def _multiple_ops_in_middle(): inputs = keras.Input(shape=(10,)) x = keras.layers.Dense(10)(inputs) x = gen_nn_ops.relu(x, name='hey') x = gen_nn_ops.relu(x, name='hey2') outputs = keras.layers.Dense(10)(x) return inputs, outputs
def test_numerical_correctness_simple(self): x = ops.convert_to_tensor([[-1., 0., -2., 1.]]) inputs = keras.Input(shape=(4,)) outputs = gen_nn_ops.relu(inputs) model = keras.Model(inputs, outputs) y = self.evaluate(model(x)) self.assertAllClose(y, [[0., 0., 0., 1.]])
def test_built(self):
inputs = keras.Input(shape=(10,))
outputs = gen_nn_ops.relu(inputs)
model = keras.Model(inputs, outputs)
model.compile('sgd', 'mse')
for layer in model.layers:
self.assertTrue(layer.built)
# Test something that requires Layers to be built.
model.summary()
def _reuse_op(): inputs = keras.Input(shape=(10,)) # This op needs to be checked multiple times. x = gen_nn_ops.relu(inputs) y = keras.layers.Dense(10)(x) x2 = x * 2 y2 = keras.layers.Dense(10)(x2) outputs = y + y2 return inputs, outputs
def test_serialization(self): x = ops.convert_to_tensor([-1., 0., -2., 1.]) inputs = keras.Input(shape=(4,)) outputs = gen_nn_ops.relu(inputs) model1 = keras.Model(inputs, outputs) y1 = self.evaluate(model1(x)) model2 = model1.from_config(model1.get_config()) y2 = self.evaluate(model2(x)) self.assertAllClose(y1, y2)
def _construct_graph_of_size(size):
start = time.time()
x = keras.backend.placeholder(shape=(10, 4))
for _ in range(size):
x = keras.layers.Dense(4)(x)
x = gen_nn_ops.relu(x)
end = time.time()
return end - start
def per_image_whitening(image):
"""Linearly scales `image` to have zero mean and unit norm.
This op computes `(x - mean) / adjusted_stddev`, where `mean` is the average
of all values in image, and
`adjusted_stddev = max(stddev, 1.0/sqrt(image.NumElements()))`.
`stddev` is the standard deviation of all values in `image`. It is capped
away from zero to protect against division by 0 when handling uniform images.
Note that this implementation is limited:
* It only whitens based on the statistics of an individual image.
* It does not take into account the covariance structure.
Args:
image: 3-D tensor of shape `[height, width, channels]`.
Returns:
The whitened image with same shape as `image`.
Raises:
ValueError: if the shape of 'image' is incompatible with this function.
"""
image = ops.convert_to_tensor(image, name='image')
_Check3DImage(image, require_static=False)
num_pixels = math_ops.reduce_prod(array_ops.shape(image))
image = math_ops.cast(image, dtype=dtypes.float32)
image_mean = math_ops.reduce_mean(image)
variance = (math_ops.reduce_mean(math_ops.square(image)) -
math_ops.square(image_mean))
variance = gen_nn_ops.relu(variance)
stddev = math_ops.sqrt(variance)
# Apply a minimum normalization that protects us against uniform images.
min_stddev = math_ops.inv(
math_ops.sqrt(math_ops.cast(num_pixels, dtypes.float32)))
pixel_value_scale = math_ops.maximum(stddev, min_stddev)
pixel_value_offset = image_mean
image = math_ops.sub(image, pixel_value_offset)
image = math_ops.div(image, pixel_value_scale)
return image
def per_image_whitening(image):
"""Linearly scales `image` to have zero mean and unit norm.
This op computes `(x - mean) / adjusted_stddev`, where `mean` is the average
of all values in image, and
`adjusted_stddev = max(stddev, 1.0/sqrt(image.NumElements()))`.
`stddev` is the standard deviation of all values in `image`. It is capped
away from zero to protect against division by 0 when handling uniform images.
Note that this implementation is limited:
* It only whitens based on the statistics of an individual image.
* It does not take into account the covariance structure.
Args:
image: 3-D tensor of shape `[height, width, channels]`.
Returns:
The whitened image with same shape as `image`.
Raises:
ValueError: if the shape of 'image' is incompatible with this function.
"""
image = ops.convert_to_tensor(image, name='image')
_Check3DImage(image, require_static=False)
num_pixels = math_ops.reduce_prod(array_ops.shape(image))
image = math_ops.cast(image, dtype=dtypes.float32)
image_mean = math_ops.reduce_mean(image)
variance = (math_ops.reduce_mean(math_ops.square(image)) -
math_ops.square(image_mean))
variance = gen_nn_ops.relu(variance)
stddev = math_ops.sqrt(variance)
# Apply a minimum normalization that protects us against uniform images.
min_stddev = math_ops.inv(
math_ops.sqrt(math_ops.cast(num_pixels, dtypes.float32)))
pixel_value_scale = math_ops.maximum(stddev, min_stddev)
pixel_value_offset = image_mean
image = math_ops.sub(image, pixel_value_offset)
image = math_ops.div(image, pixel_value_scale)
return image
def test_gradient_tape_in_function(self):
z = keras.Input((1,))
x = math_ops.matmul(z, constant_op.constant(2.0, shape=(1, 1)))
x = math_ops.reduce_mean(x, axis=0, keepdims=True)
h = gen_nn_ops.relu(x)
m = keras.Model(z, h)
@def_function.function()
def f(x):
with backprop.GradientTape() as t:
t.watch(x)
z = m(x ** 2)
grads = t.gradient(z, x)
return grads
self.assertAllEqual(f(constant_op.constant(10.0, shape=(1, 1))),
constant_op.constant(40.0, shape=(1, 1)))
f = def_function.function(f)
self.assertAllEqual(f(constant_op.constant(10.0, shape=(1, 1))),
constant_op.constant(40.0, shape=(1, 1)))
def test_gradient_tape_in_function(self):
z = keras.Input((1,))
x = math_ops.matmul(z, constant_op.constant(2.0, shape=(1, 1)))
x = math_ops.reduce_mean(x, axis=0, keepdims=True)
h = gen_nn_ops.relu(x)
m = keras.Model(z, h)
@def_function.function()
def f(x):
with backprop.GradientTape() as t:
t.watch(x)
z = m(x ** 2)
grads = t.gradient(z, x)
return grads
self.assertAllEqual(f(constant_op.constant(10.0, shape=(1, 1))),
constant_op.constant(40.0, shape=(1, 1)))
f = def_function.function(f)
self.assertAllEqual(f(constant_op.constant(10.0, shape=(1, 1))),
constant_op.constant(40.0, shape=(1, 1)))
def per_image_standardization(image):
"""Linearly scales `image` to have zero mean and unit variance.
This op computes `(x - mean) / adjusted_stddev`, where `mean` is the average
of all values in image, and
`adjusted_stddev = max(stddev, 1.0/sqrt(image.NumElements()))`.
`stddev` is the standard deviation of all values in `image`. It is capped
away from zero to protect against division by 0 when handling uniform images.
Args:
image: An n-D Tensor where the last 3 dimensions are
`[height, width, channels]`.
Returns:
The standardized image with same shape as `image`.
Raises:
ValueError: if the shape of 'image' is incompatible with this function.
"""
with ops.name_scope(None, 'per_image_standardization', [image]) as scope:
image = ops.convert_to_tensor(image, name='image')
num_pixels = math_ops.reduce_prod(array_ops.shape(image)[1:4])
image = math_ops.cast(image, dtype=dtypes.float32)
image_mean = math_ops.reduce_mean(image,
axis=[-1, -2, -3],
keepdims=True)
variance = (math_ops.reduce_mean(
math_ops.square(image), axis=[-1, -2, -3], keepdims=True) -
math_ops.square(image_mean))
variance = gen_nn_ops.relu(variance)
stddev = math_ops.sqrt(variance)
# Apply a minimum normalization that protects us against uniform images.
min_stddev = math_ops.rsqrt(math_ops.cast(num_pixels, dtypes.float32))
pixel_value_scale = math_ops.maximum(stddev, min_stddev)
pixel_value_offset = image_mean
image = math_ops.subtract(image, pixel_value_offset)
image = math_ops.div(image, pixel_value_scale, name=scope)
return image
def _single_op_in_middle(): inputs = keras.Input(shape=(10,)) x = keras.layers.Dense(10)(inputs) x = gen_nn_ops.relu(x) outputs = keras.layers.Dense(10)(x) return keras.Model(inputs, outputs)
def _multiple_ops_at_end(): inputs = keras.Input(shape=(10,)) x = keras.layers.Dense(10)(inputs) x = gen_nn_ops.relu(x) outputs = gen_nn_ops.relu(x) return keras.Model(inputs, outputs)
def predict(self, inputTensor, _):
result = gen_nn_ops.relu(inputTensor)
return (result)
def _single_op_at_end(): inputs = keras.Input(shape=(10,)) x = keras.layers.Dense(10)(inputs) outputs = gen_nn_ops.relu(x) return inputs, outputs
def _single_op_at_end(): inputs = keras.Input(shape=(10,)) x = keras.layers.Dense(10)(inputs) outputs = gen_nn_ops.relu(x, name='hey') return inputs, outputs
def _multiple_ops_at_end(): inputs = keras.Input(shape=(10,)) x = keras.layers.Dense(10)(inputs) x = gen_nn_ops.relu(x) outputs = gen_nn_ops.relu(x) return inputs, outputs
def _single_op_in_middle(): inputs = keras.Input(shape=(10,)) x = keras.layers.Dense(10)(inputs) x = gen_nn_ops.relu(x) outputs = keras.layers.Dense(10)(x) return inputs, outputs
