converter.CRI_Converter

converter.CRI_Converter(self, num_steps, input_layer, output_layer, input_shape, v_threshold, embed_dim, backend='spikingjelly')

A class to convert a neural network model into an equivalent model compatible with the CRI (Capacitive ReRAM Inverter) hardware.

Parameters

Name	Type	Description	Default
`num_steps`	int	The number of time steps in the input.	required
`input_layer`	int	The index of the first pytorch layer used as synapses.	required
`output_layer`	int	The index of the last pytorch layer used as synapses.	required
`input_shape`	tuple of int	The shape of the input data. Default is (1, 28, 28).	required
`backend`	str	The backend to use. Currently Support SpikingJelly and snnTorch. Default is ‘spikingjelly’.	`'spikingjelly'`
`v_threshold`	float	The voltage threshold for the neurons. It should be set to the v_threshold of Quantize Network.	required
`embed_dim`	int	The embedding dimension. Only used for spikeformer.	required

Name	Type	Description
HIGH_SYNAPSE_WEIGHT	float	The high synapse weight value. Default is 1e6.
axon_dict	defaultdict of list	A dictionary mapping each axon to a list of connected neurons.
neuron_dict	defaultdict of list	A dictionary mapping each neuron to a list of connected axons.
output_neurons	list	A list of output neurons.
input_shape	numpy.ndarray	The shape of the input data.
num_steps	int	The number of time steps in the input.
axon_offset	int	The current offset for axon indexing.
neuron_offset	int	The current offset for neuron indexing.
backend	str	The backend to use.
save_input	bool	Whether to save the input data.
bias_start_idx	int or None	The starting index for bias neurons.
curr_input	np.ndarray or None	The current input data.
input_layer	int	The index of the input layer.
output_layer	int	The index of the output layer.
layer_index	int	The current layer index.
total_axonSyn	int	The total number of axon synapses.
total_neuronSyn	int	The total number of neuron synapses.
max_fan	int	The maximum fan-out.
v_threshold	float	The voltage threshold for the neurons.
q	np.ndarray or None	The q matrix for attention conversion.
v	np.ndarray or None	The v matrix for attention conversion.
k	np.ndarray or None	The k matrix for attention conversion.
embed_dim	int	The embedding dimension.

>>> converter = CRI_Converter()
>>> converter.layer_converter(some_model)
>>> converter.input_converter(some_input_data)

Name	Description
input_converter	Converts input data into a spike train and then into a list of axon indices.
layer_converter	Converts a model into a CRI-compatible model.
run_CRI_hw	Runs a set of inputs through the hardware implementation of the network and gets the output predictions.
run_CRI_sw	Runs a set of inputs through the software simulation of the network and gets the output predictions.

converter.CRI_Converter.input_converter(self, input_data)

Converts input data into a spike train and then into a list of axon indices.

Name	Type	Description	Default
`input_data`	torch.Tensor	The input data.	required

Type	Description
list of list of str	The batch of spikes, with each spike represented by its axon index.

>>> converter = CRI_Converter()
>>> converter.input_converter(some_input_data)

converter.CRI_Converter.layer_converter(self, model)

Converts a model into a CRI-compatible model.

Name	Type	Description	Default
`model`	torch.nn.Module	The input model.	required

>>> converter = CRI_Converter()
>>> converter.layer_converter(some_model)

converter.CRI_Converter.run_CRI_hw(self, inputList, hardwareNetwork)

Runs a set of inputs through the hardware implementation of the network and gets the output predictions.

Name	Type	Description	Default
`inputList`	list of list of str	The input data, where each item is a list of axon indices representing the spikes.	required
`hardwareNetwork`	object	The hardware network object.	required

Type	Description
list of int	The output predictions.

>>> converter = CRI_Converter()
>>> converter.run_CRI_hw(some_inputList, some_hardwareNetwork)

converter.CRI_Converter.run_CRI_sw(self, inputList, softwareNetwork)

Runs a set of inputs through the software simulation of the network and gets the output predictions.

Name	Type	Description	Default
`inputList`	list of list of str	The input data, where each item is a list of axon indices representing the spikes.	required
`softwareNetwork`	object	The software network object.	required

Type	Description
list of int	The output predictions.

>>> converter = CRI_Converter()
>>> converter.run_CRI_sw(some_inputList, some_softwareNetwork)