def test_mul6(self):
x = d.Variable(array([[3, 1], [2, 1]]), 'x')
y = d.Variable(array([4]), 'y')
z = x * y
t = z.forward()
z.backward(array([[1, 2], [1, 2]]))
self.assertTrue(arr_equal(t, array([[12, 4], [8, 4]])))
self.assertTrue(arr_equal(x.grad, array([[4, 8], [4, 8]])))
self.assertTrue(arr_equal(y.grad, array([9])))
def test_mul4(self):
x = d.Variable(array([3, 1]), 'x')
y = d.Variable(array([4]), 'y')
z = y * x
t = z.forward()
z.backward(array([1, 2]))
self.assertTrue(arr_equal(t, array([12, 4])))
self.assertTrue(arr_equal(x.grad, array([4, 8])))
self.assertTrue(arr_equal(y.grad, array([5])))
def test_mul1(self):
x = d.Variable(array([3, 1]), 'x')
y = d.Variable(array([4, 6]), 'y')
z = x * y
t = z.forward()
z.backward(array([1, 1]))
self.assertTrue(arr_equal(t, array([12, 6])))
self.assertTrue(arr_equal(x.grad, array([4, 6])))
self.assertTrue(arr_equal(y.grad, array([3, 1])))
def test_sub_broadcast6(self):
x = d.Variable(array([[2.0, 3.0], [5.0, 6.0]]), 'x')
y = d.Variable(array([6.0]), 'y')
z = x - y
t = z.forward()
self.assertTrue(arr_equal(t, array([[-4, -3], [-1, 0]])))
z.backward(array([[1.0, 1.0], [1, 1]]))
self.assertTrue(arr_equal(x.grad, array([[1.0, 1.0], [1, 1]])))
self.assertTrue(arr_equal(y.grad, -array([4])))
def test_mul8(self):
x = d.Variable(array([[3, 1], [2, 1]]), 'x')
y = d.Variable(array([4, 5]), 'y')
z = y * x
t = z.forward()
z.backward(array([[1, 2], [1, 2]]))
self.assertTrue(arr_equal(t, array([[12, 5], [8, 5]])))
self.assertTrue(arr_equal(x.grad, array([[4, 10], [4, 10]])))
self.assertTrue(arr_equal(y.grad, array([5, 4])))
def test_div4(self):
y = d.Variable(np.array([4, 2]), 'x')
x = d.Variable(np.array([2]), 'y')
z = x / y
t = z.forward()
self.assertTrue(arr_equal(t, array([0.5, 1])))
z.backward(np.array([2, 1]))
self.assertTrue(arr_equal(x.grad, array([1])))
self.assertTrue(arr_equal(y.grad, array([-0.25, -0.5])))
def test_sub_broadcast5(self):
x = d.Variable(array([[2.0, 3.0], [5.0, 6.0]]), 'x')
y = d.Variable(array([6.0]), 'y')
z = y - x
t = z.forward()
self.assertTrue(arr_equal(t, array([[4, 3], [1, 0]])))
z.backward(array([[1.0, 1.0], [2, 1]]))
self.assertTrue(arr_equal(x.grad, -array([[1.0, 1.0], [2, 1]])))
self.assertTrue(arr_equal(y.grad, array([5])))
def test_sub_broadcast1(self):
x = d.Variable(array([2.0, 3.0]), 'x')
y = d.Variable(array([6.0]), 'y')
z = x - y
t = z.forward()
self.assertTrue(arr_equal(t, array([-4, -3])))
z.backward(array([1.0, 2.0]))
self.assertTrue(arr_equal(x.grad, array([1, 2])))
self.assertTrue(arr_equal(y.grad, array([-3])))
def test_add_broadcast6(self):
x = d.Variable(array([[2.0, 3.0], [5.0, 6.0]]), 'x')
y = d.Variable(array([6.0]), 'y')
z = y + x
t = z.forward()
self.assertTrue(arr_equal(t, array([[8, 9], [11, 12]])))
z.backward(array([[1.0, 1.0], [2, 1]]))
self.assertTrue(arr_equal(x.grad, array([[1.0, 1.0], [2, 1]])))
self.assertTrue(arr_equal(y.grad, array([5])))
def test_sub(self):
x = d.Variable(array([2.0, 3.0]), 'x')
y = d.Variable(array([6.0, 7.0]), 'y')
z = x - y
t = z.forward()
self.assertTrue(arr_equal(t, array([-4, -4])))
z.backward(array([2.0, 1.0]))
self.assertTrue(arr_equal(x.grad, array([2, 1])))
self.assertTrue(arr_equal(y.grad, -array([2, 1])))
def test_div7(self):
x = d.Variable(np.array([[4, 2], [4, 2]]), 'x')
y = d.Variable(np.array([2, 2]), 'y')
z = x / y
t = z.forward()
self.assertTrue(arr_equal(t, array([[2, 1], [2, 1]])))
z.backward(np.array([[2, 1], [2, 1]]))
self.assertTrue(arr_equal(x.grad, array([[1, 0.5], [1, 0.5]])))
self.assertTrue(arr_equal(y.grad, array([-4, -1])))
def test_add_broadcast3(self):
x = d.Variable(array([2.0, 3.0]), 'x')
y = d.Variable(array([6.0]), 'y')
z = y + x
t = z.forward()
self.assertTrue(arr_equal(t, array([8, 9])))
z.backward(array([1.0, 1.0]))
self.assertTrue(arr_equal(x.grad, array([1, 1])))
self.assertTrue(arr_equal(y.grad, array([2])))
def test_min1(self):
x = d.Variable(array([[1, 2], [-1, -2]]), 'x')
y = d.Variable(array([[2, 1], [0, -3]]), 'y')
z = d.min(x, y)
t = z.forward()
self.assertTrue(arr_equal(t, array([[1, 1], [-1, -3]])))
z.backward(array([[2, 2], [2, 2]]))
self.assertTrue(arr_equal(y.grad, array([[0, 2], [0, 2]])))
self.assertTrue(arr_equal(x.grad, array([[2, 0], [2, 0]])))
def test_div8(self):
x = d.Variable(np.array([[4, 2], [4, 2]]), 'x')
y = d.Variable(np.array([2, 2]), 'y')
z = y / x
t = z.forward()
self.assertTrue(arr_equal(t, array([[0.5, 1], [0.5, 1]])))
z.backward(np.array([[2, 1], [2, 1]]))
self.assertTrue(arr_equal(x.grad, array([[-0.25, -0.5], [-0.25, -0.5]])))
self.assertTrue(arr_equal(y.grad, array([1, 1])))
def test_add(self):
x = d.Variable(array([2.0, 3.0]), 'x')
y = d.Variable(array([6.0, 7.0]), 'y')
z = x + y
t = z.forward()
self.assertTrue(arr_equal(t, array([8, 10])))
z.backward(array([2.0, 2.0]))
self.assertTrue(arr_equal(x.grad, array([2, 2])))
self.assertTrue(arr_equal(y.grad, array([2, 2])))
def test_div2(self):
x = d.Variable(np.array([4, 8]), 'x')
y = d.Variable(np.array([2, 4]), 'y')
z = y / x
t = z.forward()
self.assertTrue(arr_equal(t, array([0.5, 0.5])))
z.backward(np.array([2, 1]))
self.assertTrue(arr_equal(x.grad, array([-0.25, -1 / 16])))
self.assertTrue(arr_equal(y.grad, array([0.5, 0.125])))
def test_div1(self):
x = d.Variable(np.array([4]), 'x')
y = d.Variable(np.array([2]), 'y')
z = x / y
t = z.forward()
self.assertTrue(arr_equal(t, array([2])))
z.backward(np.array([2]))
self.assertTrue(arr_equal(x.grad, array([1])))
self.assertTrue(arr_equal(y.grad, array([-2])))
def test_mmul2(self):
w_ = np.array([2, 3, 4])
b_ = np.array([3, 4, 5])
w = d.Variable(w_, 'w')
b = d.Variable(b_, 'b')
z = w @ b
t = z.forward()
self.assertTrue(arr_equal(t, array(38)))
z.backward(array([2]))
self.assertTrue(arr_equal(w.grad, array([6, 8, 10])))
self.assertTrue(arr_equal(b.grad, array([4, 6, 8])))
def test_mmul1(self):
w_ = np.array([[2, 3, 4], [1, 2, 3]])
b_ = np.array([3, 4, 5])
w = d.Variable(w_, 'w')
b = d.Variable(b_, 'b')
z = w @ b
t = z.forward()
self.assertTrue(arr_equal(t, array([38, 26])))
z.backward(array([2, 1]))
self.assertTrue(arr_equal(w.grad, array([[6, 8, 10], [3, 4, 5]])))
self.assertTrue(arr_equal(b.grad, array([5, 8, 11])))
def test_min2(self):
x = d.Variable(array([[1, 2], [-1, -2]]), 'x')
z = d.min(0, x)
t = z.forward()
self.assertTrue(arr_equal(t, array([[0, 0], [-1, -2]])))
z.backward(array([[2, 2], [2, 2]]))
self.assertTrue(arr_equal(x.grad, array([[0, 0], [2, 2]])))
def test_div9(self):
x = d.Variable(np.array([1, 2]), 'x')
z = 2 / x
t = z.forward()
self.assertTrue(arr_equal(t, array([2, 1])))
z.backward(np.array([1, 2]))
self.assertTrue(arr_equal(x.grad, array([-2, -1])))
x1 = np.random.randn(N, DIM)
x2 = np.random.randn(N, DIM)
y1 = np.ones(N)
y2 = -np.ones(N)
x1 += np.array([1, 1])
x2 += np.array([-1, -1])
X = np.concatenate((x1, x2), axis=0)
Y = np.concatenate((y1, y2))
data = dllib.Placeholder((2 * N, 2))
data.feed(X)
label = dllib.Placeholder((2 * N, ))
label.feed(Y)
w = dllib.Variable(np.array([0.0, 0.0]), 'w')
b = dllib.Variable(np.array([0.0]), 'b')
t = data @ w + b
loss = logistic_loss_with_logits(t, label)
optimizer = GradientDescent(loss, 0.01)
optimizer.optimize(500)
pred = np.sign(t.forward())
acc = pred[((pred - Y) == 0)].size / (2 * N)
print("accuracy is: {0}".format(acc))
plt.plot(X[:N, 0], X[:N, 1], 'r*')
plt.plot(X[N:2 * N, 0], X[N:2 * N, 1], 'g*')
plt.show()
