Cannot determine the import path of callable object

[vinnamkim]

🐛 Bug report

Same as title. Please see the following reproducible Python script.

To reproduce

# test.py
from jsonargparse import ActionConfigFile, ArgumentParser

from lightning.pytorch.cli import OptimizerCallable
from torch.optim import SGD


class MyOptimizerCallable:
    def __init__(self, lr: float = 0.1):
        self.lr = lr

    def __call__(self, params) -> SGD:
        return SGD(params, self.lr)


class MyModel:
    def __init__(self, optimizer: OptimizerCallable = MyOptimizerCallable(lr=0.2)):
        self.optimizer = optimizer


parser = ArgumentParser()
parser.add_argument("--model", type=MyModel)
parser.add_argument("--config", action=ActionConfigFile)
cfg = parser.parse_args()

Then, execute this Python script such as

python test.py --model __main__.MyModel --print_config

Expected behavior

Expect the following configuration

model:
  class_path: __main__.MyModel
  init_args:
    optimizer:
        class_path: __main__.MyOptimizerCallable
        init_args:
            lr: 0.2

However, the actual result is

usage: test.py [-h] [--model.help CLASS_PATH_OR_NAME] [--model MODEL] [--config CONFIG] [--print_config[=flags]]
error: Parser key "model":
  Not possible to determine the import path for object <__main__.MyOptimizerCallable object at 0x789147d9f0d0>.

Environment

• jsonargparse version (e.g., 4.8.0): 4.27.6
• Python version (e.g., 3.9): 3.11
• How jsonargparse was installed (e.g. pip install jsonargparse[all]): pip install jsonargparse
• OS (e.g., Linux): Linux

[mauvilsa]

There isn't any bug here. The only thing that jsonargparse knows is that the default of optimizer is an instance of MyOptimizerCallable. It doesn't know how this instance was created. And the error is because the instance is anonymous, only created in the class signature. It is not possible to import that object from that module. For a callable to be importable, it would need to be a function defined directly in the module, as:

def my_default_optimizer_callable(params) -> Optimizer: ...

or be an instance defined directly in the module, as:

my_default_optimizer_callable = MyOptimizerCallable(lr=0.2)

Then the parameter defined as optimizer: OptimizerCallable = my_default_optimizer_callable. But, when defined like this, the printed config will not be as above. It will not have class_path and init_args, but just the import path of the callable.

Also note that using class instances as defaults is discouraged. See class-type-defaults to understand why. If you use lazy_instance, then jsonargparse would be able to know how to create the instance and the printed config would be like the one above.

[vinnamkim]

Hi @mauvilsa,
Thanks for your kind explanation. From your guide, it seems a best practice to use lazy_instance for this problem. However, I faced this issue to apply lazy_instance to my callable class.

# test.py
from jsonargparse import ActionConfigFile, ArgumentParser, lazy_instance

...

class MyModel:
    def __init__(self, optimizer: OptimizerCallable = lazy_instance(MyOptimizerCallable(lr=0.2))):
        self.optimizer = optimizer

...

model = MyModel()

The instantiation (model = MyModel()) in Python code produces the following error.

ValueError: Problem with given class_path '__main__.MyModel':
  'LazyInstance_MyOptimizerCallable' object has no attribute 'lr'

This seems that MyModel.__call__() cannot do lazy initialization correctly. I could manage it by adding this overriding to

jsonargparse/jsonargparse/_typehints.py

Line 1286 in 2bc81ea

class LazyInitBaseClass:

class LazyInitBaseClass:
    ...
    def __call__(self, *args, **kwargs):
        if call := self.__dict__.get("__call__", None):
            return call(*args, **kwargs)

Can we regard this as a bug?

[mauvilsa]

In the description it says you are using v4.27.6, which is not the latest. If so, please upgrade and try again. It could be already fixed by #473.

[vinnamkim]

In the description it says you are using v4.27.6, which is not the latest. If so, please upgrade and try again. It could be already fixed by #473.

I upgraded it to 4.27.7 but I still got an error for this script

from jsonargparse import ActionConfigFile, ArgumentParser, lazy_instance

from lightning.pytorch.cli import OptimizerCallable
from torch.optim import SGD
from torch import nn


class MyOptimizerCallable:
    def __init__(self, lr: float = 0.1):
        self.lr = lr

    def __call__(self, params) -> SGD:
        return SGD(params, lr=self.lr)


class MyModel:
    def __init__(self, optimizer: OptimizerCallable = lazy_instance(MyOptimizerCallable, lr=0.2)):
        model = nn.Linear(10, 10)
        self.optimizer = optimizer(model.parameters())
        print(self.optimizer)


parser = ArgumentParser()
parser.add_argument("--model", type=MyModel)
parser.add_argument("--config", action=ActionConfigFile)
cfg = parser.parse_args()

print(parser.instantiate_classes(cfg))
print(MyModel())

$ python test5.py --model __main__.MyModel

SGD (
Parameter Group 0
    dampening: 0
    differentiable: False
    foreach: None
    lr: 0.2
    maximize: False
    momentum: 0
    nesterov: False
    weight_decay: 0
)
Namespace(model=<__main__.MyModel object at 0x74bfef32b790>, config=None)
Traceback (most recent call last):
  File "/home/vinnamki/otx/training_extensions/test5.py", line 29, in <module>
    print(MyModel())
          ^^^^^^^^^
  File "/home/vinnamki/otx/training_extensions/test5.py", line 19, in __init__
    self.optimizer = optimizer(model.parameters())
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/vinnamki/otx/training_extensions/test5.py", line 13, in __call__
    return SGD(params, lr=self.lr)
                          ^^^^^^^
AttributeError: 'LazyInstance_MyOptimizerCallable' object has no attribute 'lr'

[mauvilsa]

Okay, I need to look at this.

[mauvilsa]

This is a bug. But the fix can't be like suggested in #481 (comment), since this would make all lazy instances callable, which shouldn't be the case. I will push a fix in the next days.

[mauvilsa]

@vinnamkim I have pushed a fix in #487. Please try it out.