def test_forward(self):
val = np.random.random((1, 2, 3))
x = ad.variable(val)
y = x[0, 1, 0]
actual = y.forward()
expect = val[0, 1, 0]
self.assertEqual(expect.shape, y.shape)
self.assertTrue(np.allclose(expect, actual), (expect, actual))
y = x[0]
actual = y.forward()
expect = val[0]
self.assertEqual(expect.shape, y.shape)
self.assertTrue(np.allclose(expect, actual), (expect, actual))
y = x[0, 1:]
actual = y.forward()
expect = val[0, 1:]
self.assertEqual(expect.shape, y.shape)
self.assertTrue(np.allclose(expect, actual), (expect, actual))
y = x[0, 1:, :2]
actual = y.forward()
expect = val[0, 1:, :2]
self.assertEqual(expect.shape, y.shape)
self.assertTrue(np.allclose(expect, actual), (expect, actual))
y = x[0, :, :2:]
actual = y.forward()
expect = val[0, :, :2:]
self.assertEqual(expect.shape, y.shape)
self.assertTrue(np.allclose(expect, actual), (expect, actual))
y = x[0, :-2:-1, ::-2]
actual = y.forward()
expect = val[0, :-2:-1, ::-2]
self.assertEqual(expect.shape, y.shape)
self.assertTrue(np.allclose(expect, actual), (expect, actual))
x = ad.placeholder(shape=(None, 12))
y = x[:, 3:5]
self.assertEqual((None, 2), y.shape)
x = ad.placeholder(shape=(None, 12))
y = x[1:, 3:5]
self.assertEqual((None, 2), y.shape)
x = ad.placeholder(shape=(None, 12))
y = x[4:6, 3:5]
self.assertEqual(
(None, 2),
y.shape) # Because the first dimension may be less than 6
def test_forward(self):
x = ad.placeholder(shape=(2, 3))
feed_dict = {x: np.random.random((2, 3))}
actual = ad.Session().run(x, feed_dict=feed_dict)
expect = feed_dict[x]
self.assertTrue(np.allclose(expect, actual), (expect, actual))
feed_dict = {x: np.random.random((2, 3))}
actual = ad.Session().run(x, feed_dict=feed_dict)
expect = feed_dict[x]
self.assertTrue(np.allclose(expect, actual), (expect, actual))
def gen_linear_model(config: dict, verbose=False):
"""Generate a linear model.
:param config: Configuration.
:param verbose: Print loss and gradients if it is True.
:return: Model, loss, placeholders and variables.
"""
x = ad.placeholder(shape=(None, config['input_len']), name='X')
y = ad.placeholder(shape=(None, ), name='Y')
w1 = ad.variable(
initializer=ad.inits.random_normal(),
shape=(config['input_len'], config['hidden_dim']),
name='W1',
)
b1 = ad.variable(
initializer=ad.inits.zeros,
shape=config['hidden_dim'],
name='b1',
)
v = ad.acts.leaky_relu(ad.dot(x, w1) + b1)
w2 = ad.variable(
initializer=ad.inits.random_normal(),
shape=(config['hidden_dim'], 2),
name='W2',
)
b2 = ad.variable(
initializer=ad.inits.zeros,
shape=2,
name='b2',
)
y_pred = ad.acts.softmax(ad.dot(v, w2) + b2)
loss = ad.square(y - y_pred).mean()
if verbose:
print('Loss:', loss)
return y_pred, loss, [x, y], [w1, b1, w2, b2]
def gen_linear_model(config: dict, verbose=False):
"""Generate a linear model.
:param config: Configuration.
:param verbose: Print loss and gradients if it is True.
:return: Model, loss, placeholders and variables.
"""
x = ad.placeholder(shape=(None, config['input_len']), name='X')
y = ad.placeholder(shape=(None, ), name='Y')
w = ad.variable(
initializer=ad.inits.random_normal(),
shape=config['input_len'],
name='W',
)
b = ad.variable(0.0, name='b')
y_pred = ad.dot(x, w) + b
loss = ad.square(y - y_pred).mean()
if verbose:
print('Loss:', loss)
return y_pred, loss, [x, y], [w, b]
def test_forward(self):
val = np.random.random((3, 5))
x = ad.variable(val)
y = ad.pad(x, 2)
actual = y.forward()
expect = np.pad(val, 2, mode='constant')
self.assertEqual(expect.shape, y.shape)
self.assertTrue(np.allclose(expect, actual), (expect, actual))
val = np.random.random((3, 5))
x = ad.variable(val)
y = ad.pad(x, (1, 2))
actual = y.forward()
expect = np.pad(val, (1, 2), mode='constant')
self.assertEqual(expect.shape, y.shape)
self.assertTrue(np.allclose(expect, actual), (expect, actual))
val = np.random.random((3, 5))
x = ad.variable(val)
y = ad.pad(x, ((1,), (2,)))
actual = y.forward()
expect = np.pad(val, ((1,), (2,)), mode='constant')
self.assertEqual(expect.shape, y.shape)
self.assertTrue(np.allclose(expect, actual), (expect, actual))
val = np.random.random((3, 5))
x = ad.variable(val)
y = ad.pad(x, ((1, 2), (3, 4)))
actual = y.forward()
expect = np.pad(val, ((1, 2), (3, 4)), mode='constant')
self.assertEqual(expect.shape, y.shape)
self.assertTrue(np.allclose(expect, actual), (expect, actual))
val = np.random.random((3, 5))
x = ad.placeholder(shape=(None, 5))
y = ad.pad(x, 1)
actual = y.forward({x: val})
expect = np.pad(val, 1, mode='constant')
self.assertEqual((None, 7), y.shape)
self.assertTrue(np.allclose(expect, actual), (expect, actual))
def test_backward(self):
ad.placeholder(shape=(2, 3)).backward()
def __init__(self, shape, **kwargs):
super(Input, self).__init__(**kwargs)
self.shape = shape
self.placeholder = ad.placeholder(shape)
self._outputs = self.call(None)
self._output_shapes = self.compute_output_shape(None)
def test_forward_variable_shape(self):
x = ad.placeholder(shape=(None, 3))
y = ad.random(ad.shape(x))
z = y.forward({x: np.random.random((5, 3))})
self.assertEqual((5, 3), z.shape)