experiment

logspace_density(low=-8, high=-4, density=3)

Similar as np.logspace but the datapoint always contains $10^{-n}$

Parameters:

Name	Type	Description	Default
low	int	begin point at 10**low	-8
high	int	end point at 10**high	-4
density	int	how many data point per 10x range.	3

Returns:

Type	Description
np.array	Array of uniformed logspace data point.

Acc_experiment(model, dataloader)

Return classification accuracy on dataset.

Parameters:

Name	Type	Description	Default
model	Union[MRFI, torch.nn.Module]	Target model.	required
dataloader	DataLoader	Yields a series of tuple of (batched) input images and classification label.	required

Returns:

Type	Description
float	Accuracy result among all data, between 0 and 1.

observeFI_experiment(fi_model, input_data, use_golden=True, observers_dict=None)

Run fault inject experiment and return internal oberver results.

Parameters:

Name	Type	Description	Default
fi_model	MRFI	Target model.	required
input_data	Union[torch.Tensor, DataLoader]	A input tensor or a dataloader.	required
use_golden	bool	If True, run both golden run and fault injection per batch, most of fault injection observer requires golden run before fault inject.	True
observers_dict	Optional[dict]	A dict of observer config handlers to get result. if None, return all observer result.	None

Returns:

Type	Description
dict	Dictionary of observer result, keys are observer name and values are result of observer.result().

BER_Acc_experiment(fi_model, fi_selectors, dataloader, BER_range=logspace_density(), bit_width=16)

Conduct a series of experimet under different bit error rate and get accuracy.

Parameters:

Name	Type	Description	Default
fi_model	MRFI	Target model.	required
fi_selectors	ConfigItemList	A target selector list from fi_model.get_configs().	required
dataloader	DataLoader	Yields a series of tuple of (batched) input images and classification label.	required
BER_range	Union[list, np.array]	A list or a array for target bit error rate.	logspace_density()
bit_width	int	To calculate bit error rate by tensor value selected rate.	16

Returns:

Type	Description
tuple	Current bit error rate list and accuracy result among all data of each bit error list.

BER_observe_experiment(fi_model, fi_selectors, dataloader, BER_range=logspace_density(), bit_width=16, use_golden=True)

Conduct a series of experimet under different bit error rate and get all observers result.

Parameters:

Name	Type	Description	Default
fi_model	MRFI	Target model.	required
fi_selectors	ConfigItemList	A target selector list from fi_model.get_configs().	required
dataloader	DataLoader	Yields a series of tuple of (batched) input images and classification label.	required
BER_range	Union[list, np.array]	A list or a array for target bit error rate.	logspace_density()
bit_width	int	To calculate bit error rate by tensor value selected rate.	16
use_golden	bool	If True, run both golden run and fault injection per batch.	True

Returns:

Type	Description
tuple	Current bit error rate list and observation result dictionary list.

Acc_golden(fi_model, dataloader, disable_quantization=False)

Evaluate model accuracy without error injection.

Parameters:

Name	Type	Description	Default
fi_model	MRFI	Target model.	required
dataloader	torch.utils.data.DataLoader	Yields a series of tuple of (batched) input images and classification label.	required
disable_quantization	bool	If True, also disable quantization if it exists. Note that quantization also affects model accuracy.	False

Returns:

Type	Description
float	Golden run accuracy result among all data, between 0 and 1.

observeFI_experiment_plus(fi_model, input_data, method='RMSE', pre_hook=False, **kwargs)

Run fault injection experiment and observe its internal effect.

Compare with observeFI_experiment(), this function add temporary target observers of to model.

Parameters:

Name	Type	Description	Default
fi_model	MRFI	Target model.	required
input_data	Union[torch.Tensor, DataLoader]	A input tensor or a dataloader.	required
method	str	Name of the fault injection observer in mrfi.observer.	'RMSE'
pre_hook	bool	Observer value in module input, not output.	False
**kwargs	dict	Target module to do observe, module_type, module_name or module_fullname	{}

Returns:

Type	Description
dict	Dictionary of observer result, keys are observer name and values are result of observer.result().

get_activation_info(fi_model, input_data, method='MinMax', pre_hook=False, **kwargs)

Observe model activations without fault injection.

Parameters:

Name	Type	Description	Default
fi_model	MRFI	Target model.	required
input_data	Union[torch.Tensor, DataLoader]	A input tensor or a dataloader.	required
method	str	Name of the basic observer in mrfi.observer.	'MinMax'
pre_hook	bool	Observer value in module input, not output.	False
**kwargs	dict	Target module to observe, module_type, module_name or module_fullname	{}

Returns:

Type	Description
dict	Dictionary of observer result, keys are activation observation name and values are result of observer.result().

get_weight_info(model, method='MinMax', weight_name='weight', **kwargs)

Observe model weights.

Parameters:

Name	Type	Description	Default
model	Union[MRFI, torch.nn.Module]	Target model.	required
method	str	Name of the basic observer in mrfi.observer.	'MinMax'
weight_name	Union[str, List[str]]	Target weight name.	'weight'
**kwargs	dict	Target module to observe, module_type, module_name or module_fullname	{}

Returns:

Type	Description
dict	Dictionary of observer result, keys are weight observation and values are result of observer.result().