def test_crossentropy():
X = np.random.rand(100, 10).astype(np.float32)
O = np.random.rand(100, 10).astype(np.float32)
X /= X.sum(1)[:, None]
O /= O.sum(1)[:, None]
Y_expected = -np.sum(X * np.log(O)) / X.shape[0]
rtol=1e-4
Y = np.empty_like(X)
Yhr = op.cross_entropy(X, O)
assert_allclose(Y_expected, Yhr, err_msg="CPU, no target", rtol=rtol)
Xd = op.to_gpu(X)
Od = op.to_gpu(O)
Yd = op.cross_entropy(Xd, Od)
assert_allclose(Y_expected, op.to_cpu(Yd), err_msg="GPU, no target", rtol=rtol)
def _get_loss(self, target, pred):
if self.loss == "crossentropy":
op.streams[0].synchronize()
return op.cross_entropy(target, pred, stream=op.streams[3])
elif self.loss == "squarederror":
return op.mean_squared_error(target, pred)
else:
raise NotImplementedError()
def _get_loss(self, target, pred):
if self.loss == "crossentropy":
op.streams[0].synchronize()
return op.cross_entropy(target, pred, stream=op.streams[3])
elif self.loss == "squarederror":
return op.mean_squared_error(target, pred)
else:
raise NotImplementedError()
def test_crossentropy():
X = np.random.rand(100, 10).astype(np.float32)
O = np.random.rand(100, 10).astype(np.float32)
X /= X.sum(1)[:, None]
O /= O.sum(1)[:, None]
Y_expected = -np.sum(X * np.log(O)) / X.shape[0]
rtol = 1e-4
Y = np.empty_like(X)
Yhr = op.cross_entropy(X, O)
assert_allclose(Y_expected, Yhr, err_msg="CPU, no target", rtol=rtol)
Xd = op.to_gpu(X)
Od = op.to_gpu(O)
Yd = op.cross_entropy(Xd, Od)
assert_allclose(Y_expected,
op.to_cpu(Yd),
err_msg="GPU, no target",
rtol=rtol)
def cross_entropy_loss(w, x, y, net):
unserialize(w, net)
a = net.forward_pass(x)
retval = op.cross_entropy(y, a)
return retval
def cross_entropy_loss(w, x, y, net):
unserialize(w, net)
a = net.forward_pass(x)
retval = op.cross_entropy(y, a)
return retval
