Shortcuts

Lightning CLI and config files

Another source of boilerplate code that Lightning can help to reduce is in the implementation of command line tools. Furthermore, it provides a standardized way to configure trainings using a single file that includes settings for Trainer and user extended LightningModule and LightningDataModule classes. The full configuration is automatically saved in the log directory. This has the benefit of greatly simplifying the reproducibility of experiments.

The main requirement for user extended classes to be made configurable is that all relevant init arguments must have type hints. This is not a very demanding requirement since it is good practice to do anyway. As a bonus if the arguments are described in the docstrings, then the help of the command line tool will display them.

Warning

LightningCLI is in beta and subject to change.


LightningCLI

The implementation of training command line tools is done via the LightningCLI class. The minimal installation of pytorch-lightning does not include this support. To enable it, either install lightning with the all extras require or install the package jsonargparse[signatures].

The case in which the user’s LightningModule class implements all required *_dataloader methods, a trainer.py tool can be as simple as:

from pytorch_lightning.utilities.cli import LightningCLI

cli = LightningCLI(MyModel)

The help of the tool describing all configurable options and default values can be shown by running python trainer.py --help. Default options can be changed by providing individual command line arguments. However, it is better practice to create a configuration file and provide this to the tool. A way to do this would be:

# Dump default configuration to have as reference
python trainer.py --print_config > default_config.yaml
# Create config including only options to modify
nano config.yaml
# Run training using created configuration
python trainer.py --config config.yaml

The instantiation of the LightningCLI class takes care of parsing command line and config file options, instantiating the classes, setting up a callback to save the config in the log directory and finally running the trainer. The resulting object cli can be used for example to get the instance of the model, (cli.model).

After multiple trainings with different configurations, each run will have in its respective log directory a config.yaml file. This file can be used for reference to know in detail all the settings that were used for each particular run, and also could be used to trivially reproduce a training, e.g.:

python trainer.py --config lightning_logs/version_7/config.yaml

If a separate LightningDataModule class is required, the trainer tool just needs a small modification as follows:

from pytorch_lightning.utilities.cli import LightningCLI

cli = LightningCLI(MyModel, MyDataModule)

The start of a possible implementation of MyModel including the recommended argument descriptions in the docstring could be the one below. Note that by using type hints and docstrings there is no need to duplicate this information to define its configurable arguments.

class MyModel(LightningModule):

    def __init__(
        self,
        encoder_layers: int = 12,
        decoder_layers: List[int] = [2, 4]
    ):
        """Example encoder-decoder model

        Args:
            encoder_layers: Number of layers for the encoder
            decoder_layers: Number of layers for each decoder block
        """
        super().__init__()
        self.save_hyperparameters()

With this model class, the help of the trainer tool would look as follows:

$ python trainer.py --help
usage: trainer.py [-h] [--print_config] [--config CONFIG]
                  [--trainer.logger LOGGER]
                  ...

pytorch-lightning trainer command line tool

optional arguments:
  -h, --help            show this help message and exit
  --print_config        print configuration and exit
  --config CONFIG       Path to a configuration file in json or yaml format.
                        (default: null)

Customize every aspect of training via flags:
  ...
  --trainer.max_epochs MAX_EPOCHS
                        Stop training once this number of epochs is reached.
                        (type: int, default: 1000)
  --trainer.min_epochs MIN_EPOCHS
                        Force training for at least these many epochs (type: int,
                        default: 1)
  ...

Example encoder-decoder model:
  --model.encoder_layers ENCODER_LAYERS
                        Number of layers for the encoder (type: int, default: 12)
  --model.decoder_layers DECODER_LAYERS
                        Number of layers for each decoder block (type: List[int],
                        default: [2, 4])

The default configuration that option --print_config gives is in yaml format and for the example above would look as follows:

$ python trainer.py --print_config
model:
  decoder_layers:
  - 2
  - 4
  encoder_layers: 12
trainer:
  accelerator: null
  accumulate_grad_batches: 1
  amp_backend: native
  amp_level: O2
  ...

Note that there is a section for each class (model and trainer) including all the init parameters of the class. This grouping is also used in the formatting of the help shown previously.

Use of command line arguments

For every CLI implemented, users are encouraged to learn how to run it by reading the documentation printed with the --help option and use the --print_config option to guide the writing of config files. A few more details that might not be clear by only reading the help are the following.

LightningCLI is based on argparse and as such follows the same arguments style as many POSIX command line tools. Long options are prefixed with two dashes and its corresponding values should be provided with an empty space or an equal sign, as --option value or --option=value. Command line options are parsed from left to right, therefore if a setting appears multiple times the value most to the right will override the previous ones. If a class has an init parameter that is required (i.e. no default value), it is given as --option which makes it explicit and more readable instead of relying on positional arguments.

When calling a CLI, all options can be provided using individual arguments. However, given the large amount of options that the CLIs have, it is recommended to use a combination of config files and individual arguments. Therefore, a common pattern could be a single config file and only a few individual arguments that override defaults or values in the config, for example:

$ python trainer.py --config experiment_defaults.yaml --trainer.max_epochs 100

Another common pattern could be having multiple config files:

$ python trainer.py --config config1.yaml --config config2.yaml [...]

As explained before, config1.yaml is parsed first and then config2.yaml. Therefore, if individual settings are defined in both files, then the ones in config2.yaml will be used. Settings in config1.yaml that are not in config2.yaml are be kept.

Groups of options can also be given as independent config files:

$ python trainer.py --trainer trainer.yaml --model model.yaml --data data.yaml [...]

When running experiments in clusters it could be desired to use a config which needs to be accessed from a remote location. LightningCLI comes with fsspec support which allows reading from many types of remote file systems. One example is if you have installed the gcsfs then a config could be stored in an S3 bucket and accessed as:

$ python trainer.py --config s3://bucket/config.yaml [...]

In some cases people might what to pass an entire config in an environment variable, which could also be used instead of a path to a file, for example:

$ python trainer.py --trainer "$TRAINER_CONFIG" --model "$MODEL_CONFIG" [...]

An alternative for environment variables could be to instantiate the CLI with env_parse=True. In this case the help shows the names of the environment variables for all options. A global config would be given in PL_CONFIG and there wouldn’t be a need to specify any command line argument.

It is also possible to set a path to a config file of defaults. If the file exists it would be automatically loaded without having to specify any command line argument. Arguments given would override the values in the default config file. Loading a defaults file my_cli_defaults.yaml in the current working directory would be implemented as:

cli = LightningCLI(
    MyModel,
    MyDataModule,
    parser_kwargs={'default_config_files': ['my_cli_defaults.yaml']},
)

To load a file in the user’s home directory would be just changing to ~/.my_cli_defaults.yaml. Note that this setting is given through parser_kwargs. More parameters are supported. For details see the ArgumentParser API documentation.

Trainer Callbacks and arguments with class type

A very important argument of the Trainer class is the callbacks. In contrast to other more simple arguments which just require numbers or strings, callbacks expects a list of instances of subclasses of Callback. To specify this kind of argument in a config file, each callback must be given as a dictionary including a class_path entry with an import path of the class, and optionally an init_args entry with arguments required to instantiate it. Therefore, a simple configuration file example that defines a couple of callbacks is the following:

trainer:
  callbacks:
    - class_path: pytorch_lightning.callbacks.EarlyStopping
      init_args:
        patience: 5
    - class_path: pytorch_lightning.callbacks.LearningRateMonitor
      init_args:
        ...

Similar to the callbacks, any arguments in Trainer and user extended LightningModule and LightningDataModule classes that have as type hint a class can be configured the same way using class_path and init_args.

Multiple models and/or datasets

In the previous examples LightningCLI works only for a single model and datamodule class. However, there are many cases in which the objective is to easily be able to run many experiments for multiple models and datasets. For these cases the tool can be configured such that a model and/or a datamodule is specified by an import path and init arguments. For example, with a tool implemented as:

from pytorch_lightning.utilities.cli import LightningCLI

cli = LightningCLI(
    MyModelBaseClass,
    MyDataModuleBaseClass,
    subclass_mode_model=True,
    subclass_mode_data=True
)

A possible config file could be as follows:

model:
  class_path: mycode.mymodels.MyModel
  init_args:
    decoder_layers:
    - 2
    - 4
    encoder_layers: 12
data:
  class_path: mycode.mydatamodules.MyDataModule
  init_args:
    ...
trainer:
  callbacks:
    - class_path: pytorch_lightning.callbacks.EarlyStopping
      init_args:
        patience: 5
    ...

Only model classes that are a subclass of MyModelBaseClass would be allowed, and similarly only subclasses of MyDataModuleBaseClass. If as base classes LightningModule and LightningDataModule are given, then the tool would allow any lightning module and data module.

Tip

Note that with the subclass modes the --help option does not show information for a specific subclass. To get help for a subclass the options --model.help and --data.help can be used, followed by the desired class path. Similarly --print_config does not include the settings for a particular subclass. To include them the class path should be given before the --print_config option. Examples for both help and print config are:

$ python trainer.py --model.help mycode.mymodels.MyModel
$ python trainer.py --model mycode.mymodels.MyModel --print_config

Models with multiple submodules

Many use cases require to have several modules each with its own configurable options. One possible way to handle this with LightningCLI is to implement a single module having as init parameters each of the submodules. Since the init parameters have as type a class, then in the configuration these would be specified with class_path and init_args entries. For instance a model could be implemented as:

class MyMainModel(LightningModule):

    def __init__(
        self,
        encoder: EncoderBaseClass,
        decoder: DecoderBaseClass
    ):
        """Example encoder-decoder submodules model

        Args:
            encoder: Instance of a module for encoding
            decoder: Instance of a module for decoding
        """
        super().__init__()
        self.encoder = encoder
        self.decoder = decoder

If the CLI is implemented as LightningCLI(MyMainModel) the configuration would be as follows:

model:
  encoder:
    class_path: mycode.myencoders.MyEncoder
    init_args:
      ...
  decoder:
    class_path: mycode.mydecoders.MyDecoder
    init_args:
      ...

It is also possible to combine subclass_mode_model=True and submodules, thereby having two levels of class_path.

Customizing LightningCLI

The init parameters of the LightningCLI class can be used to customize some things, namely: the description of the tool, enabling parsing of environment variables and additional arguments to instantiate the trainer and configuration parser.

Nevertheless the init arguments are not enough for many use cases. For this reason the class is designed so that can be extended to customize different parts of the command line tool. The argument parser class used by LightningCLI is LightningArgumentParser which is an extension of python’s argparse, thus adding arguments can be done using the add_argument() method. In contrast to argparse it has additional methods to add arguments, for example add_class_arguments() adds all arguments from the init of a class, though requiring parameters to have type hints. For more details about this please refer to the respective documentation.

The LightningCLI class has the add_arguments_to_parser() method which can be implemented to include more arguments. After parsing, the configuration is stored in the config attribute of the class instance. The LightningCLI class also has two methods that can be used to run code before and after trainer.fit is executed: before_fit() and after_fit(). A realistic example for these would be to send an email before and after the execution of fit. The code would be something like:

from pytorch_lightning.utilities.cli import LightningCLI

class MyLightningCLI(LightningCLI):

    def add_arguments_to_parser(self, parser):
        parser.add_argument('--notification_email', default='will@email.com')

    def before_fit(self):
        send_email(
            address=self.config['notification_email'],
            message='trainer.fit starting'
        )

    def after_fit(self):
        send_email(
            address=self.config['notification_email'],
            message='trainer.fit finished'
        )

cli = MyLightningCLI(MyModel)

Note that the config object self.config is a dictionary whose keys are global options or groups of options. It has the same structure as the yaml format described previously. This means for instance that the parameters used for instantiating the trainer class can be found in self.config['trainer'].

Tip

Have a look at the LightningCLI class API reference to learn about other methods that can be extended to customize a CLI.

Configurable callbacks

As explained previously, any callback can be added by including it in the config via class_path and init_args entries. However, there are other cases in which a callback should always be present and be configurable. This can be implemented as follows:

from pytorch_lightning.callbacks import EarlyStopping
from pytorch_lightning.utilities.cli import LightningCLI

class MyLightningCLI(LightningCLI):

    def add_arguments_to_parser(self, parser):
        parser.add_lightning_class_args(EarlyStopping, 'my_early_stopping')
        parser.set_defaults({'my_early_stopping.patience': 5})

cli = MyLightningCLI(MyModel)

To change the configuration of the EarlyStopping in the config it would be:

model:
  ...
trainer:
  ...
my_early_stopping:
  patience: 5

Note

The example above overrides a default in add_arguments_to_parser. This is included to show that defaults can be changed if needed. However, note that overriding of defaults in the source code is not intended to be used to store the best hyperparameters for a task after experimentation. To ease reproducibility the source code should be stable. It is better practice to store the best hyperparameters for a task in a configuration file independent from the source code.

Class type defaults

The support for classes as type hints allows to try many possibilities with the same CLI. This is a useful feature, but it can make it tempting to use an instance of a class as a default. For example:

class MyMainModel(LightningModule):

    def __init__(
        self,
        backbone: torch.nn.Module = MyModel(encoder_layers=24),  # BAD PRACTICE!
    ):
        super().__init__()
        self.backbone = backbone

Normally classes are mutable as it is in this case. The instance of MyModel would be created the moment that the module that defines MyMainModel is first imported. This means that the default of backbone will be initialized before the CLI class runs seed_everything making it non-reproducible. Furthermore, if MyMainModel is used more than once in the same Python process and the backbone parameter is not overridden, the same instance would be used in multiple places which very likely is not what the developer intended. Having an instance as default also makes it impossible to generate the complete config file since for arbitrary classes it is not known which arguments were used to instantiate it.

A good solution to these problems is to not have a default or set the default to a special value (e.g. a string) which would be checked in the init and instantiated accordingly. If a class parameter has no default and the CLI is subclassed then a default can be set as follows:

default_backbone = {
    'class_path': 'import.path.of.MyModel',
    'init_args': {
        'encoder_layers': 24,
    },
}

class MyLightningCLI(LightningCLI):
    def add_arguments_to_parser(self, parser):
        parser.set_defaults({'model.backbone': default_backbone})

A more compact version that avoids writing a dictionary would be:

from jsonargparse import lazy_instance

class MyLightningCLI(LightningCLI):
    def add_arguments_to_parser(self, parser):
        parser.set_defaults({'model.backbone': lazy_instance(MyModel, encoder_layers=24)})

Argument linking

Another case in which it might be desired to extend LightningCLI is that the model and data module depend on a common parameter. For example in some cases both classes require to know the batch_size. It is a burden and error prone giving the same value twice in a config file. To avoid this the parser can be configured so that a value is only given once and then propagated accordingly. With a tool implemented like shown below, the batch_size only has to be provided in the data section of the config.

from pytorch_lightning.utilities.cli import LightningCLI

class MyLightningCLI(LightningCLI):

    def add_arguments_to_parser(self, parser):
        parser.link_arguments('data.batch_size', 'model.batch_size')

cli = MyLightningCLI(MyModel, MyDataModule)

The linking of arguments is observed in the help of the tool, which for this example would look like:

$ python trainer.py --help
  ...
    --data.batch_size BATCH_SIZE
                          Number of samples in a batch (type: int, default: 8)

  Linked arguments:
    model.batch_size <-- data.batch_size
                          Number of samples in a batch (type: int)

Sometimes a parameter value is only available after class instantiation. An example could be that your model requires the number of classes to instantiate its fully connected layer (for a classification task) but the value is not available until the data module has been instantiated. The code below illustrates how to address this.

from pytorch_lightning.utilities.cli import LightningCLI

class MyLightningCLI(LightningCLI):

    def add_arguments_to_parser(self, parser):
        parser.link_arguments('data.num_classes', 'model.num_classes', apply_on='instantiate')

cli = MyLightningCLI(MyClassModel, MyDataModule)

Instantiation links are used to automatically determine the order of instantiation, in this case data first.

Tip

The linking of arguments can be used for more complex cases. For example to derive a value via a function that takes multiple settings as input. For more details have a look at the API of link_arguments.

Optimizers and learning rate schedulers

Optimizers and learning rate schedulers can also be made configurable. The most common case is when a model only has a single optimizer and optionally a single learning rate scheduler. In this case the model’s LightningModule could be left without implementing the configure_optimizers method since it is normally always the same and just adds boilerplate. The following code snippet shows how to implement it:

import torch
from pytorch_lightning.utilities.cli import LightningCLI

class MyLightningCLI(LightningCLI):

    def add_arguments_to_parser(self, parser):
        parser.add_optimizer_args(torch.optim.Adam)
        parser.add_lr_scheduler_args(torch.optim.lr_scheduler.ExponentialLR)

cli = MyLightningCLI(MyModel)

With this the configure_optimizers method is automatically implemented and in the config the optimizer and lr_scheduler groups would accept all of the options for the given classes, in this example Adam and ExponentialLR. Therefore, the config file would be structured like:

optimizer:
  lr: 0.01
lr_scheduler:
  gamma: 0.2
model:
  ...
trainer:
  ...

And any of these arguments could be passed directly through command line. For example:

$ python train.py --optimizer.lr=0.01 --lr_scheduler.gamma=0.2

There is also the possibility of selecting among multiple classes by giving them as a tuple. For example:

class MyLightningCLI(LightningCLI):

    def add_arguments_to_parser(self, parser):
        parser.add_optimizer_args((torch.optim.SGD, torch.optim.Adam))

In this case in the config the optimizer group instead of having directly init settings, it should specify class_path and optionally init_args. Sub-classes of the classes in the tuple would also be accepted. A corresponding example of the config file would be:

optimizer:
  class_path: torch.optim.Adam
  init_args:
    lr: 0.01

And the same through command line:

$ python train.py --optimizer.class_path=torch.optim.Adam --optimizer.init_args.lr=0.01

The automatic implementation of configure_optimizers can be disabled by linking the configuration group. An example can be ReduceLROnPlateau which requires to specify a monitor. This would be:

from pytorch_lightning.utilities.cli import instantiate_class, LightningCLI

class MyModel(LightningModule):

    def __init__(self, optimizer_init: dict, lr_scheduler_init: dict):
        super().__init__()
        self.optimizer_init = optimizer_init
        self.lr_scheduler_init = lr_scheduler_init

    def configure_optimizers(self):
        optimizer = instantiate_class(self.parameters(), self.optimizer_init)
        scheduler = instantiate_class(optimizer, self.lr_scheduler_init)
        return {"optimizer": optimizer, "lr_scheduler": scheduler, "monitor": "metric_to_track"}

class MyLightningCLI(LightningCLI):

    def add_arguments_to_parser(self, parser):
        parser.add_optimizer_args(
            torch.optim.Adam,
            link_to='model.optimizer_init',
        )
        parser.add_lr_scheduler_args(
            torch.optim.lr_scheduler.ReduceLROnPlateau,
            link_to='model.lr_scheduler_init',
        )

cli = MyLightningCLI(MyModel)

For both possibilities of using pytorch_lightning.utilities.cli.LightningArgumentParser.add_optimizer_args() with a single class or a tuple of classes, the value given to optimizer_init will always be a dictionary including class_path and init_args entries. The function instantiate_class() takes care of importing the class defined in class_path and instantiating it using some positional arguments, in this case self.parameters(), and the init_args. Any number of optimizers and learning rate schedulers can be added when using link_to.

Read the Docs v: 1.4.0
Versions
latest
stable
1.4.0
1.3.8
1.3.7
1.3.6
1.3.5
1.3.4
1.3.3
1.3.2
1.3.1
1.3.0
1.2.10
1.2.8
1.2.7
1.2.6
1.2.5
1.2.4
1.2.3
1.2.2
1.2.1
1.2.0
1.1.8
1.1.7
1.1.6
1.1.5
1.1.4
1.1.3
1.1.2
1.1.1
1.1.0
1.0.8
1.0.7
1.0.6
1.0.5
1.0.4
1.0.3
1.0.2
1.0.1
1.0.0
0.10.0
0.9.0
0.8.5
0.8.4
0.8.3
0.8.2
0.8.1
0.8.0
0.7.6
0.7.5
0.7.4
0.7.3
0.7.2
0.7.1
0.7.0
0.6.0
0.5.3
0.4.9
ipynb-update
docs-search
Downloads
html
On Read the Docs
Project Home
Builds

Free document hosting provided by Read the Docs.