samna.dynapcnnB.configuration

Submodules

Module content

class samna.dynapcnnB.configuration.CnnLayerConfig

Bases: pybind11_object

Configuration options for CNN layers.

property biases

Contains bias values for features in a layer.

Formatted as [feature].

Type:

tensor array

property destinations

List of 2 CnnLayerDestination which describe target destination layers for events produced in this layer.

Type:

Array of 2 fixed length

property dimensions

Specifies padding, stride, kernel size and I/O shape of the layer.

Type:

Instance of CnnLayerDimensions

from_json(self: samna.dynapcnnB.configuration.CnnLayerConfig, arg0: str) → None

property input_congestion_balancer_enable

Enable input congestion balancer.

Type:

bool

property leak_enable

Enable leak/bias processing for this layer. A slow clk must be provided to SLOW_CLK pin.

CAUTION: May impact performance significantly.

Type:

bool

property leak_internal_slow_clk_enable

forward internally generated slow clock to convolutional layers’ leak generator to replace the need of external slow clk. The leakEnable must be set to True first

Type:

bool

property monitor_enable

Enable monitoring of this layer. When True all events produced by this layer will also be output on the monitoring bus. When False, events will only go to layers selected in the destinations list.

Monitoring should only be enabled on development kits that support fast readout of events.

Type:

bool

property neurons_initial_value

Contains initial neuron membrane potential values for neurons in a layer.

Formatted as [feature][x][y].

Type:

tensor array

property output_decimator_enable

Enable output decimator.

Type:

bool

property output_decimator_interval

Output decimator block interval.

Type:

int

property return_to_zero

If True neurons in this layer will return to 0 when firing. Otherwise their threshold will be subtracted from the value their membrane potential reached before firing.

Type:

bool

property threshold_high

Set the high threshold. When the neuron membrane potential becomes equal to or great than this value the neuron will fire.

Type:

int

property threshold_low

The minimum value of the neuron membrane potential which cannot be surpassed by leak.

Type:

int

to_json(self: samna.dynapcnnB.configuration.CnnLayerConfig) → str

property weights

4 dimensional list of lists representing kernel parameters.

Formatted as [feature][channel][x][y].

Type:

tensor array

class samna.dynapcnnB.configuration.CnnLayerDestination

Bases: pybind11_object

Configuration options for CNN layer destinations.

property enable

Enable this destination. If False no events will be forwarded, if True all events will be forwarded to this destination.

Type:

bool

property feature_shift

Shift feature IDs. The feature ID observed in the target layer will be the ID in this layer + feature_shift.

Type:

int

from_json(self: samna.dynapcnnB.configuration.CnnLayerDestination, arg0: str) → None

property layer

Target layer ID. Allowed values [0, 4] and 6, where [0,4] is the CNN layers and layer 6 is the readout layer. If you want to send Spike events to the readout block from outside, you need to set the target layer of the Spike event to 5.

Type:

int

property pooling

{1, 2, 4, 8}

Type:

int

Type:

Pooling will be applied to events before reaching the selected destination layer. Allowed values

to_json(self: samna.dynapcnnB.configuration.CnnLayerDestination) → str

class samna.dynapcnnB.configuration.CnnLayerDimensions

Bases: pybind11_object

Describes the dimensions of a Cnn layer as the configurable parameters.

from_json(self: samna.dynapcnnB.configuration.CnnLayerDimensions, arg0: str) → None

get_biases_array_dimensions(self: samna.dynapcnnB.configuration.CnnLayerDimensions) → List[int[1]]

Function to get the dimension of the bias array.

Returns:

A list of the dimension in the format of [feature].

get_neuron_states_array_dimensions(self: samna.dynapcnnB.configuration.CnnLayerDimensions) → List[int[3]]

Function to get the dimension of the neuron states array.

Returns:

A list of the dimension in the format of [feature, x, y].

get_weights_array_dimensions(self: samna.dynapcnnB.configuration.CnnLayerDimensions) → List[int[4]]

Function to get the dimension of the weights array.

Returns:

A list of the dimension in the format of [feature, channel, x, y].

property input_shape

Shape of features on the input of this layer.

Type:

Instance of IOShape

property kernel_size

Kernel size for all features in this layer. All kernels are square. Allowable values [1, 16]

Type:

int

property output_shape

Shape of features on the output of this layer.

Type:

Instance of IOShape

property padding

Padding of all features in this layer. Allowable values [0, 7].

Type:

Instance of Vec2

property stride

Stride for all features in this layer. Allowable values {1, 2, 4, 8}.

Type:

Instance of Vec2

to_json(self: samna.dynapcnnB.configuration.CnnLayerDimensions) → str

class samna.dynapcnnB.configuration.DvsFilterConfig

Bases: pybind11_object

The DVS Filter Block is included in the DVS pre-processing layer of Speck 2e in order to filter events depending on neighboring pixel activity. In general, an event at position (x,y) is forwarded by the filter when at least one pixel in the vicinity of (x,y) has spiked in a defined time window before the (x,y) event.

property enable

Enable the filter. A slow clk must be provided to SLOW_CLK pin.

Type:

bool

property filter_size

Filter window size in the X/Y dimension.

Type:

Instance of Vec2

from_json(self: samna.dynapcnnB.configuration.DvsFilterConfig, arg0: str) → None

property hot_pixel_filter_enable

Enable hot pixel filtering, default is enabled. The enable has to be set True first.

Type:

bool

property internal_slow_clk

Forward internally generated slow clock to DVS filter to replace the need of external slow clk. The enable must be set to True first.

Type:

bool

property threashold

Defines the threshold of the filter. This value is used to compare the current counter value with the value of the pixels neighboring the current spiking pixel. If any of the neighboring pixels has a value difference withing the delta value, the filter emits the current spike. Otherwise, the current spike is blocked.

Type:

int

to_json(self: samna.dynapcnnB.configuration.DvsFilterConfig) → str

class samna.dynapcnnB.configuration.DvsLayerConfig

Bases: pybind11_object

Configuration options for event pre-process layer. The event pre-processing layer receives as input the events coming from the internal DVS, or from an external source through the Input Serial Interface

property cut

Select the cut range, the roi cut happens concurrently, the range needs to be defined from the sensors (original) origin).

Actually it cuts an area using the given size based on (0, 0), then adjust this area according to origin.

Type:

Instance of Vec2

property destinations

Destinations

Type:

list of 2 DvsLayerDestination

from_json(self: samna.dynapcnnB.configuration.DvsLayerConfig, arg0: str) → None

property merge

Whether merge the two polarity from dvs (On/Off).

If True, Merge all events marking them with Polarity = ‘1’. If False, Pass all events and maintain the polarity information.

Type:

bool

property mirror

Enable mirroring (1) or not (0) of the X or Y dimension. mirroring is reversing image on X or Y dimension.

Type:

object(x bool, y bool)

property mirror_diagonal

Enables mirroring of the input events along the diagonal. when set to ‘1’, it swaps X with Y.

Type:

bool

property monitor_enable

Enables the DVS Pre-Processing Layer Monitor Bus. When enabled, output messages of the dvs pre-processing block are forwarded on the monitor bus to the output serial interface.

Type:

bool

property off_channel

Whether pass events with OFF polarity, only make sense when merge equal to True.

If True, pass events with OFF polarity. If False, block events with OFF polarity

Type:

bool

property on_channel

Whether pass events with ON polarity, only make sense when merge equal to True.

If True, pass events with ON polarity. If False, block events with ON polarity

Type:

bool

property origin

Select the origin offset the roi cut happens concurrently, the offset needs to be defined from the sensors (original) origin).

Actually this origin will redefined the area after cut.

Type:

Instance of Vec2

property pooling

Pooling is restricting dvs events in an area, available values are {1, 2, 4}, you can set pooling value on X/Y dimension seperately.

For example, if you set config.dvs_layer.pooling.x = 2 and config.dvs_layer.pooling.y = 2, you will see a 1/4 area on top left of visualizer which show pixels and other areas will be blank

Type:

Instance of Vec2

to_json(self: samna.dynapcnnB.configuration.DvsLayerConfig) → str

class samna.dynapcnnB.configuration.DvsLayerDestination

Bases: pybind11_object

Configuration options for dvs layer destination. It can have 2 destinations.

property enable

if this destination is enabled.

Type:

bool

from_json(self: samna.dynapcnnB.configuration.DvsLayerDestination, arg0: str) → None

property layer

destination layer that is going to receive output data.

Type:

int

to_json(self: samna.dynapcnnB.configuration.DvsLayerDestination) → str

class samna.dynapcnnB.configuration.DynapcnnConfiguration

Bases: pybind11_object

Contains configuration options and network parameters for a DynapCNN-B device. Applied to a device via apply_configuration() on the device model.

The configuration is divided into settings for the 9 CNN layers, the DVS event pre-processing layer and the readout layer. The factory configuration and debug configuration sub structs are pre-populated with default values and are only used for debug scenarios.

property cnn_layers

List of 5 CnnLayerConfig.

property debug_config

Instace of DebugConfig

property dvs_filter

Instance of DvsFilterConfig

property dvs_layer

Instance of DvsLayerConfig

property factory_config

Instance of FactoryConfig.

from_json(self: samna.dynapcnnB.configuration.DynapcnnConfiguration, arg0: str) → None

property readout

Instance of ReadoutConfig.

to_json(self: samna.dynapcnnB.configuration.DynapcnnConfiguration) → str

class samna.dynapcnnB.configuration.IOShape

Bases: pybind11_object

Describes the input or output shape of a CNN layer.

property feature_count

Number of features in the input or output”

Type:

int

from_json(self: samna.dynapcnnB.configuration.IOShape, arg0: str) → None

property size

Size of each feature.

Type:

Instance of Vec2

to_json(self: samna.dynapcnnB.configuration.IOShape) → str

class samna.dynapcnnB.configuration.ReadoutConfig

Bases: pybind11_object

The main use of the post-processing block is to calculate the moving average over a time window for a maximum of 15 neurons, provide the maximum average of the 15 neurons and compare the value of the calculated moving averages against a specified threshold.

property enable

enable the readout block, a slow clk (frequency depends on the algorithm model) must be provided to SLOW_CLK pin.

Type:

bool

from_json(self: samna.dynapcnnB.configuration.ReadoutConfig, arg0: str) → None

property internal_slow_clk

forward internally generated slow clk to the readout block to replace the need of external slow clk. The enable must be set to True first.

Type:

bool

property low_pass_filter32_not16

Selects readout moving average length (16 or 32 clock periods).

Value	Configuration
False	time window = 16* MOVING_AVERAGE_CLK period
True	time window = 32* MOVING_AVERAGE_CLK period

Type:

bool

property low_pass_filter_disable

Disables readout low pass filtering.

Value	Configuration
False	time window = selected by low_pass_filter32_not16
True	time window = 1 MOVING_AVERAGE_CLK period

This overrides low_pass_filter32_not16 when set to True.

Type:

bool

property monitor_enable

Allow monitoring of readout events, you will receive ReadoutValue from the model.

Please note that if you turn this on, you can’t get any events from CNN layers or the DVS sensor anymore.

Type:

bool

property output_mod_sel

There are 4 different modes where the output data are multiplexed. In the following table the different data output modes are described:

output_mode_sel	data_out[20]	data_out[19:16]	data_out[15:0]
0b00	data valid	neuron index of max average	power down (clock gating)
0b01	data valid	neuron index of max average	threshold compare output
0b10	data valid	neuron index of max average	average output of the selected neuron
0b11	max spiking neuron above threshold	neuron index of max average	average output of max spiking neuron

If output_mode_sel equals to ‘0b00’ then the clk_gated signal and the data_out[15:0] are equal to 0.

If output_mode_sel equals to ‘0b01’ then the data_out[15:0] consists of the data of the threshold comparison, the index of the maximum moving average neuron and the data valid signal. The threshold comparison data is the 16 bit value of the comparison of each neurons’ moving average with the threshold.

If output_mode_sel equals to ‘0b10’ then the data_out[15:0] consists of the moving average of the selected neuron, the index of the maximum moving average neuron and the data valid signal. The data valid signal is asserted after all the computations have finished in order to ensure correct sampling of the data.

If output_mode_sel equals to ‘0b11’ then the data_out[15:0] consists of the maximum average of the 15 neurons, the index of the maximum moving average neuron and the comparison output between the maximum moving average and the threshold.

Type:

int

property output_neuron_num

The internal signal output_mode_sel enable the selection of one of the 15 output neurons. The neuron selection is used when signals output_mode_sel == ‘0b10’.

Selected neuron number = average_sel.

Type:

int

property override_threshold_max

Override Output Layer Threshold compare. Sends interrupt at each slow clock cycle.

Type:

bool

property readout_configuration_sel

Addressing modes of the incoming spikes.

There are 4 different addressing modes that could be selected. The valid configurations are the following:

Value	Configuration
0b00	2x*2y*4f
0b10	4x*4y*1f
0b01	2x*4y*2f
0b11	1x*1y*16f

Where x, y and f represent, respectively, the X, Y and Feature dimensions of the events arriving at the Readout Layer.

Type:

int

property readout_pin_monitor_enable

enable/disable the monitor of the readout pins. If enabled, the ReadoutPinValue will be outputted. Default is False.

Type:

bool

property threshold

Readout neurons threshold value.

The moving average of the output neurons is compared to the threshold value to produce an output if the received number of spikes is greater than the threshold.

Type:

int

to_json(self: samna.dynapcnnB.configuration.ReadoutConfig) → str