LBFGS¶

class paddle.incubate.optimizer. LBFGS ( learning_rate=1.0, max_iter=20, max_eval=None, tolerance_grad=1e-07, tolerance_change=1e-09, history_size=100, line_search_fn=None, parameters=None, weight_decay=None, grad_clip=None, name=None ) [source]

The L-BFGS is a quasi-Newton method for solving an unconstrained optimization problem over a differentiable function. Closely related is the Newton method for minimization. Consider the iterate update formula:

\[x_{k+1} = x_{k} + H_k \nabla{f_k}\]

If \(H_k\) is the inverse Hessian of \(f\) at \(x_k\), then it’s the Newton method. If \(H_k\) is symmetric and positive definite, used as an approximation of the inverse Hessian, then it’s a quasi-Newton. In practice, the approximated Hessians are obtained by only using the gradients, over either whole or part of the search history, the former is BFGS, the latter is L-BFGS.

Reference:

Jorge Nocedal, Stephen J. Wright, Numerical Optimization, Second Edition, 2006. pp179: Algorithm 7.5 (L-BFGS).

Parameters
• learning_rate (float, optional) – learning rate .The default value is 1.

• max_iter (int, optional) – maximal number of iterations per optimization step. The default value is 20.

• max_eval (int, optional) – maximal number of function evaluations per optimization step. The default value is max_iter * 1.25.

• tolerance_grad (float, optional) – termination tolerance on first order optimality The default value is 1e-5.

• tolerance_change (float, optional) – termination tolerance on function value/parameter changes. The default value is 1e-9.

• history_size (int, optional) – update history size. The default value is 100.

• line_search_fn (string, optional) – either ‘strong_wolfe’ or None. The default value is strong_wolfe.

• parameters (list|tuple, optional) – List/Tuple of `Tensor` names to update to minimize `loss`. This parameter is required in dygraph mode. The default value is None.

• weight_decay (float|WeightDecayRegularizer, optional) – The strategy of regularization. It canbe a float value as coeff of L2 regularization or L1Decay, L2Decay. If a parameter has set regularizer using ParamAttr already, the regularization setting here in optimizer will be ignored for this parameter. Otherwise, the regularization setting here in optimizer will take effect. Default None, meaning there is no regularization.

• grad_clip (GradientClipBase, optional) – Gradient cliping strategy, it’s an instance of some derived class of `GradientClipBase` . There are three cliping strategies ( ClipGradByGlobalNorm , ClipGradByNorm , ClipGradByValue ). Default None, meaning there is no gradient clipping.

• name (str, optional) – Normally there is no need for user to set this property. For more information, please refer to Name. The default value is None.

Returns

the final loss of closure.

Return type

loss (Tensor)

Examples

```>>> import paddle
>>> import numpy as np

>>> np.random.seed(0)
>>> np_w = np.random.rand(1).astype(np.float32)
>>> np_x = np.random.rand(1).astype(np.float32)

>>> inputs = [np.random.rand(1).astype(np.float32) for i in range(10)]
>>> # y = 2x
>>> targets = [2 * x for x in inputs]

...     def __init__(self):
...         super().__init__()
...     def forward(self, x):
...         return self.w * x

>>> net = Net()
>>> opt = LBFGS(learning_rate=1, max_iter=1, max_eval=None, tolerance_grad=1e-07, tolerance_change=1e-09, history_size=100, line_search_fn='strong_wolfe', parameters=net.parameters())
>>> def train_step(inputs, targets):
...     def closure():
...         outputs = net(inputs)
...         print('loss: ', loss.item())
...         loss.backward()
...         return loss
...     opt.step(closure)

>>> for input, target in zip(inputs, targets):
...     train_step(input, target)
```
state_dict ( )

Returns the state of the optimizer as a `dict`.

Returns

state, a dict holding current optimization state. Its content

differs between optimizer classes.

Create and add backward regularization Operators

Creates and adds backward regularization operators in the BlockDesc. This will add gradients of the regularizer function to the gradients of the parameters and return these modified gradients. This is the same as implementing weight decay in optimizers for regularization.

Parameters
• parameters_and_grads – A list of (parameters, gradients) pairs that need to be regularized.

• regularization – A global regularizer. If the parameter is not set. It will be applied with regularizer.

Returns

Return type

list[(Variable, Variable)]

Raises

Exception – Unknown regularization type

Second part of minimize, appending optimization operators for given params_grads pairs.

Parameters

Returns

A list of operators appended to the current program.

Return type

list

Examples

```>>> import paddle

>>> inp = paddle.uniform([10, 10], dtype="float32", min=-0.1, max=0.1)
>>> out = linear(inp)
...         parameters=linear.parameters())
```
backward ( loss, startup_program=None, parameters=None, no_grad_set=None, callbacks=None )

The first part of `minimize`, do auto-diff to append backward operations for the current program.

Parameters
• loss (Tensor) – `loss` tensor to run optimizations.

• startup_program (Program, optional) – Program for initializing parameters in `parameters`. The default value is None, at this time default_startup_program will be used.

• parameters (list, optional) – List of `Tensor` or `Tensor.name` to update to minimize `loss`. The default value is None, at this time all parameters will be updated.

• no_grad_set (set, optional) – Set of `Tensor` or `Tensor.name` that don’t need to be updated. The default value is None.

• callbacks (list, optional) – list of callable objects to run when appending backward operator for one parameter. The default value is None.

Returns

list of (param, grad) tensor pairs, param is `Parameter`,

Return type

list

Examples

```>>> import paddle
>>> x = paddle.arange(26, dtype="float32").reshape([2, 13])

>>> # This can be any optimizer supported by dygraph.
...                             parameters = linear.parameters())
>>> out = linear(x)
>>> out.backward()
```

Clear the gradients of all optimized parameters for model.

Parameters

set_to_zero (bool, optional) – If set grads to zero or not, default is True.

Returns

None

Examples

```>>> import paddle

>>> a = paddle.arange(26, dtype="float32").reshape([2, 13])
>>> # This can be any optimizer supported by dygraph.
...                             parameters = linear.parameters())
>>> out = linear(a)
>>> out.backward()
```
get_lr ( )

Get current learning rate of optimizer. If ‘LRScheduler’ is not used, the return value is all the same. If ‘LRScheduler’ is used, the return value is the current scheduled learing rete.

Returns

The current learning rate of optimizer.

Return type

float

Examples

```>>> # train on default dynamic graph mode
>>> import numpy as np

>>> ## example1: LRScheduler is not used, return the same value is all the same
>>> for batch in range(10):
...     input = paddle.randint(low=0, high=5, shape=[5])
...     out = emb(input)
...     out.backward()
...     print("Learning rate of step{}: {}".format(batch, adam.get_lr())) # 0.01
Learning rate of step0: 0.01
Learning rate of step1: 0.01
Learning rate of step2: 0.01
Learning rate of step3: 0.01
Learning rate of step4: 0.01
Learning rate of step5: 0.01
Learning rate of step6: 0.01
Learning rate of step7: 0.01
Learning rate of step8: 0.01
Learning rate of step9: 0.01

>>> ## example2: StepDecay is used, return the scheduled learning rate
>>> scheduler = paddle.optimizer.lr.StepDecay(learning_rate=0.5, step_size=2, gamma=0.1)
>>> for batch in range(10):
...     input = paddle.randint(low=0, high=5, shape=[5])
...     out = emb(input)
...     out.backward()
...     print("Learning rate of step{}: {}".format(batch, adam.get_lr())) # 0.5->0.05...
...     scheduler.step()
Learning rate of step0: 0.5
Learning rate of step1: 0.5
Learning rate of step2: 0.05
Learning rate of step3: 0.05
Learning rate of step4: 0.005000000000000001
Learning rate of step5: 0.005000000000000001
Learning rate of step6: 0.0005000000000000001
Learning rate of step7: 0.0005000000000000001
Learning rate of step8: 5.000000000000001e-05
Learning rate of step9: 5.000000000000001e-05

>>> # train on static graph mode
...     x = paddle.static.data(name='x', shape=[None, 10])
...     scheduler = paddle.optimizer.lr.StepDecay(learning_rate=0.5, step_size=2, gamma=0.1)

>>> exe.run(start_prog)
>>> for batch in range(10):
...     print("Learning rate of step{}: {}".format(batch, adam.get_lr())) # 0.5->0.05->0.005...
...     out = exe.run(main_prog, feed={'x': np.random.randn(3, 10).astype('float32')})
...     scheduler.step()
Learning rate of step0: 0.5
Learning rate of step1: 0.5
Learning rate of step2: 0.05
Learning rate of step3: 0.05
Learning rate of step4: 0.005000000000000001
Learning rate of step5: 0.005000000000000001
Learning rate of step6: 0.0005000000000000001
Learning rate of step7: 0.0005000000000000001
Learning rate of step8: 5.000000000000001e-05
Learning rate of step9: 5.000000000000001e-05
```
minimize ( loss, startup_program=None, parameters=None, no_grad_set=None )

Add operations to minimize `loss` by updating `parameters`.

Parameters
• loss (Tensor) – A `Tensor` containing the value to minimize.

• startup_program (Program, optional) – Program for initializing parameters in `parameters`. The default value is None, at this time default_startup_program will be used.

• parameters (list, optional) – List of `Tensor` or `Tensor.name` to update to minimize `loss`. The default value is None, at this time all parameters will be updated.

• no_grad_set (set, optional) – Set of `Tensor` or `Tensor.name` that don’t need to be updated. The default value is None.

Returns

tuple (optimize_ops, params_grads), A list of operators appended by minimize and a list of (param, grad) tensor pairs, param is `Parameter`, grad is the gradient value corresponding to the parameter. In static graph mode, the returned tuple can be passed to `fetch_list` in `Executor.run()` to indicate program pruning. If so, the program will be pruned by `feed` and `fetch_list` before run, see details in `Executor`.

Return type

tuple

Examples

```>>> import paddle
>>> input = paddle.uniform(shape=[10, 10], min=-0.1, max=0.1)
>>> out = linear(input)

...         parameters=linear.parameters(),
...         weight_decay=0.01)
>>> loss.backward()
```
set_lr ( value )
Api_attr

imperative

Set the value of the learning rate manually in the optimizer. If the optimizer use LRScheduler, this API cannot be invoked, because it will lead to conflict.

Parameters

value (float) – the value of learning rate

Returns

None

Examples

```>>> import paddle

>>> # set learning rate manually by python float value
>>> lr_list = [0.2, 0.3, 0.4, 0.5, 0.6]
>>> for i in range(5):
...     print("current lr is {}".format(lr))
current lr is 0.2
current lr is 0.3
current lr is 0.4
current lr is 0.5
current lr is 0.6
```
set_lr_scheduler ( scheduler )
Api_attr

imperative

Set the LRScheduler of the learning rate manually in the optimizer. If the optimizer already used LRScheduler previously, this API will set it be the new one.

Parameters

scheduler (LRScheduler) – the LRScheduler of learning rate

Returns

None

Examples

```>>> import paddle

>>> # set learning rate manually by class LRScheduler
>>> scheduler = paddle.optimizer.lr.MultiStepDecay(learning_rate=0.5, milestones=[2,4,6], gamma=0.8)
>>> print("current lr is {}".format(lr))
current lr is 0.5

>>> # set learning rate manually by another LRScheduler
>>> scheduler = paddle.optimizer.lr.StepDecay(learning_rate=0.1, step_size=5, gamma=0.6)
>>> print("current lr is {}".format(lr))
current lr is 0.1
```
set_state_dict ( state_dict )

Load optimizer state dict. For Adam optimizer, contains beta1, beta2, momentum etc. If LRScheduler have been used, global_step will be changed.

Parameters

state_dict (dict) – Dict contains all the Tensor needed by optimizer

Returns

None

Examples

```>>> import paddle

>>> layer_state_dict = emb.state_dict()

...     d_model=0.01, warmup_steps=100, verbose=True)
...     learning_rate=scheduler,
...     parameters=emb.parameters())