本文整理汇总了Python中tensorflow.compat.v2.GradientTape方法的典型用法代码示例。如果您正苦于以下问题:Python v2.GradientTape方法的具体用法?Python v2.GradientTape怎么用?Python v2.GradientTape使用的例子?那么, 这里精选的方法代码示例或许可以为您提供帮助。您也可以进一步了解该方法所在类tensorflow.compat.v2
的用法示例。
在下文中一共展示了v2.GradientTape方法的15个代码示例,这些例子默认根据受欢迎程度排序。您可以为喜欢或者感觉有用的代码点赞,您的评价将有助于系统推荐出更棒的Python代码示例。
示例1: testGrad
# 需要导入模块: from tensorflow.compat import v2 [as 别名]
# 或者: from tensorflow.compat.v2 import GradientTape [as 别名]
def testGrad(self):
def f(a, b):
return tf_np.sum(tf_np.sqrt(tf_np.exp(a)) + b)
g = extensions.grad(f)
def compare(a, b):
with tf.GradientTape() as tape:
tape.watch(a.data)
r = f(a, b)
expected = tape.gradient(r.data, a.data)
self.assertAllEqual(expected, g(a, b))
shape = [10]
a = tf_np.random.randn(*shape)
b = tf_np.random.randn(*shape)
compare(a, b)
示例2: hessian
# 需要导入模块: from tensorflow.compat import v2 [as 别名]
# 或者: from tensorflow.compat.v2 import GradientTape [as 别名]
def hessian(function: Callable[[Parameters], tf.Tensor],
parameters: Parameters) -> Parameters:
"""Computes the Hessian of a given function.
Useful for testing, although scales very poorly.
Args:
function: A function for which we want to compute the Hessian.
parameters: Parameters with respect to the Hessian should be computed.
Returns:
A tensor or list of tensors of same nested structure as `Parameters`,
representing the Hessian.
"""
with tf.GradientTape() as outer_tape:
with tf.GradientTape() as inner_tape:
value = function(parameters)
grads = inner_tape.gradient(value, parameters)
grads = tensor_list_util.tensor_list_to_vector(grads)
return outer_tape.jacobian(grads, parameters)
示例3: test_multiple_state_vars
# 需要导入模块: from tensorflow.compat import v2 [as 别名]
# 或者: from tensorflow.compat.v2 import GradientTape [as 别名]
def test_multiple_state_vars(self):
x = tf.constant([3.0, 4.0])
y = tf.constant([5.0, 6.0])
z = tf.constant([7.0, 8.0])
alpha = tf.constant(2.0)
beta = tf.constant(1.0)
with tf.GradientTape(persistent=True) as tape:
tape.watch([alpha, beta])
def body(i, state):
x, y, z = state
k = tf.cast(i + 1, tf.float32)
return [x * alpha - beta, y * k * alpha * beta, z * beta + x]
out = for_loop(body, [x, y, z], [alpha, beta], 3)
with self.subTest("independent_vars"):
grad = tape.gradient(out[1], alpha)
self.assertAllEqual(792, grad)
with self.subTest("dependent_vars"):
grad = tape.gradient(out[2], beta)
self.assertAllEqual(63, grad)
示例4: test_batching
# 需要导入模块: from tensorflow.compat import v2 [as 别名]
# 或者: from tensorflow.compat.v2 import GradientTape [as 别名]
def test_batching(self):
x = tf.constant([[3.0, 4.0], [30.0, 40.0]])
y = tf.constant([[5.0, 6.0], [50.0, 60.0]])
z = tf.constant([[7.0, 8.0], [70.0, 80.0]])
alpha = tf.constant(2.0)
beta = tf.constant(1.0)
with tf.GradientTape(persistent=True) as tape:
tape.watch([alpha, beta])
def body(i, state):
x, y, z = state
k = tf.cast(i + 1, tf.float32)
return [x * alpha - beta, y * k * alpha * beta, z * beta + x]
out = for_loop(body, [x, y, z], [alpha, beta], 3)
with self.subTest("independent_vars"):
grad = tape.gradient(out[1], alpha)
self.assertAllEqual(8712, grad)
with self.subTest("dependent_vars"):
grad = tape.gradient(out[2], beta)
self.assertAllEqual(783, grad)
示例5: test_with_xla
# 需要导入模块: from tensorflow.compat import v2 [as 别名]
# 或者: from tensorflow.compat.v2 import GradientTape [as 别名]
def test_with_xla(self):
@tf.function
def fn():
x = tf.constant([[3.0, 4.0], [30.0, 40.0]])
y = tf.constant([[7.0, 8.0], [70.0, 80.0]])
alpha = tf.constant(2.0)
beta = tf.constant(1.0)
with tf.GradientTape(persistent=True) as tape:
tape.watch([alpha, beta])
def body(i, state):
del i
x, y = state
return [x * alpha - beta, y * beta + x]
out = for_loop(body, [x, y], [alpha, beta], 3)
return tape.gradient(out[1], beta)
grad = self.evaluate(tf.xla.experimental.compile(fn))[0]
self.assertAllEqual(783, grad)
示例6: train_step
# 需要导入模块: from tensorflow.compat import v2 [as 别名]
# 或者: from tensorflow.compat.v2 import GradientTape [as 别名]
def train_step(model, loss_fn, optimizer_fn, metric, image, label):
"""Perform one training step for the model.
Args:
model: Keras model to train.
loss_fn: Loss function to use.
optimizer_fn: Optimizer function to use.
metric: keras.metric to use.
image: Tensor of training images of shape [batch_size, 28, 28, 1].
label: Tensor of class labels of shape [batch_size].
"""
with tf.GradientTape() as tape:
preds = model(image)
label_onehot = tf.one_hot(label, 10)
loss_ = loss_fn(label_onehot, preds)
grads = tape.gradient(loss_, model.trainable_variables)
optimizer_fn.apply_gradients(zip(grads, model.trainable_variables))
metric(loss_)
示例7: run_one_epoch
# 需要导入模块: from tensorflow.compat import v2 [as 别名]
# 或者: from tensorflow.compat.v2 import GradientTape [as 别名]
def run_one_epoch(
self, minibatches: Iterable[np.ndarray], training: bool = False,
):
total_loss, num_samples, num_tokens, num_correct_tokens = 0.0, 0, 0, 0
for step, minibatch_data in enumerate(minibatches):
with tf.GradientTape() as tape:
model_outputs = self.compute_logits(minibatch_data, training=training)
result = self.compute_loss_and_acc(model_outputs, minibatch_data)
total_loss += result.token_ce_loss
num_samples += minibatch_data.shape[0]
num_tokens += result.num_predictions
num_correct_tokens += result.num_correct_token_predictions
if training:
gradients = tape.gradient(
result.token_ce_loss, self.trainable_variables
)
self.optimizer.apply_gradients(zip(gradients, self.trainable_variables))
print(
" Batch %4i: Epoch avg. loss: %.5f || Batch loss: %.5f | acc: %.5f"
% (
step,
total_loss / num_samples,
result.token_ce_loss,
result.num_correct_token_predictions
/ (float(result.num_predictions) + 1e-7),
),
end="\r",
)
print("\r\x1b[K", end="")
return (
total_loss / num_samples,
num_correct_tokens / (float(num_tokens) + 1e-7),
)
示例8: train
# 需要导入模块: from tensorflow.compat import v2 [as 别名]
# 或者: from tensorflow.compat.v2 import GradientTape [as 别名]
def train(self, x, y, learning_rate=0.01):
"""Runs a single training pass.
Args:
x: 2-d array of size batch_size x image_size.
y: 2-d array of size batch_size x num_classes in one-hot notation.
learning_rate: The learning rate.
"""
x = np.array(x, copy=False)
y = np.array(y, copy=False)
def mean_squared_error(x, y):
diff = x - y
return np.sum(diff * diff) / len(x)
wb_tensors = [p.data for p in self.weights + self.biases]
with tf.GradientTape() as g:
g.watch(wb_tensors)
loss = mean_squared_error(self.forward(x), y)
gradients = g.gradient(loss.data, wb_tensors)
gradients = [np.asarray(grad) for grad in gradients]
new_weights_and_biases = []
for v, dv in zip(self.weights + self.biases, gradients):
new_weights_and_biases.append(v - learning_rate * dv)
total_len = len(new_weights_and_biases)
self.weights = new_weights_and_biases[:total_len // 2]
self.biases = new_weights_and_biases[total_len // 2:]
示例9: testVjp
# 需要导入模块: from tensorflow.compat import v2 [as 别名]
# 或者: from tensorflow.compat.v2 import GradientTape [as 别名]
def testVjp(self, has_aux):
x_shape = (tf.TensorShape([10]), tf.TensorShape([1, 10]))
y_shape = (tf.TensorShape([]))
dtype = np.float32
def f(a, b):
y = tf_np.sum(tf_np.sqrt(tf_np.exp(a)) + b)
if has_aux:
return y, tf_np.asarray(1)
else:
return y
rng = tf.random.Generator.from_seed(1234)
x, dy_list = tf.nest.map_structure(lambda shape: uniform(rng, shape, dtype),
[x_shape, [y_shape] * 2])
tf_x = to_tf(x)
outputs = extensions.vjp(f, *x, has_aux=has_aux)
if has_aux:
y, vjp, aux = outputs
else:
y, vjp = outputs
with tf.GradientTape(persistent=True) as tape:
tape.watch(tf_x)
outputs = f(*x)
if has_aux:
expected_y, expected_aux = outputs
self.assertAllClose(to_tf(expected_aux), to_tf(aux))
else:
expected_y = outputs
self.assertAllClose(to_tf(expected_y), to_tf(y))
for dy in dy_list:
expected_dx = tape.gradient(
to_tf(expected_y), tf_x, output_gradients=to_tf(dy))
self.assertAllClose(expected_dx, to_tf(vjp(dy)))
示例10: test_setitem
# 需要导入模块: from tensorflow.compat import v2 [as 别名]
# 或者: from tensorflow.compat.v2 import GradientTape [as 别名]
def test_setitem(self):
# Single integer index.
a = array_ops.array([1., 2., 3.])
b = array_ops.array(5.)
c = array_ops.array(10.)
tensors = [arr.data for arr in [a, b, c]]
with tf.GradientTape() as g:
g.watch(tensors)
a[1] = b + c
loss = array_ops.sum(a)
gradients = g.gradient(loss.data, tensors)
self.assertSequenceEqual(
array_ops.array(gradients[0]).tolist(), [1., 0., 1.])
self.assertEqual(array_ops.array(gradients[1]).tolist(), 1.)
self.assertEqual(array_ops.array(gradients[2]).tolist(), 1.)
# Tuple index.
a = array_ops.array([[[1., 2.], [3., 4.]], [[5., 6.],
[7., 8.]]]) # 2x2x2 array.
b = array_ops.array([10., 11.])
tensors = [arr.data for arr in [a, b]]
with tf.GradientTape() as g:
g.watch(tensors)
a[(1, 0)] = b
loss = array_ops.sum(a)
gradients = g.gradient(loss.data, tensors)
self.assertSequenceEqual(
array_ops.array(gradients[0]).tolist(),
[[[1., 1.], [1., 1.]], [[0., 0.], [1., 1.]]])
self.assertEqual(array_ops.array(gradients[1]).tolist(), [1., 1.])
示例11: _hessian_vector_product
# 需要导入模块: from tensorflow.compat import v2 [as 别名]
# 或者: from tensorflow.compat.v2 import GradientTape [as 别名]
def _hessian_vector_product(
function: Callable[[Parameters], tf.Tensor],
parameters: Parameters,
v: Parameters) -> Parameters:
"""Computes Hessian-vector products.
Computes the product H.v where v is an arbitrary vector and H is the Hessian
of a function evaluated at `parameters`.
The result is the same as if the Hessian was computed explicitly and
multiplied the vector. However, this function uses the autograd in backward
then forward mode in order to compute this Hessian vector product without
having to explicitly compute the Hessian.
Args:
function: A (twice) differentiable function that takes as input a tensor or
a list of tensors and returns a scalar.
parameters: The parameters with respect to which we want to compute the
Hessian for the hessian vector product.
v: An arbitrary vector or list of vectors of the same nested structure as
`parameters`.
Returns:
A vector or list of vectors of the same nested structure as
`parameters`, equal to H.v.
"""
with tf.autodiff.ForwardAccumulator(
primals=parameters, tangents=v) as acc:
with tf.GradientTape() as tape:
tape.watch(parameters)
value = function(parameters)
backward = tape.gradient(value, parameters)
return acc.jvp(backward)
示例12: test_simple_grad_wrt_parameter
# 需要导入模块: from tensorflow.compat import v2 [as 别名]
# 或者: from tensorflow.compat.v2 import GradientTape [as 别名]
def test_simple_grad_wrt_parameter(self):
x = tf.constant([3.0])
sigma = tf.constant(2.0)
with tf.GradientTape() as tape:
tape.watch(sigma)
def body(i, state):
del i
x = state[0]
return [x * sigma]
out = for_loop(body, [x], [sigma], 3)[0]
grad = tape.gradient(out, sigma)
self.assertAllEqual(36, grad)
示例13: test_simple_grad_wrt_initial_state
# 需要导入模块: from tensorflow.compat import v2 [as 别名]
# 或者: from tensorflow.compat.v2 import GradientTape [as 别名]
def test_simple_grad_wrt_initial_state(self):
x = tf.constant([3.0])
sigma = tf.constant(2.0)
with tf.GradientTape() as tape:
tape.watch(x)
def body(i, state):
del i
x = state[0]
return [x * sigma]
out = for_loop(body, [x], [sigma], 3)[0]
grad = tape.gradient(out, x)
self.assertAllEqual([8], grad)
示例14: test_shapes
# 需要导入模块: from tensorflow.compat import v2 [as 别名]
# 或者: from tensorflow.compat.v2 import GradientTape [as 别名]
def test_shapes(self, state_dims, num_params, times):
# Checks that the loop can handle various shapes and outputs correct shapes.
def test_with_batch_shape(batch_shape):
initial_state = [tf.ones(shape=batch_shape + (d,)) for d in state_dims]
params = [tf.constant(1.0) for _ in range(num_params)]
with tf.GradientTape(persistent=True) as tape:
tape.watch(initial_state)
tape.watch(params)
def body(i, state):
del i
if not params:
return state
sum_params = tf.add_n(params)
state = [s * sum_params for s in state]
return state
final_state = for_loop(body, initial_state, params, times)
for s_in in initial_state:
for s_out in final_state:
grad = tape.gradient(s_out, s_in)
self.assertAllEqual(s_in.shape, grad.shape)
for p in params:
for s_out in final_state:
grad = tape.gradient(s_out, p)
self.assertAllEqual([], grad.shape)
with self.subTest("no_batch"):
test_with_batch_shape(batch_shape=())
with self.subTest("simple_batch"):
test_with_batch_shape(batch_shape=(5,))
with self.subTest("complex_batch"):
test_with_batch_shape(batch_shape=(2, 8, 3))
示例15: _jacobian_wrt_parameter
# 需要导入模块: from tensorflow.compat import v2 [as 别名]
# 或者: from tensorflow.compat.v2 import GradientTape [as 别名]
def _jacobian_wrt_parameter(y, param, tape):
"""Computes a Jacobian w.r.t. a parameter."""
# For input shapes (b, dy), yields shape (b, dy, 1) (1 is added for
# convenience elsewhere).
# To avoid having to broadcast param to y's shape, we need to take a forward
# gradient.
with tf.GradientTape() as w_tape:
w = tf.zeros_like(y)
w_tape.watch(w)
vjp = tape.gradient(y, param, output_gradients=w)
if vjp is None: # Unconnected.
return tf.expand_dims(tf.zeros_like(y), axis=-1)
return tf.expand_dims(w_tape.gradient(vjp, w), axis=-1)