def test_train(self):
np.random.seed(0)
ws = list()
ws.append(Variable(np.random.rand(2, 3)))
ws.append(Variable(np.random.rand(3, 1)))
data = Variable(np.array([[0, 0], [0, 1], [1, 0], [1, 1]]))
target = Variable(np.array([[0], [1], [0], [1]]))
expected = [
6.210704901174886, 1.4434549528818301, 0.809644434846779,
0.7585458232291216, 0.7437140298400363, 0.7316218334889659,
0.7198939685708763, 0.708293823629362, 0.6967362454336858,
0.6851547179015602
]
log = []
for i in range(10):
t = matmul(data, ws[0])
t = sigmoid2(t)
pred = matmul(t, ws[1])
diff = (pred - target)
loss = sum(diff * diff)
loss.backward()
for w in ws:
w.data -= w.grad * 0.1
w.cleargrad()
#print(loss.data)
log.append(loss.data)
self.assertTrue(np.allclose(np.array(log), np.array(expected)))
def test_forward(self):
x = Variable(np.random.rand(10))
y = F.sum(x)
expected = np.sum(x.data)
self.assertTrue(np.allclose(y.data, expected))
def test_backward3(self):
x_data = np.random.rand(2, 3)
f = lambda x: F.sum(x, axis=1)
check_backward(f, x_data, testcase=self)
def test_backward3(self):
x_data = np.random.rand(10, 20, 20)
f = lambda x: F.sum(x, axis=(0, 2), keepdims=True)
check_backward(f, x_data, testcase=self)
def test_backward(self):
x_data = np.random.rand(10)
f = lambda x: F.sum(x)
check_backward(f, x_data, testcase=self)
def test_forward2(self):
x = Variable(np.random.rand(10, 20, 30))
y = F.sum(x, axis=1)
expected = np.sum(x.data, axis=1)
self.assertTrue(np.allclose(y.data, expected))
def mean_squared_error(x, t):
diff = x - t
total = sum(diff * diff)
N = x.data.shape[0]
return total / N
def training_loss(weights):
# Training loss is the negative log-likelihood of the training labels.
preds = logistic_predictions(weights, inputs)
label_probabilities = preds * targets + (1 - preds) * (1 - targets)
return -sum(log(label_probabilities))
def logprob(model, x, t, noise_scale=0.1):
noise_scale = np.array(noise_scale)
pred = model(x)
logp = D.Normal(t, noise_scale).log_prob(pred)
return -1 * F.sum(logp)
def log_gaussian(model, scale):
t = 0
normal = D.Normal(0, scale)
for p in model.params():
t += F.sum(normal.log_prob(p))
return -1 * t