/build

/build

https://github.com/wyli/MONAI/blob/ac26db9ba5672c45cb904c5ad426383846f95a58/tests/test_config_parser.py#L253

This doesn't cover the following case where the model is an instantiated object and not a partial function. Picked EfficientNetBN for no particular reason. I've changed $@model().forward to @model.forward as I believe that that would be the expected format.

configs = {
            "fwd": {"_target_": "@model.forward", "x": "$torch.rand(1, 3, 256, 256)"},
            "model": {"_target_": "monai.networks.nets.EfficientNetBN",  "model_name": "efficientnet-b0", "pretrained": False, "spatial_dims": 2},
        }

thanks @ibro45, in that case we still need the $ so that .forward is treated as a python dot-notation to access the instance method:

"fwd": {"_target_": "[email protected]", "inputs": "$torch.rand(1, 3, 256, 256)"},

also because an instance of EfficientNetBN is also a callable, "@model" works as well:

configs = {
    "fwd": {"_target_": "@model", "inputs": "$torch.rand(1, 3, 256, 256)"},
    "model": {
        "_target_": "monai.networks.nets.EfficientNetBN",
        "model_name": "efficientnet-b0",
        "pretrained": False,
        "spatial_dims": 2,
    },
}
print(ConfigParser(configs).fwd())

Please feel free to make a PR to add more test cases if you are interested :)

@wyli when you parse the config it returns a partial function that you then manually call here print(ConfigParser(configs).fwd())

Reproduce

config.yaml:

fwd1:
    _target_: "[email protected]"
    inputs: "$torch.zeros(1, 3, 256, 256)"

fwd2: "[email protected](inputs=torch.zeros(1, 3, 256, 256))"

model:
    _target_: "monai.networks.nets.EfficientNetBN"
    model_name: "efficientnet-b0"
    pretrained: True
    spatial_dims: 2

Run fwd1:
python -m monai.bundle run fwd1 --config_file config.yaml
Returns a partial function that didn't run.

Run fwd2:
python -m monai.bundle run fwd2 --config_file config.yaml
Returns the output tensor of the model.

Do you agree that the fwd1 and fwd2 should return the same output, or is it just that I don't understand how its meant to work?

Reasoning

I want to call a method using _target_ and pass the arguments to it through CLI. Let's say:
python -m monai.bundle run fwd1 --config_file config.yaml --fwd1#inputs "$torch.rand(1,3,512,512)". That isn't possible using the format in fwd2 unless you rewrite it something like this:

fwd2:
    _method_: "[email protected](@fwd#inputs)
    inputs: "$torch.zeros(1, 3, 256, 256)"

This allows you to pass the arguments from CLI at the expense of making it very cluttered.

_partial_

I might just be used to Hydra's _partial_, but I think it makes sense. If you want a function, method, or even class init to be partial, we could have a _partial_ flag. Otherwise, it will just be instantiated/called.

Let me know if I should open an issue for this.

This seems like a helpful feature to me as well! Having the ability to both call a function through instantiate and create a partial function.

I would've expected the default config to call the function because I mostly worked with hydra and am used to that being their default behavior.

I think _target_ is mainly to create the object instance, to get the concrete result, we can create an "infer" command:

configs = {
    "fwd": {"_target_": "@model", "inputs": "$torch.rand(1, 3, 256, 256)"},
    "model": {
        "_target_": "monai.networks.nets.EfficientNetBN",
        "model_name": "efficientnet-b0",
        "pretrained": False,
        "spatial_dims": 2,
    },
    "infer": "$@fwd()",
}
print(ConfigParser(configs).infer)

cc @Nic-Ma @ericspod