def test_forward(self):
x_val = np.random.random((3, 4))
x = ad.variable(x_val)
y = ad.shape(x)
actual = y.forward()
expect = np.array([3, 4])
self.assertEqual(expect.shape, y.shape)
self.assertTrue(np.allclose(expect, actual), (expect, actual))
def call(self, inputs, **kwargs):
if 0.0 < self.rate < 1.0:
if self.noise_shape is not None:
noise_shape = self.noise_shape
else:
noise_shape = ad.shape(inputs)
return ad.where(ad.in_train_phase(),
inputs * (ad.random(noise_shape) > self.rate),
inputs)
return inputs
def call(self, inputs, **kwargs):
padded = ad.pad(inputs, ((0, ), (self.pad_width[0], ),
(self.pad_width[1], ), (0, )))
batch_size = ad.shape(inputs)[0]
reshaped = ad.map_fn(lambda i: self.call_batch(padded, i),
ad.arange(batch_size))
y = ad.dot(reshaped, self.w)
if self.use_bias:
y += self.b
if self.activation is not None:
y = self.activation(y)
return y
def call(self, inputs, **kwargs):
initial_val = ad.dot(
ad.zeros_like(inputs)[:, 0, :],
ad.zeros_like(self.wx[:, :self.units]))
outputs = ad.while_loop(
lambda body_inputs: ad.less(body_inputs[0],
ad.shape(inputs)[1]),
lambda x: self.step(inputs, x),
[ad.variable(0.0), initial_val, initial_val],
output_index=-1,
)
if self.return_sequences:
return outputs.transpose(axes=[1, 0, 2])
return outputs[-1]
def test_backward(self):
x_val = np.random.random((3, 4))
x = ad.variable(x_val)
y = ad.shape(x)
self.numeric_gradient_check(y, {}, [x])
def call_batch(self, padded: ad.Operation, i: int):
height = ad.shape(padded)[1]
new_height = (height -
self.dilated_kernel_size[0]) // self.strides[0] + 1
return ad.map_fn(lambda r: self.call_row(padded, i, r),
ad.arange(new_height))
def call_row(self, padded: ad.Operation, i: int, r: int):
width = ad.shape(padded)[2]
new_width = (width -
self.dilated_kernel_size[1]) // self.strides[1] + 1
return ad.map_fn(lambda c: self.call_column(padded, i, r, c),
ad.arange(new_width))
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)