Some model training parameters in a configuration are not being used

[TS-Lee]

Some parameters are not being drawn from the configuration file, but hard coded instead. For example,

• input shape: https://github.com/twosixlabs/armory/blob/53a7ccbd122ce3e05118416ba0d5442dee737fb9/armory/baseline_models/pytorch/resnet18.py#L56
• input shape: https://github.com/twosixlabs/armory/blob/53a7ccbd122ce3e05118416ba0d5442dee737fb9/armory/baseline_models/pytorch/micronnet_gtsrb.py#L76
• learning rate: https://github.com/twosixlabs/armory/blob/53a7ccbd122ce3e05118416ba0d5442dee737fb9/armory/baseline_models/pytorch/micronnet_gtsrb.py#L74

[TS-Lee]

• learning rate: https://github.com/twosixlabs/armory/blob/53a7ccbd122ce3e05118416ba0d5442dee737fb9/armory/baseline_models/pytorch/cifar.py#L60

[TS-Lee]

Maybe, a desired behavior is that a certain parameter is not used but specified, it should warn or error.

[Nathalie-B]

@davidslater ^^ I wanted to get this in your radar. It may lead to errors in the future

[davidslater]

Some of those mentioned above are actually using the config wrapper_kwargs, if they exist. For example, in the resnet18 one: https://github.com/twosixlabs/armory/blob/53a7ccbd122ce3e05118416ba0d5442dee737fb9/armory/baseline_models/pytorch/resnet18.py#L55-L57

        input_shape=wrapper_kwargs.pop(
            "input_shape", (224, 224, 3)
        ),  # default to imagenet shape

Many of the others are indeed hard coded. However, if they are added to the wrapper_kwargs in the config, they will error here (or similar places in the other files): https://github.com/twosixlabs/armory/blob/53a7ccbd122ce3e05118416ba0d5442dee737fb9/armory/baseline_models/pytorch/micronnet_gtsrb.py#L79

        **wrapper_kwargs

For instance, if I add this line to the cifar10_baseline.json:

        "wrapper_kwargs": {
            "input_shape": [64, 64, 3]
        }

I will get this error when running armory:

armory run scenario_configs/cifar10_baseline.json --check
...
File "/workspace/armory/baseline_models/pytorch/cifar.py", line 57, in get_art_model
wrapped_model = PyTorchClassifier(
                └ <class 'art.estimators.classification.pytorch.PyTorchClassifier'>
TypeError:
InputFilter object got multiple values for keyword argument 'input_shape'