pytorch/docs/source/optim.rst
Ilqar Ramazanli cec08e7032 To add warm-up scheduler to optim (#60836)
Summary:
Warm up of learning rate scheduling has initially been discussed  by Priya et. al. in the paper: https://arxiv.org/pdf/1706.02677.pdf .

In the section 2.2 of the paper they discussed and proposed idea of warming up learning schedulers in order to prevent big variance / noise in the learning rate. Then idea has been further discussed in the following papers:
  * Akilesh Gotmare et al. https://arxiv.org/abs/1810.13243
  * Bernstein et al  http://proceedings.mlr.press/v80/bernstein18a/bernstein18a.pdf
  * Liyuan Liu et al: https://arxiv.org/pdf/1908.03265.pdf

There are two type of popularly used learning rate warm up ideas
  * Constant warmup  (start with very small constant learning rate)
  * Linear Warmup        ( start with small learning rate and gradually increase)

In this PR we are adding warm up as learning rate scheduler. Note that learning rates are chainable, which means that we can merge warmup scheduler with any other learning rate scheduler to make more sophisticated learning rate scheduler.

## Linear Warmup

Linear Warmup is multiplying learning rate with pre-defined constant - warmup_factor in the first epoch (epoch 0). Then targeting to increase this multiplication constant to one in warmup_iters many epochs. Hence we can derive the formula at i-th step to have multiplication constant equal to:

                    warmup_factor + (1-warmup_factor) * i /  warmup_iters

Moreover, the fraction of this quantity at point i to point i-1 will give us

           1 + (1.0 - warmup_factor) / [warmup_iters*warmup_factor+(i-1)*(1-warmup_factor)]

which is used in get_lr() method in our implementation. Below we provide an example how to use linear warmup scheduler and to give an example to show how does it works.

```python
import torch
from torch.nn import Parameter
from torch.optim import SGD
from torch.optim.lr_scheduler import WarmUpLR

model = [Parameter(torch.randn(2, 2, requires_grad=True))]
optimizer = SGD(model, 0.1)
scheduler = WarmUpLR(optimizer, warmup_factor=0.1, warmup_iters=10, warmup_method="linear")

for epoch in range(15):

    print(epoch, scheduler.get_last_lr()[0])

    optimizer.step()
    scheduler.step()
```

```
0 0.010000000000000002
1 0.019000000000000003
2 0.028000000000000008
3 0.03700000000000001
4 0.04600000000000001
5 0.055000000000000014
6 0.06400000000000002
7 0.07300000000000002
8 0.08200000000000003
9 0.09100000000000004
10 0.10000000000000005
11 0.10000000000000005
12 0.10000000000000005
13 0.10000000000000005
14 0.10000000000000005
```

## Constant Warmup

Constant warmup has straightforward idea, to multiply learning rate by warmup_factor until we reach to epoch warmup_factor, then do nothing for following epochs

```python
import torch
from torch.nn import Parameter
from torch.optim import SGD
from torch.optim.lr_scheduler import WarmUpLR

model = [Parameter(torch.randn(2, 2, requires_grad=True))]
optimizer = SGD(model, 0.1)
scheduler = WarmUpLR(optimizer, warmup_factor=0.1, warmup_iters=5, warmup_method="constant")

for epoch in range(10):

    print(epoch, scheduler.get_last_lr()[0])

    optimizer.step()
    scheduler.step()
```

```
0 0.010000000000000002
1 0.010000000000000002
2 0.010000000000000002
3 0.010000000000000002
4 0.010000000000000002
5 0.10000000000000002
6 0.10000000000000002
7 0.10000000000000002
8 0.10000000000000002
9 0.10000000000000002
```

Pull Request resolved: https://github.com/pytorch/pytorch/pull/60836

Reviewed By: saketh-are

Differential Revision: D29537615

Pulled By: iramazanli

fbshipit-source-id: d910946027acc52663b301f9c56ade686e62cb69
2021-08-15 12:31:45 -07:00

324 lines
11 KiB
ReStructuredText

torch.optim
===================================
.. automodule:: torch.optim
How to use an optimizer
-----------------------
To use :mod:`torch.optim` you have to construct an optimizer object, that will hold
the current state and will update the parameters based on the computed gradients.
Constructing it
^^^^^^^^^^^^^^^
To construct an :class:`Optimizer` you have to give it an iterable containing the
parameters (all should be :class:`~torch.autograd.Variable` s) to optimize. Then,
you can specify optimizer-specific options such as the learning rate, weight decay, etc.
.. note::
If you need to move a model to GPU via ``.cuda()``, please do so before
constructing optimizers for it. Parameters of a model after ``.cuda()`` will
be different objects with those before the call.
In general, you should make sure that optimized parameters live in
consistent locations when optimizers are constructed and used.
Example::
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.9)
optimizer = optim.Adam([var1, var2], lr=0.0001)
Per-parameter options
^^^^^^^^^^^^^^^^^^^^^
:class:`Optimizer` s also support specifying per-parameter options. To do this, instead
of passing an iterable of :class:`~torch.autograd.Variable` s, pass in an iterable of
:class:`dict` s. Each of them will define a separate parameter group, and should contain
a ``params`` key, containing a list of parameters belonging to it. Other keys
should match the keyword arguments accepted by the optimizers, and will be used
as optimization options for this group.
.. note::
You can still pass options as keyword arguments. They will be used as
defaults, in the groups that didn't override them. This is useful when you
only want to vary a single option, while keeping all others consistent
between parameter groups.
For example, this is very useful when one wants to specify per-layer learning rates::
optim.SGD([
{'params': model.base.parameters()},
{'params': model.classifier.parameters(), 'lr': 1e-3}
], lr=1e-2, momentum=0.9)
This means that ``model.base``'s parameters will use the default learning rate of ``1e-2``,
``model.classifier``'s parameters will use a learning rate of ``1e-3``, and a momentum of
``0.9`` will be used for all parameters.
Taking an optimization step
^^^^^^^^^^^^^^^^^^^^^^^^^^^
All optimizers implement a :func:`~Optimizer.step` method, that updates the
parameters. It can be used in two ways:
``optimizer.step()``
~~~~~~~~~~~~~~~~~~~~
This is a simplified version supported by most optimizers. The function can be
called once the gradients are computed using e.g.
:func:`~torch.autograd.Variable.backward`.
Example::
for input, target in dataset:
optimizer.zero_grad()
output = model(input)
loss = loss_fn(output, target)
loss.backward()
optimizer.step()
``optimizer.step(closure)``
~~~~~~~~~~~~~~~~~~~~~~~~~~~
Some optimization algorithms such as Conjugate Gradient and LBFGS need to
reevaluate the function multiple times, so you have to pass in a closure that
allows them to recompute your model. The closure should clear the gradients,
compute the loss, and return it.
Example::
for input, target in dataset:
def closure():
optimizer.zero_grad()
output = model(input)
loss = loss_fn(output, target)
loss.backward()
return loss
optimizer.step(closure)
.. _optimizer-algorithms:
Base class
----------
.. autoclass:: Optimizer
.. autosummary::
:toctree: generated
:nosignatures:
Optimizer.add_param_group
Optimizer.load_state_dict
Optimizer.state_dict
Optimizer.step
Optimizer.zero_grad
Algorithms
----------
.. autosummary::
:toctree: generated
:nosignatures:
Adadelta
Adagrad
Adam
AdamW
SparseAdam
Adamax
ASGD
LBFGS
NAdam
RAdam
RMSprop
Rprop
SGD
How to adjust learning rate
---------------------------
:mod:`torch.optim.lr_scheduler` provides several methods to adjust the learning
rate based on the number of epochs. :class:`torch.optim.lr_scheduler.ReduceLROnPlateau`
allows dynamic learning rate reducing based on some validation measurements.
Learning rate scheduling should be applied after optimizer's update; e.g., you
should write your code this way:
Example::
model = [Parameter(torch.randn(2, 2, requires_grad=True))]
optimizer = SGD(model, 0.1)
scheduler = ExponentialLR(optimizer, gamma=0.9)
for epoch in range(20):
for input, target in dataset:
optimizer.zero_grad()
output = model(input)
loss = loss_fn(output, target)
loss.backward()
optimizer.step()
scheduler.step()
Most learning rate schedulers can be called back-to-back (also referred to as
chaining schedulers). The result is that each scheduler is applied one after the
other on the learning rate obtained by the one preceding it.
Example::
model = [Parameter(torch.randn(2, 2, requires_grad=True))]
optimizer = SGD(model, 0.1)
scheduler1 = ExponentialLR(optimizer, gamma=0.9)
scheduler2 = MultiStepLR(optimizer, milestones=[30,80], gamma=0.1)
for epoch in range(20):
for input, target in dataset:
optimizer.zero_grad()
output = model(input)
loss = loss_fn(output, target)
loss.backward()
optimizer.step()
scheduler1.step()
scheduler2.step()
In many places in the documentation, we will use the following template to refer to schedulers
algorithms.
>>> scheduler = ...
>>> for epoch in range(100):
>>> train(...)
>>> validate(...)
>>> scheduler.step()
.. warning::
Prior to PyTorch 1.1.0, the learning rate scheduler was expected to be called before
the optimizer's update; 1.1.0 changed this behavior in a BC-breaking way. If you use
the learning rate scheduler (calling ``scheduler.step()``) before the optimizer's update
(calling ``optimizer.step()``), this will skip the first value of the learning rate schedule.
If you are unable to reproduce results after upgrading to PyTorch 1.1.0, please check
if you are calling ``scheduler.step()`` at the wrong time.
.. autosummary::
:toctree: generated
:nosignatures:
lr_scheduler.LambdaLR
lr_scheduler.MultiplicativeLR
lr_scheduler.StepLR
lr_scheduler.MultiStepLR
lr_scheduler.WarmUpLR
lr_scheduler.ExponentialLR
lr_scheduler.CosineAnnealingLR
lr_scheduler.ReduceLROnPlateau
lr_scheduler.CyclicLR
lr_scheduler.OneCycleLR
lr_scheduler.CosineAnnealingWarmRestarts
Stochastic Weight Averaging
---------------------------
:mod:`torch.optim.swa_utils` implements Stochastic Weight Averaging (SWA). In particular,
:class:`torch.optim.swa_utils.AveragedModel` class implements SWA models,
:class:`torch.optim.swa_utils.SWALR` implements the SWA learning rate scheduler and
:func:`torch.optim.swa_utils.update_bn` is a utility function used to update SWA batch
normalization statistics at the end of training.
SWA has been proposed in `Averaging Weights Leads to Wider Optima and Better Generalization`_.
.. _`Averaging Weights Leads to Wider Optima and Better Generalization`: https://arxiv.org/abs/1803.05407
Constructing averaged models
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
`AveragedModel` class serves to compute the weights of the SWA model. You can create an
averaged model by running:
>>> swa_model = AveragedModel(model)
Here the model ``model`` can be an arbitrary :class:`torch.nn.Module` object. ``swa_model``
will keep track of the running averages of the parameters of the ``model``. To update these
averages, you can use the :func:`update_parameters` function:
>>> swa_model.update_parameters(model)
SWA learning rate schedules
^^^^^^^^^^^^^^^^^^^^^^^^^^^
Typically, in SWA the learning rate is set to a high constant value. :class:`SWALR` is a
learning rate scheduler that anneals the learning rate to a fixed value, and then keeps it
constant. For example, the following code creates a scheduler that linearly anneals the
learning rate from its initial value to 0.05 in 5 epochs within each parameter group:
>>> swa_scheduler = torch.optim.swa_utils.SWALR(optimizer, \
>>> anneal_strategy="linear", anneal_epochs=5, swa_lr=0.05)
You can also use cosine annealing to a fixed value instead of linear annealing by setting
``anneal_strategy="cos"``.
Taking care of batch normalization
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
:func:`update_bn` is a utility function that allows to compute the batchnorm statistics for the SWA model
on a given dataloader ``loader`` at the end of training:
>>> torch.optim.swa_utils.update_bn(loader, swa_model)
:func:`update_bn` applies the ``swa_model`` to every element in the dataloader and computes the activation
statistics for each batch normalization layer in the model.
.. warning::
:func:`update_bn` assumes that each batch in the dataloader ``loader`` is either a tensors or a list of
tensors where the first element is the tensor that the network ``swa_model`` should be applied to.
If your dataloader has a different structure, you can update the batch normalization statistics of the
``swa_model`` by doing a forward pass with the ``swa_model`` on each element of the dataset.
Custom averaging strategies
^^^^^^^^^^^^^^^^^^^^^^^^^^^
By default, :class:`torch.optim.swa_utils.AveragedModel` computes a running equal average of
the parameters that you provide, but you can also use custom averaging functions with the
``avg_fn`` parameter. In the following example ``ema_model`` computes an exponential moving average.
Example:
>>> ema_avg = lambda averaged_model_parameter, model_parameter, num_averaged:\
>>> 0.1 * averaged_model_parameter + 0.9 * model_parameter
>>> ema_model = torch.optim.swa_utils.AveragedModel(model, avg_fn=ema_avg)
Putting it all together
^^^^^^^^^^^^^^^^^^^^^^^
In the example below, ``swa_model`` is the SWA model that accumulates the averages of the weights.
We train the model for a total of 300 epochs and we switch to the SWA learning rate schedule
and start to collect SWA averages of the parameters at epoch 160:
>>> loader, optimizer, model, loss_fn = ...
>>> swa_model = torch.optim.swa_utils.AveragedModel(model)
>>> scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=300)
>>> swa_start = 160
>>> swa_scheduler = SWALR(optimizer, swa_lr=0.05)
>>>
>>> for epoch in range(300):
>>> for input, target in loader:
>>> optimizer.zero_grad()
>>> loss_fn(model(input), target).backward()
>>> optimizer.step()
>>> if epoch > swa_start:
>>> swa_model.update_parameters(model)
>>> swa_scheduler.step()
>>> else:
>>> scheduler.step()
>>>
>>> # Update bn statistics for the swa_model at the end
>>> torch.optim.swa_utils.update_bn(loader, swa_model)
>>> # Use swa_model to make predictions on test data
>>> preds = swa_model(test_input)