mrfi

Core module of MRFI.

EasyConfig(config=None)

EasyConfig object.

Use classmethod EasyConfig.load_[file|string|preset]() to construct a EasyConfig.

Properties

• faultinject: List of fault injectors.
• observe: List of observers.

Load a dict-like configuration or a EasyConfig preset.

Parameters:

Name	Type	Description	Default
config	Optional[dict]	if dict, load the dict config. if None, make a empty EasyConfig.	None

load_file(filename) classmethod

Load EasyConfig from .yaml File.

load_string(string) classmethod

Load EasyConfig from a yaml string.

load_preset(name) classmethod

Load EasyConfig from presets.

ConfigTreeNodeType

Bases: Enum

Config tree node types.

Note

A config tree has recursion structure:

• MODULE
  • FI_LIST (e.g. activation_in, weights)
    • FI_ATTR
      • FI_STAGE (e.g. quantization, selector)
        • METHOD_ARGS (e.g. dynamic_range, rate)
  • OBSERVER_LIST
    • OBSERVER_ATTR
  • MODULE_LIST (sub_modules)
    • MODULE
      • FI_LIST
      • ...

ConfigTree(config_node, mrfi, nodetype=ConfigTreeNodeType.MODULE, name='model')

Make a detail config tree.

Detail config tree is a dict-like object.

Parameters:

Name	Type	Description	Default
config_node	dict	A dict with ConfigTreeNodeType structrue.	required
mrfi	MRFI	MRFI object	required
name	str	config node name for visualize.	'model'

append(obj)

Append a object on a list node

remove(index)

Remove object by index

state_dict()

Visualize or export config tree node and all sub node.

hasattr(name)

Determine if an attr is in config node.

getattr(name)

Return an attr in config node, if not exist, return None

set_internal_attr(key, name)

__setattr__ reject bind new attr, use this if necessary.

ConfigItemList

Bases: list

A convient list of detail config tree allow broadcasting assignment.

example

cfg = fi_model.get_configs('weights.0.selector')
cfg.enable = False
cfg[1].enable = True

MRFI(model, config)

MRFI core object, a wrapper of network module.

__change_arg(fiattr, keep_arg=['method'])

Add an args level

golden_run()

A context manager for golden run.

Examples:

>>> with fi_model.golden_run():
>>>    out_golden = fi_model(inputs)

get_configs(configname=None, strict=False, **kwargs)

Find config tree nodes in current detail config.

Parameters:

Name	Type	Description	Default
configname	str	A string splited by dot '.' to match ConfigTreeNode, e.g. 'activations.0.selector'.	None
strict	bool	Use strict match config name on found modules.	False
module_name	Union[str, List[str]]	Found modules by module name, e.g. 'conv'.	required
module_type	Union[str, List[str]]	Found modules by module type, e.g. 'Conv2d'.	required
module_fullname	Union[str, List[str]]	Found modules by strict module name match, e.g. ['model.conv1', 'model.conv2'].	required

get_activation_configs(fi_configname=None, strict=False, activation_id=0, is_out=False, **kwargs)

A convenient function of get_configs to get activation config.

Equivalent to call get_configs('activation.0.{fi_configname}') or get_configs('activation_out.0.{fi_configname}'). when activation_id=0.

If first parameter fi_configname is not specified, return the root nodes of FI_CONFIG, usually used for settingenabled.

get_weights_configs(fi_configname=None, strict=False, weight_id=0, **kwargs)

A convenient function of get_configs to get weights config.

Equivalent to call get_configs('weights.0.{fi_configname}') when weight_id=0.

If first parameter fi_configname is not specified, return the root nodes of FI_CONFIG, usually used for settingenabled.

save_config(filename)

Save current detail config tree to a file.

observers_reset()

Reset all observers in MRFI, usually executed before an epoch of observation.

observers_result()

A convient function to get all observer results in MRFI.

load_package_functions(name)

Load a package of functions.

Parameters:

Name	Type	Description	Default
name	str	module name.	required

add_function(name, function_object)

Add a custom function, may be a observer, quantization, selector or error_mode.

Parameters:

Name	Type	Description	Default
name	str	To be used in config file, method field of observers, quantization, selector, error_mode.	required
function_object	Callable[..., Any]	A callable object consists with implementions of observers, quantization, selector, error_mode.	required

Examples:

>>> mrfi.add_function('my_empty_selector', lambda shape: [])
Then you can use this empty selector by setting selector.method 

to 'my_empty_selector' in .yaml config file.

find_modules(model, attrdict)

Find by module_type, module_name, module_fullname, also remove these attr in attrdict.