def check_forward(self, h_data):
h = chainer.Variable(h_data)
loss = functions.decov(h)
self.assertEqual(loss.data.shape, ())
self.assertEqual(loss.data.dtype, numpy.float32)
loss_value = float(loss.data)
# Compute expected value
h_data = cuda.to_cpu(h_data)
h_mean = h_data.mean(axis=0)
N = h_data.shape[0]
loss_expect = 0
for i in six.moves.range(h_data.shape[1]):
for j in six.moves.range(h_data.shape[1]):
ij_loss = 0.
if i != j:
for n in six.moves.range(N):
ij_loss += (h_data[n, i] - h_mean[i]) * (h_data[n, j] -
h_mean[j])
ij_loss /= N
loss_expect += ij_loss**2
loss_expect *= 0.5
self.assertAlmostEqual(loss_expect, loss_value, places=5)
def check_forward(self, h_data):
h = chainer.Variable(h_data)
loss = functions.decov(h)
self.assertEqual(loss.data.shape, ())
self.assertEqual(loss.data.dtype, numpy.float32)
loss_value = float(loss.data)
# Compute expected value
h_data = cuda.to_cpu(h_data)
h_mean = h_data.mean(axis=0)
N = h_data.shape[0]
loss_expect = 0
for i in six.moves.range(h_data.shape[1]):
for j in six.moves.range(h_data.shape[1]):
ij_loss = 0.
if i != j:
for n in six.moves.range(N):
ij_loss += (h_data[n, i] - h_mean[i]) * (
h_data[n, j] - h_mean[j])
ij_loss /= N
loss_expect += ij_loss ** 2
loss_expect *= 0.5
self.assertAlmostEqual(loss_expect, loss_value, places=5)
def check_forward(self, h_data):
h = chainer.Variable(h_data)
loss = functions.decov(h, self.reduce)
self.assertEqual(loss.shape, self.gloss.shape)
self.assertEqual(loss.data.dtype, self.dtype)
loss_value = cuda.to_cpu(loss.data)
# Compute expected value
h_data = cuda.to_cpu(h_data)
loss_expect = _deconv(h_data)
if self.reduce == 'half_squared_sum':
loss_expect = (loss_expect ** 2).sum() * 0.5
numpy.testing.assert_allclose(
loss_expect, loss_value, **self.forward_options)
def check_forward(self, h_data):
h = chainer.Variable(h_data)
loss = functions.decov(h, self.reduce)
self.assertEqual(loss.shape, self.gloss.shape)
self.assertEqual(loss.data.dtype, self.dtype)
loss_value = cuda.to_cpu(loss.data)
# Compute expected value
h_data = cuda.to_cpu(h_data)
loss_expect = _deconv(h_data)
if self.reduce == 'half_squared_sum':
loss_expect = (loss_expect**2).sum() * 0.5
numpy.testing.assert_allclose(loss_expect, loss_value,
**self.forward_options)
def check_type(self, h_data, gloss_data):
h = chainer.Variable(h_data)
loss = functions.decov(h, self.reduce)
loss.grad = gloss_data
loss.backward()
self.assertEqual(h_data.dtype, h.grad.dtype)
def f(h):
return functions.decov(h, self.reduce)
def check_invalid_option(self, xp):
h = xp.asarray(self.h)
with self.assertRaises(ValueError):
functions.decov(h, 'invalid_option')
def check_type(self, h_data):
h = chainer.Variable(h_data)
loss = functions.decov(h)
loss.backward()
self.assertEqual(h_data.dtype, h.grad.dtype)
def check_invalid_option(self, xp):
h = xp.asarray(self.h)
with self.assertRaises(ValueError):
functions.decov(h, 'invalid_option')
def forward(self, inputs, device):
h, = inputs
loss = functions.decov(h, self.reduce)
return loss,
def f(h):
return functions.decov(h, self.reduce)