class Configurable[source]#

Class for load/save configuration. It will automatically record every argument passed to the __init__ function.

This class is inspired by state_dict() in PyTorch, but designed for hyperparameters.

Inherit this class to construct a configurable class.

>>> class MyClass(nn.Module, core.Configurable):

Note Configurable only applies to the current class rather than any derived class. For example, the following definition only records the arguments of MyClass.

>>> class DerivedClass(MyClass):

In order to record the arguments of DerivedClass, explicitly specify the inheritance.

>>> class DerivedClass(MyClass, core.Configurable):

To get the configuration of an instance, use config_dict(), which returns a dict of argument names and values. If an argument is also an instance of Configurable, it will be recursively expanded in the dict. The configuration dict can be passed to load_config_dict() to create a copy of the instance.

For classes already registered in Registry, they can be directly created from the Configurable class. This is convenient for building models from configuration files.

>>> config = models.GCN(128, [128]).config_dict()
>>> gcn = Configurable.load_config_dict(config)


class Engine(task, train_set, valid_set, test_set, optimizer, scheduler=None, gpus=None, batch_size=1, gradient_interval=1, num_worker=0, logger='logging', log_interval=100)[source]#

General class that handles everything about training and test of a task.

This class can perform synchronous distributed parallel training over multiple CPUs or GPUs. To invoke parallel training, launch with one of the following commands.

  1. Single-node multi-process case.

python -m torch.distributed.launch --nproc_per_node={number_of_gpus} {} {your_arguments...}
  1. Multi-node multi-process case.

python -m torch.distributed.launch --nnodes={number_of_nodes} --node_rank={rank_of_this_node}
--nproc_per_node={number_of_gpus} {} {your_arguments...}

If preprocess() is defined by the task, it will be applied to train_set, valid_set and test_set.

  • task (nn.Module) – task

  • train_set (data.Dataset) – training set

  • valid_set (data.Dataset) – validation set

  • test_set (data.Dataset) – test set

  • optimizer (optim.Optimizer) – optimizer

  • scheduler (lr_scheduler._LRScheduler, optional) – scheduler

  • gpus (list of int, optional) – GPU ids. By default, CPUs will be used. For multi-node multi-process case, repeat the GPU ids for each node.

  • batch_size (int, optional) – batch size of a single CPU / GPU

  • gradient_interval (int, optional) – perform a gradient update every n batches. This creates an equivalent batch size of batch_size * gradient_interval for optimization.

  • num_worker (int, optional) – number of CPU workers per GPU

  • logger (str or core.LoggerBase, optional) – logger type or logger instance. Available types are logging and wandb.

  • log_interval (int, optional) – log every n gradient updates

evaluate(split, log=True)[source]#

Evaluate the model.

  • split (str) – split to evaluate. Can be train, valid or test.

  • log (bool, optional) – log metrics or not



Return type


load(checkpoint, load_optimizer=True, strict=True)[source]#

Load a checkpoint from file.

  • checkpoint (file-like) – checkpoint file

  • load_optimizer (bool, optional) – load optimizer state or not

  • strict (bool, optional) – whether to strictly check the checkpoint matches the model parameters

classmethod load_config_dict(config)[source]#

Construct an instance from the configuration dict.


Save checkpoint to file.


checkpoint (file-like) – checkpoint file

train(num_epoch=1, batch_per_epoch=None)[source]#

Train the model.

If batch_per_epoch is specified, randomly draw a subset of the training set for each epoch. Otherwise, the whole training set is used for each epoch.

  • num_epoch (int, optional) – number of epochs

  • batch_per_epoch (int, optional) – number of batches per epoch

property epoch#

Current epoch.


class Meter(log_interval=100, silent=False, logger=None)[source]#

Meter for recording metrics and training progress.

  • log_interval (int, optional) – log every n updates

  • silent (int, optional) – surpress all outputs or not

  • logger (core.LoggerBase, optional) – log handler

log(record, category='train/batch')[source]#

Log a record.

  • record (dict) – dict of any metric

  • category (str, optional) – log category. Available types are train/batch, train/epoch, valid/epoch and test/epoch.


Log a hyperparameter config.


config (dict) – hyperparameter config


Step an epoch for this meter.

Instead of manually invoking step(), it is suggested to use the following line

>>> for epoch in meter(num_epoch):
>>>     # do something

Update the meter with a record.


record (dict) – dict of any metric


class Registry[source]#

Registry class for managing all call-by-name access to objects.

Typical scenarios:

  1. Create a model according to a string.

>>> gcn ="GCN")(128, [128])
  1. Register a customize hook to the package.

>>> @R.register("features.atom.my_feature")
>>> def my_featurizer(atom):
>>>     ...
>>> data.Molecule.from_smiles("C1=CC=CC=C1", atom_feature="my_feature")
classmethod get(name)[source]#

Get an object with a canonical name. Hierarchical names are separated by ..

classmethod register(name)[source]#

Register an object with a canonical name. Hierarchical names are separated by ..

classmethod search(name)[source]#

Search an object with the given name. The name doesn’t need to be canonical.

For example, we can search GCN and get the object of models.GCN.

Meta Container#

class _MetaContainer(meta_dict=None, **kwargs)[source]#

Meta container that maintains meta types about members.

The meta type of each member is tracked when a member is assigned. We use a context manager to define the meta types for a bunch of assignment.

The meta types are stored as a dict in instance.meta_dict, where keys are member names and values are meta types.

>>> class MyClass(_MetaContainer):
>>>     ...
>>> instance = MyClass()
>>> with instance.context("important"):
>>>     instance.value = 1
>>> assert instance.meta_dict["value"] == "important"

Members assigned with context(None) or without a context won’t be tracked.

>>> instance.random = 0
>>> assert "random" not in instance.meta_dict

You can also restrict available meta types by defining a set _meta_types in the derived class.


Meta container also supports auto inference of meta types. This can be enabled by setting enable_auto_context to True in the derived class.

Once auto inference is on, any member without an explicit context will be recognized through their name prefix. For example, instance.node_value will be recognized as node if node is defined in meta_types.

This may make code hard to maintain. Use with caution.


Context manager for assigning members with a specific meta type.

data_by_meta(include=None, exclude=None)[source]#

Return members based on the specific meta types.

  • include (list of string, optional) – meta types to include

  • exclude (list of string, optional) – meta types to exclude


data member dict and meta type dict

Return type

(dict, dict)

property data_dict#

A dict that maps tracked names to members.