selector

MRFI selector methods

A selector is a call back function with fixed argument shape, and other optional args specified in config file.

The selector determines which positions of tensor are send to fault injector i.e. error_mode.

Selector returns a 1-d index tensor with type torch.long, indicates the index of target tensor after flatten.

Note

If no selector specified in config file, all target tensor will be sent to fault injector.

EmptySelector(shape)

No position selected by this, for debug or special use.

FixPosition(shape, position)

Select fixed one position by coordinate.

Parameters:

Name	Type	Description	Default
position	Union[int, List[int]]	if int, stands for index of target position on flattened tensor. if List[int], stands for n-d coordinate of target position.	required

FixPositions(shape, positions)

Select a list of fixed positions by coordinate.

Parameters:

Name	Type	Description	Default
positions	Union[List[int], List[List[int]]]	if List[int], stands for index of target tensor after flatten. if List[List[int]], stands for n-d coordinate of target position.	required

RandomPositionByNumber(shape, n=1, per_instance=False)

Select n random positions.

Parameters:

Name	Type	Description	Default
n	int	Number of target positions.	1
per_instance	bool	if True, perform n inject on per instance (i.e. ignore dim 0).	False

RandomPositionByRate(shape, rate=0.0001, poisson=True)

Select random positions by rate.

This function generate positions by n ~ Possion(N_all * rate).

Parameters:

Name	Type	Description	Default
rate	float	Rate of each position to be chosen.	0.0001
poisson	bool	Enable poisson samping, which is more accurate when rate is quite small.	True

Info

rate stands for tensor value selected rate here. To calculate the bit error rate on bit flip experiment, it needs to be additionally divided by the bit width.

RandomPositionByRate_classic(shape, rate=0.0001)

Select random positions by rate.

This function generate positions by sampling on all value.

Parameters:

Name	Type	Description	Default
rate	float	Rate of each position to be chosen.	0.0001

Deprecated

This function is deprecated and only for test because of low performance.

MaskedDimRandomPositionByNumber(shape, n=1, **kwargs)

Select n positions after specifed dimensions are masked.

Use a list to indicate which values on a certain dimension are not selected.

For a 2-D tensor, it is equivalent to selecting on a submatrix where some rows and cols are masked.

Tip

MaskedDim- and SelectedDim- selectors can be used for fine-grained evaluate and selective protect experiments, including instance-wise, channel-wise and spatial-wise.

Info

Mask argeuments follow common pytorch memory layout:

• CNN feature map: (instance, channel, height, width)

• CNN weight: (out_channel, in_channel, height, width)

• Full connect layer: (instance, neuron)

• Full connect weight: (out, in)

Parameters:

Name	Type	Description	Default
n	int	number of position to select.	1
**kwargs	dict	instance(dim = 0), type list[int] channel(dim = 1), type list[int] height(dim = 2), type list[int] width(dim = 3), type list[int] out_channel(dim = 0), type list[int] in_channel(dim = 1), type list[int] out(dim = 0), type list[int] in(dim = 1), type list[int] neuron(dim = 1), type list[int]	{}

SelectedDimRandomPositionByNumber(shape, n=1, **kwargs)

Select n positions on selected coordinate.

For argument list, please refer MaskedDimRandomPositionByNumber.

Note if one dimension selection list is not specified, it stands for all of this dimension are possible selected.

MaskedDimRandomPositionByRate(shape, rate, poisson=True, **kwargs)

Select by rate where some coordinate are masked.

For argument list, please refer MaskedDimRandomPositionByNumber.

SelectedDimRandomPositionByRate(shape, rate, poisson=True, **kwargs)

Select on some coordinate by rate.

For argument list, please refer MaskedDimRandomPositionByNumber.

Note if one dimension selection list is not specified, it stands for all of this dimension are possible selected.

FixedPixelByNumber(shape, n, pixel, per_instance=False)

Select n random channel on one fixed pixel.

Pixel is on H * W dimension.

Parameters:

Name	Type	Description	Default
n	int	number of positions	required
pixel	Union[int, tuple]	An 2-d coordinate tuple specified target pixel	required
per_instance	bool	if True, perform n inject on per instance (i.e. ignore dim 0)	False