Instantiating objects with Hydra
One of the best ways to drive different behavior in an application is to instantiate different implementations of an interface. The code using the instantiated object only knows the interface which remains constant, but the behavior is determined by the actual object instance.
Hydra provides hydra.utils.instantiate() (and its alias hydra.utils.call()) for instantiating objects and calling functions. Prefer instantiate for creating objects and call for invoking functions.
Call/instantiate supports:
- Constructing an object by calling the
__init__method - Calling functions, static functions, class methods and other callable global objects
Instantiate API (Expand for details)
def instantiate(config: Any, *args: Any, **kwargs: Any) -> Any:
"""
:param config: An config object describing what to call and what params to use.
In addition to the parameters, the config must contain:
_target_ : target class or callable name (str)
And may contain:
_args_: List-like of positional arguments to pass to the target
_recursive_: Construct nested objects as well (bool).
True by default.
may be overridden via a _recursive_ key in
the kwargs
_convert_: Conversion strategy
none : Passed objects are DictConfig and ListConfig, default
partial : Passed objects are converted to dict and list, with
the exception of Structured Configs (and their fields).
object : Passed objects are converted to dict and list.
Structured Configs are converted to instances of the
backing dataclass / attr class.
all : Passed objects are dicts, lists and primitives without
a trace of OmegaConf containers. Structured configs
are converted to dicts / lists too.
_partial_: If True, return functools.partial wrapped method or object
False by default. Configure per target.
:param args: Optional positional parameters pass-through
:param kwargs: Optional named parameters to override
parameters in the config object. Parameters not present
in the config objects are being passed as is to the target.
IMPORTANT: dataclasses instances in kwargs are interpreted as config
and cannot be used as passthrough
:return: if _target_ is a class name: the instantiated object
if _target_ is a callable: the return value of the call
"""
# Alias for instantiate
call = instantiate
The config passed to these functions must have a key called _target_, with the value of a fully qualified class name, class method, static method or callable.
For convenience, None config results in a None object.
Named arguments : Config fields (except reserved fields like _target_) are passed as named arguments to the target.
Named arguments in the config can be overridden by passing named argument with the same name in the instantiate() call-site.
Positional arguments : The config may contain a _args_ field representing positional arguments to pass to the target.
The positional arguments can be overridden together by passing positional arguments in the instantiate() call-site.
Simple usage
Your application might have an Optimizer class:
class Optimizer:
algo: str
lr: float
def __init__(self, algo: str, lr: float) -> None:
self.algo = algo
self.lr = lr
optimizer:
_target_: my_app.Optimizer
algo: SGD
lr: 0.01
opt = instantiate(cfg.optimizer)
print(opt)
# Optimizer(algo=SGD,lr=0.01)
# override parameters on the call-site
opt = instantiate(cfg.optimizer, lr=0.2)
print(opt)
# Optimizer(algo=SGD,lr=0.2)
Recursive instantiation
Let's add a Dataset and a Trainer class. The trainer holds a Dataset and an Optimizer instances.
class Dataset:
name: str
path: str
def __init__(self, name: str, path: str) -> None:
self.name = name
self.path = path
class Trainer:
def __init__(self, optimizer: Optimizer, dataset: Dataset) -> None:
self.optimizer = optimizer
self.dataset = dataset
With the following config, you can instantiate the whole thing with a single call:
trainer:
_target_: my_app.Trainer
optimizer:
_target_: my_app.Optimizer
algo: SGD
lr: 0.01
dataset:
_target_: my_app.Dataset
name: Imagenet
path: /datasets/imagenet
Hydra will instantiate nested objects recursively by default.
trainer = instantiate(cfg.trainer)
print(trainer)
# Trainer(
# optimizer=Optimizer(algo=SGD,lr=0.01),
# dataset=Dataset(name=Imagenet, path=/datasets/imagenet)
# )
You can override parameters for nested objects:
trainer = instantiate(
cfg.trainer,
optimizer={"lr": 0.3},
dataset={"name": "cifar10", "path": "/datasets/cifar10"},
)
print(trainer)
# Trainer(
# optimizer=Optimizer(algo=SGD,lr=0.3),
# dataset=Dataset(name=cifar10, path=/datasets/cifar10)
# )
Similarly, positional arguments of nested objects can be overridden:
obj = instantiate(
cfg.object,
# pass 1 and 2 as positional arguments to the target object
1, 2,
# pass 3 and 4 as positional arguments to a nested child object
child={"_args_": [3, 4]},
)
Disable recursive instantiation
You can disable recursive instantiation by setting _recursive_ to False in the config node or in the call-site
In that case the Trainer object will receive an OmegaConf DictConfig for nested dataset and optimizer instead of the instantiated objects.
optimizer = instantiate(cfg.trainer, _recursive_=False)
print(optimizer)
Output:
Trainer(
optimizer={
'_target_': 'my_app.Optimizer', 'algo': 'SGD', 'lr': 0.01
},
dataset={
'_target_': 'my_app.Dataset', 'name': 'Imagenet', 'path': '/datasets/imagenet'
}
)
Parameter conversion strategies
By default, the parameters passed to the target are either primitives (int,
float, bool etc) or OmegaConf containers (DictConfig, ListConfig).
OmegaConf containers have many advantages over primitive dicts and lists,
including convenient attribute access for keys,
duck-typing as instances of dataclasses or attrs classes, and
support for variable interpolation
and custom resolvers.
If the callable targeted by instantiate leverages OmegaConf's features, it
will make sense to pass DictConfig and ListConfig instances directly to
that callable.
That being said, in many cases it's desired to pass normal Python dicts and
lists, rather than DictConfig or ListConfig instances, as arguments to your
callable. You can change instantiate's argument conversion strategy using the
_convert_ parameter. Supported values are:
"none": Default behavior, Use OmegaConf containers"partial": Convert OmegaConf containers to dict and list, except Structured Configs, which remain as DictConfig instances."object": Convert OmegaConf containers to dict and list, except Structured Configs, which are converted to instances of the backing dataclass / attr class usingOmegaConf.to_object."all": Convert everything to primitive containers
The conversion strategy applies recursively to all subconfigs of the instantiation target. Here is an example demonstrating the various conversion strategies:
from dataclasses import dataclass
from omegaconf import DictConfig, OmegaConf
from hydra.utils import instantiate
@dataclass
class Foo:
a: int = 123
class MyTarget:
def __init__(self, foo, bar):
self.foo = foo
self.bar = bar
cfg = OmegaConf.create(
{
"_target_": "__main__.MyTarget",
"foo": Foo(),
"bar": {"b": 456},
}
)
obj_none = instantiate(cfg, _convert_="none")
assert isinstance(obj_none, MyTarget)
assert isinstance(obj_none.foo, DictConfig)
assert isinstance(obj_none.bar, DictConfig)
obj_partial = instantiate(cfg, _convert_="partial")
assert isinstance(obj_partial, MyTarget)
assert isinstance(obj_partial.foo, DictConfig)
assert isinstance(obj_partial.bar, dict)
obj_object = instantiate(cfg, _convert_="object")
assert isinstance(obj_object, MyTarget)
assert isinstance(obj_object.foo, Foo)
assert isinstance(obj_object.bar, dict)
obj_all = instantiate(cfg, _convert_="all")
assert isinstance(obj_none, MyTarget)
assert isinstance(obj_all.foo, dict)
assert isinstance(obj_all.bar, dict)
Passing the _convert_ keyword argument to instantiate has the same effect as defining
a _convert_ attribute on your config object. Here is an example creating
instances of MyTarget that are equivalent to the above:
cfg_none = OmegaConf.create({..., "_convert_": "none"})
obj_none = instantiate(cfg_none)
cfg_partial = OmegaConf.create({..., "_convert_": "partial"})
obj_partial = instantiate(cfg_partial)
cfg_object = OmegaConf.create({..., "_convert_": "object"})
obj_object = instantiate(cfg_object)
cfg_all = OmegaConf.create({..., "_convert_": "all"})
obj_all = instantiate(cfg_all)
Partial Instantiation
Sometimes you may not set all parameters needed to instantiate an object from the configuration, in this case you can set
_partial_ to be True to get a functools.partial wrapped object or method, then complete initializing the object in
the application code. Here is an example:
class Optimizer:
algo: str
lr: float
def __init__(self, algo: str, lr: float) -> None:
self.algo = algo
self.lr = lr
def __repr__(self) -> str:
return f"Optimizer(algo={self.algo},lr={self.lr})"
class Model:
def __init__(self, optim_partial: Any, lr: float):
super().__init__()
self.optim = optim_partial(lr=lr)
self.lr = lr
def __repr__(self) -> str:
return f"Model(Optimizer={self.optim},lr={self.lr})"
model:
_target_: my_app.Model
optim_partial:
_partial_: true
_target_: my_app.Optimizer
algo: SGD
lr: 0.01
model = instantiate(cfg.model)
print(model)
# "Model(Optimizer=Optimizer(algo=SGD,lr=0.01),lr=0.01)
If you are repeatedly instantiating the same config,
using _partial_=True may provide a significant speedup as compared with regular (non-partial) instantiation.
factory = instantiate(config, _partial_=True)
obj = factory()
In the above example, repeatedly calling factory would be faster than repeatedly calling instantiate(config).
A caveat of this approach is that the same keyword arguments would be re-used in each call to factory.
class Foo:
...
class Bar:
def __init__(self, foo):
self.foo = foo
bar_conf = {
"_target_": "__main__.Bar",
"foo": {"_target_": "__main__.Foo"},
}
bar_factory = instantiate(bar_conf, _partial_=True)
bar1 = bar_factory()
bar2 = bar_factory()
assert bar1 is not bar2
assert bar1.foo is bar2.foo # the `Foo` instance is re-used here
This does not apply if _partial_=False,
in which case a new Foo instance would be created with each call to instantiate.
Instantiation of builtins
The value of _target_ passed to instantiate should be a "dotpath" pointing
to some callable that can be looked up via a combination of import and getattr.
If you want to target one of Python's built-in functions (such as len or print or divmod),
you will need to provide a dotpath looking up that function in Python's builtins module.
from hydra.utils import instantiate
# instantiate({"_target_": "len"}, [1,2,3]) # this gives an InstantiationException
instantiate({"_target_": "builtins.len"}, [1,2,3]) # this works, returns the number 3
Dotpath lookup machinery
Hydra looks up a given _target_ by attempting to find a module that
corresponds to a prefix of the given dotpath and then looking for an object in
that module corresponding to the dotpath's tail. For example, to look up a _target_
given by the dotpath "my_module.my_nested_module.my_object", hydra first locates
the module my_module.my_nested_module, then find my_object inside that nested module.
Hydra exposes an API allowing direct use of this dotpath lookup machinery. The following two functions, which can be imported from the
def get_class(path: str) -> type:
"""
Look up a class based on a dotpath.
Fails if the path does not point to a class.
>>> import my_module
>>> from hydra.utils import get_class
>>> assert get_class("my_module.MyClass") is my_module.MyClass
"""
...
def get_method(path: str) -> Callable[..., Any]:
"""
Look up a callable based on a dotpath.
Fails if the path does not point to a callable object.
>>> import my_module
>>> from hydra.utils import get_method
>>> assert get_method("my_module.my_function") is my_module.my_function
"""
...
# Alias for get_method
get_static_method = get_method