Custom method/arguments

Method arguments

Arguments of method can be variety type, including number, string, list, dict. It depends on the definition of the method. In addition, argument an be emitted in configuration if a default value is given in the definition of method.

For example, random rate position selector RandomPositionByRate receives a float parameter rate, while fixed positions selector FixPositions receives a list of positions.

Use custom fault injection method/function

Custom methods can be defined according to corresponding requirements, see observer doc, selector doc, error mode doc, quantization doc. Then use mrfi.add_function to let MRFI know it. The custom method can be used in both EasyConfig and fine-grained ConfigTree.

The following code adds two trivial method to MRFI.

from mrfi import add_function

def MyTestErrorMode(x, custom_val = 0):
    # set all selected 
    return torch.full_like(x, val)

class MyNoQuantization:
    @staticmethod
    def quantize(x, custom_arg):
        pass

    @staticmethod
    def dequantize(x, custom_arg):
        pass

if __name__ == '__main__':
    add_function('NoQuantization', MyNoQuantization)
    add_function('TestErrorMode', MyTestErrorMode)

To use the custom method, the EasyConfig can be like this:

faultinject:
  - type: activation
    quantization:
      method: NoQuantization
      custom_arg: xxx  # can not be emitted
    error_mode:
      method: MyTestErrorMode
      custom_val: 16 # can be emitted because defaut value gived

If you think your function is commonly used and not provided by MRFI, welcome to add it to MRFI's built-in methods through Pull Request on Github.