu, cpu)
def test_invalid_shape(self):
with self.assertRaises(type_check.InvalidType):
F.batch_det(chainer.Variable(numpy.zeros((1, 2))))
def test_answer_gpu_cpu(self):
x = cuda.to_gpu(self.x)
y = F.batch_det(chainer.Variable(x))
gpu = cuda.to_cpu(y.data)
cpu = numpy.linalg.det(self.x)
testing.assert_allclose(gpu, cpu)
def test_invalid_shape(self):
with self.assertRaises(type_check.InvalidType):
F.batch_det(chainer.Variable(numpy.zeros((1, 2))))
def test_answer_gpu_cpu(self):
x = cuda.to_gpu(self.x)
y = F.batch_det(chainer.Variable(x))
gpu = cuda.to_cpu(y.data)
cpu = numpy.linalg.det(self.x)
gradient_check.assert_allclose(gpu, cpu)
def fwd(self, x, isTraining=False, gpu=-1):
if gpu >= 0:
xp = cuda.cupy
else:
xp = np
# Compression Network
# エンコード
zc = self.Encoder(x)
# デコード
y = self.Decoder(zc)
# 再構築誤差を計算 relativeユークリッド距離とコサイン類似度
rEucDist = self.relativeEuclideanDistance(x, y)
CosSim = self.cosineSimilarity(x, y)
#潜在変数と再構築誤差を合体
z = F.concat((zc, rEucDist, CosSim), axis=1)
# Estimation Network
gamma = self.Estimation(z)
NumOfData, NumOfClass = gamma.shape
_, zDim = z.shape
# GMM
# 各サンプルの各分布への帰属確率から、各分布の混合比、平均ベクトル、共分散行列を得る
if chainer.config.train:
phi = self._phi(gamma)
mu = self._mu(z, gamma)
sigma = self._sigma(z, gamma, mu)
os.makedirs(self.dirGMMparameters, exist_ok=True)
if chainer.cuda.available:
np.savetxt(self.dirGMMparameters + "phi.csv",
np.array((phi.data).get()),
delimiter=",")
np.savetxt(self.dirGMMparameters + "mu.csv",
np.array((mu.data).get()),
delimiter=",")
np.savetxt(self.dirGMMparameters + "sigma.csv",
np.array(
(sigma.data).get()).reshape(NumOfClass, -1),
delimiter=",")
else:
np.savetxt(self.dirGMMparameters + "phi.csv",
np.array(phi.data),
delimiter=",")
np.savetxt(self.dirGMMparameters + "mu.csv",
np.array(mu.data),
delimiter=",")
np.savetxt(self.dirGMMparameters + "sigma.csv",
np.array(sigma.data).reshape(NumOfClass, -1),
delimiter=",")
else:
phi = Variable(
xp.array(
np.loadtxt(self.dirGMMparameters + "phi.csv",
delimiter=",").astype(np.float32)))
mu = Variable(
xp.array(
np.loadtxt(self.dirGMMparameters + "mu.csv",
delimiter=",").astype(np.float32)))
sigma = Variable(
xp.array((np.loadtxt(self.dirGMMparameters + "sigma.csv",
delimiter=",")).reshape(
NumOfClass, zDim,
zDim).astype(np.float32)))
# エネルギーを計算
# eps = 1e-3 #ランク落ちまたは、行列式が0になってしまう対策
# sigma = sigma + Variable(xp.array(np.array(list(np.eye(zDim))*NumOfClass).reshape(NumOfClass,zDim,zDim).astype(np.float32)) * eps)
sigmaInv = F.batch_inv(sigma) # shape(3D) -> NumOfClass, zDim, zDim
z_broadcast = F.broadcast_to(
z, (NumOfClass, NumOfData,
zDim)) # shape(3D) -> NumOfClass, NumOfData, zDim
mu_broadcast = F.transpose(F.broadcast_to(
mu, (NumOfData, NumOfClass, zDim)),
axes=(1, 0, 2)) # shape(3D) -> NumOfClass,
sa = z_broadcast - mu_broadcast
listForEnr1 = [
F.matmul(sa[i], sigmaInv[i]) for i in range(NumOfClass)
] # shape(3D) -> NumOfClass, NumOfData, zDim
listForEnr2 = [
F.sum(listForEnr1[i] * sa[i], axis=1) for i in range(NumOfClass)
] # shape(2D) -> NumOfClass, NumOfData
varForEnr = F.stack(
[listForEnr2[i] for i in range(len(listForEnr2))],
axis=0) # リストからVariableへ変換 # shape(2D) -> NumOfClass, NumOfData
numer = F.exp(-(1 / 2) *
varForEnr) # 分子の計算 # shape(2D) -> NumOfClass, NumOfData
denom = F.transpose(
F.broadcast_to(
F.sqrt(F.batch_det(2 * math.pi * sigma)),
(NumOfData,
NumOfClass))) # 分母の計算 # shape(2D) -> NumOfClass, NumOfData
phi_broadcast = F.transpose(
F.broadcast_to(
phi,
(NumOfData, NumOfClass))) # shape(2D) -> NumOfClass, NumOfData
energy = -1 * F.log(
F.sum(phi_broadcast * (numer / denom), axis=0,
keepdims=True)) # shape(2D) -> 1, NumOfData
energy = F.transpose(energy) # shape(2D) -> NumOfData,1
if isTraining:
return y, energy, sigma
else:
return z, energy, gamma, y
