MRFI function table

experiment

Name	Description	Args
logspace_density	Similar as np.logspace but the datapoint always contains $10^{-n}$	low,high,density,return
Acc_experiment	Return classification accuracy on dataset.	model,dataloader,return
observeFI_experiment	Run fault inject experiment and return internal oberver results.	fi_model,input_data,use_golden,observers_dict,return
BER_Acc_experiment	Conduct a series of experimet under different bit error rate and get accuracy.	fi_model,fi_selectors,dataloader,BER_range,bit_width,return
BER_observe_experiment	Conduct a series of experimet under different bit error rate and get all observers result.	fi_model,fi_selectors,dataloader,BER_range,bit_width,use_golden,return
Acc_golden	Evaluate model accuracy without error injection.	fi_model,dataloader,disable_quantization,return
observeFI_experiment_plus	Run fault injection experiment and observe its internal effect.	fi_model,input_data,method,pre_hook,kwargs,return
get_activation_info	Observe model activations without fault injection.	fi_model,input_data,method,pre_hook,kwargs,return
get_weight_info	Observe model weights.	model,method,weight_name,kwargs,return

observer

Name	Description	Args
BaseObserver	Basement of observers.
MinMax	Observe min/max range of tensors.
RMSE	Root Mean Square Error metric between golden run and fault inject run.
SaveLast	Simply save last inference internal tensor.
MaxAbs	Observe max abs range of tensors.
MeanAbs	Mean of abs, a metric of scale of values
Std	Standard deviation of zero-mean values.
Shape	Simply record tensor shape of last inference
MAE	Mean Absolute Error between golden run and fault inject run.
EqualRate	Compare how many value unchanged between golden run and fault inject run.
UniformSampling	Uniform sampling from tensors between all inference, up to 10000 samples.

selector

Name	Description	Args
EmptySelector	No position selected by this, for debug or special use.
FixPosition	Select fixed one position by coordinate.	position
FixPositions	Select a list of fixed positions by coordinate.	positions
RandomPositionByNumber	Select n random positions.	n,per_instance
RandomPositionByRate	Select random positions by rate.	rate,poisson
RandomPositionByRate_classic	Select random positions by rate.	rate
MaskedDimRandomPositionByNumber	Select n positions after specifed dimensions are masked.	n,kwargs
SelectedDimRandomPositionByNumber	Select n positions on selected coordinate.	n,kwargs
MaskedDimRandomPositionByRate	Select by rate where some coordinate are masked.	rate,poisson,kwargs
SelectedDimRandomPositionByRate	Select on some coordinate by rate.	rate,poisson,kwargs
FixedPixelByNumber	Select random channel on one fixed pixel (i.e. H * W dimension).	n,pixel,per_instance

error_mode

Name	Description	Args
IntSignBitFlip	Flip integer tensor on sign bit	x_in,bit_width
IntRandomBitFlip	Flip integer tensor on random bit	x_in,bit_width
IntFixedBitFlip	Flip integer tensor on specified bit	x_in,bit_width,bit
SetZero	Fault inject selected value to zero	x
SetValue	Fault inject value to specified value(s)	x,value
FloatRandomBitFlip	Flip bit randomly on float values	x_in,floattype
FloatFixedBitFlip	Flip specific bit on float values	x_in,bit,floattype
IntRandom	Uniformly set to random integer in range.	low,high
IntRandomBit	Set fault value to a random bit_width-bit integer.	bit_width,signed
UniformRandom	Uniform set to float number in range.	low,high
NormalRandom	Set fault injection value to a normal distribution.	mean,std
UniformDisturb	Add uniform distribution noise to current value.	low,high
NormalDisturb	Add normal distribution noise to current value.	std,bias

quantization

Name	Description	Args
SymmericQuantization	Simple symmeric quantization.	dynamic_range,scale_factor
PositiveQuantization	Simple positive quantization.	dynamic_range,scale_factor
FixPointQuantization	Fixpoint quantization.	decimal_bit