ml_genn_tf.converters package

mlGeNN TF converters take an ANN trained in TensorFlow and convert it to an mlGeNN Network for SNN inference.

class ml_genn_tf.converters.Converter

Bases: object

Base class for all converters

convert(tf_model)

Convert a TensorFlow model to an mlGeNN network. Returns network, list of input populations, list of output populations and dictionary mapping TF layers to populations.

Parameters:

tf_model (tensorflow.keras.Model) – TensorFlow model to be converted

Return type:

Tuple[Network, List[Population], List[Population], Dict[tensorflow.keras.layers.Layer, Population]]

create_compiler(**kwargs)

Create suitable compiler to compile networks produced by this converter

Return type:

Compiler

create_input_neurons(pre_convert_output)

Create converter-specific input neuron model

Parameters:

pre_convert_output – Compiler-specific state created by pre_convert().

Return type:

Neuron

create_neurons(tf_layer, pre_convert_output, is_output)

Create converter-specific neuron model from TF layer

Parameters:

  • tf_layer (tensorflow.keras.layers.Layer) – TF layer to convert

  • pre_convert_output – Compiler-specific state created by pre_convert().

  • is_output (bool) – Is this an output layer?

Return type:

Neuron

post_convert(mlg_network, mlg_network_inputs, mlg_model_outputs)

If any post-processing is required to the network after adding all layers, converters should implement it here.

Parameters:

  • mlg_network (Network) – Populated network

  • mlg_network_inputs (List[Population]) – List of input populations

  • mlg_model_outputs (List[Population]) – List of output populations

pre_convert(tf_model)

If any pre-processing is required before converting TF model, converters should implement it here. Any converter-specific state that should be persistent across conversion should be encapsulated in an object returned from this method.

Parameters:

tf_model (tensorflow.keras.Model) – TensorFlow model to be converted

validate_tf_layer(tf_layer, config)
Parameters:

tf_layer (tensorflow.keras.layers.Layer)

class ml_genn_tf.converters.DataNorm(evaluate_timesteps, signed_input=False, norm_data=None, input_type=InputType.POISSON)

Bases: Converter

Converts ANNs to network of integrate-and-fire neurons, operating in a rate-based regime If normalisation data is provided, thresholds are balancing using the algorithm proposed by [Diehl2015].

Parameters:

  • evaluate_timesteps (int) – ss

  • signed_input – ss

  • norm_data – paa

  • input_type – sss

create_compiler(**kwargs)

Create suitable compiler to compile networks produced by this converter

create_input_neurons(pre_convert_output)

Create converter-specific input neuron model

Parameters:

pre_convert_output (PreConvertOutput) – Compiler-specific state created by pre_convert().

Return type:

Neuron

create_neurons(tf_layer, pre_convert_output, is_output)

Create converter-specific neuron model from TF layer

Parameters:

  • tf_layer (tensorflow.keras.layers.Layer) – TF layer to convert

  • pre_convert_output (PreConvertOutput) – Compiler-specific state created by pre_convert().

  • is_output (bool) – Is this an output layer?

Return type:

Neuron

pre_convert(tf_model)

If any pre-processing is required before converting TF model, converters should implement it here. Any converter-specific state that should be persistent across conversion should be encapsulated in an object returned from this method.

Parameters:

tf_model – TensorFlow model to be converted

class ml_genn_tf.converters.FewSpike(k=10, alpha=25, signed_input=False, norm_data=None)

Bases: Converter

Parameters:

  • k (int)

  • alpha (float)

create_compiler(**kwargs)

Create suitable compiler to compile networks produced by this converter

create_input_neurons(pre_convert_output)

Create converter-specific input neuron model

Parameters:

pre_convert_output – Compiler-specific state created by pre_convert().

create_neurons(tf_layer, pre_convert_output, is_output)

Create converter-specific neuron model from TF layer

Parameters:

  • tf_layer – TF layer to convert

  • pre_convert_output – Compiler-specific state created by pre_convert().

  • is_output – Is this an output layer?

post_convert(mlg_network, mlg_network_inputs, mlg_model_outputs)

If any post-processing is required to the network after adding all layers, converters should implement it here.

Parameters:

  • mlg_network – Populated network

  • mlg_network_inputs – List of input populations

  • mlg_model_outputs – List of output populations

pre_convert(tf_model)

If any pre-processing is required before converting TF model, converters should implement it here. Any converter-specific state that should be persistent across conversion should be encapsulated in an object returned from this method.

Parameters:

tf_model – TensorFlow model to be converted

class ml_genn_tf.converters.InputType(value)

Bases: Enum

An enumeration.

IF = 'if'
POISSON = 'poisson'
SPIKE = 'spike'
class ml_genn_tf.converters.Simple(evaluate_timesteps, signed_input=False, input_type=InputType.POISSON)

Bases: Converter

Parameters:
create_compiler(**kwargs)

Create suitable compiler to compile networks produced by this converter

create_input_neurons(pre_compile_output)

Create converter-specific input neuron model

Parameters:

pre_convert_output – Compiler-specific state created by pre_convert().

create_neurons(tf_layer, pre_compile_output, is_output)

Create converter-specific neuron model from TF layer

Parameters:

  • tf_layer – TF layer to convert

  • pre_convert_output – Compiler-specific state created by pre_convert().

  • is_output – Is this an output layer?

pre_compile(mlg_network)
pre_convert(tf_model)

If any pre-processing is required before converting TF model, converters should implement it here. Any converter-specific state that should be persistent across conversion should be encapsulated in an object returned from this method.

Parameters:

tf_model – TensorFlow model to be converted

Submodules

ml_genn_tf.converters.converter module

class ml_genn_tf.converters.converter.Converter

Bases: object

Base class for all converters

convert(tf_model)

Convert a TensorFlow model to an mlGeNN network. Returns network, list of input populations, list of output populations and dictionary mapping TF layers to populations.

Parameters:

tf_model (tensorflow.keras.Model) – TensorFlow model to be converted

Return type:

Tuple[Network, List[Population], List[Population], Dict[tensorflow.keras.layers.Layer, Population]]

create_compiler(**kwargs)

Create suitable compiler to compile networks produced by this converter

Return type:

Compiler

create_input_neurons(pre_convert_output)

Create converter-specific input neuron model

Parameters:

pre_convert_output – Compiler-specific state created by pre_convert().

Return type:

Neuron

create_neurons(tf_layer, pre_convert_output, is_output)

Create converter-specific neuron model from TF layer

Parameters:

  • tf_layer (tensorflow.keras.layers.Layer) – TF layer to convert

  • pre_convert_output – Compiler-specific state created by pre_convert().

  • is_output (bool) – Is this an output layer?

Return type:

Neuron

post_convert(mlg_network, mlg_network_inputs, mlg_model_outputs)

If any post-processing is required to the network after adding all layers, converters should implement it here.

Parameters:

  • mlg_network (Network) – Populated network

  • mlg_network_inputs (List[Population]) – List of input populations

  • mlg_model_outputs (List[Population]) – List of output populations

pre_convert(tf_model)

If any pre-processing is required before converting TF model, converters should implement it here. Any converter-specific state that should be persistent across conversion should be encapsulated in an object returned from this method.

Parameters:

tf_model (tensorflow.keras.Model) – TensorFlow model to be converted

validate_tf_layer(tf_layer, config)
Parameters:

tf_layer (tensorflow.keras.layers.Layer)

ml_genn_tf.converters.data_norm module

class ml_genn_tf.converters.data_norm.DataNorm(evaluate_timesteps, signed_input=False, norm_data=None, input_type=InputType.POISSON)

Bases: Converter

Converts ANNs to network of integrate-and-fire neurons, operating in a rate-based regime If normalisation data is provided, thresholds are balancing using the algorithm proposed by [Diehl2015].

Parameters:

  • evaluate_timesteps (int) – ss

  • signed_input – ss

  • norm_data – paa

  • input_type – sss

create_compiler(**kwargs)

Create suitable compiler to compile networks produced by this converter

create_input_neurons(pre_convert_output)

Create converter-specific input neuron model

Parameters:

pre_convert_output (PreConvertOutput) – Compiler-specific state created by pre_convert().

Return type:

Neuron

create_neurons(tf_layer, pre_convert_output, is_output)

Create converter-specific neuron model from TF layer

Parameters:

  • tf_layer (tensorflow.keras.layers.Layer) – TF layer to convert

  • pre_convert_output (PreConvertOutput) – Compiler-specific state created by pre_convert().

  • is_output (bool) – Is this an output layer?

Return type:

Neuron

pre_convert(tf_model)

If any pre-processing is required before converting TF model, converters should implement it here. Any converter-specific state that should be persistent across conversion should be encapsulated in an object returned from this method.

Parameters:

tf_model – TensorFlow model to be converted

class ml_genn_tf.converters.data_norm.PreConvertOutput(thresholds)

Bases: tuple

Create new instance of PreConvertOutput(thresholds,)

thresholds

Alias for field number 0

ml_genn_tf.converters.enum module

class ml_genn_tf.converters.enum.InputType(value)

Bases: Enum

An enumeration.

IF = 'if'
POISSON = 'poisson'
SPIKE = 'spike'

ml_genn_tf.converters.few_spike module

class ml_genn_tf.converters.few_spike.FewSpike(k=10, alpha=25, signed_input=False, norm_data=None)

Bases: Converter

Parameters:

  • k (int)

  • alpha (float)

create_compiler(**kwargs)

Create suitable compiler to compile networks produced by this converter

create_input_neurons(pre_convert_output)

Create converter-specific input neuron model

Parameters:

pre_convert_output – Compiler-specific state created by pre_convert().

create_neurons(tf_layer, pre_convert_output, is_output)

Create converter-specific neuron model from TF layer

Parameters:

  • tf_layer – TF layer to convert

  • pre_convert_output – Compiler-specific state created by pre_convert().

  • is_output – Is this an output layer?

post_convert(mlg_network, mlg_network_inputs, mlg_model_outputs)

If any post-processing is required to the network after adding all layers, converters should implement it here.

Parameters:

  • mlg_network – Populated network

  • mlg_network_inputs – List of input populations

  • mlg_model_outputs – List of output populations

pre_convert(tf_model)

If any pre-processing is required before converting TF model, converters should implement it here. Any converter-specific state that should be persistent across conversion should be encapsulated in an object returned from this method.

Parameters:

tf_model – TensorFlow model to be converted

class ml_genn_tf.converters.few_spike.PreConvertOutput(layer_alpha, input_alpha)

Bases: tuple

Create new instance of PreConvertOutput(layer_alpha, input_alpha)

input_alpha

Alias for field number 1

layer_alpha

Alias for field number 0

ml_genn_tf.converters.simple module

class ml_genn_tf.converters.simple.Simple(evaluate_timesteps, signed_input=False, input_type=InputType.POISSON)

Bases: Converter

Parameters:
create_compiler(**kwargs)

Create suitable compiler to compile networks produced by this converter

create_input_neurons(pre_compile_output)

Create converter-specific input neuron model

Parameters:

pre_convert_output – Compiler-specific state created by pre_convert().

create_neurons(tf_layer, pre_compile_output, is_output)

Create converter-specific neuron model from TF layer

Parameters:

  • tf_layer – TF layer to convert

  • pre_convert_output – Compiler-specific state created by pre_convert().

  • is_output – Is this an output layer?

pre_compile(mlg_network)
pre_convert(tf_model)

If any pre-processing is required before converting TF model, converters should implement it here. Any converter-specific state that should be persistent across conversion should be encapsulated in an object returned from this method.

Parameters:

tf_model – TensorFlow model to be converted

ml_genn_tf.converters.spike_norm module

class ml_genn_tf.converters.spike_norm.NormCompiler(evaluate_timesteps, dt=1.0, batch_size=1, rng_seed=0, kernel_profiling=False, prefer_in_memory_connect=True, reset_time_between_batches=True, reset_vars_between_batches=True, reset_in_syn_between_batches=False, communicator=None, **genn_kwargs)

Bases: InferenceCompiler

Parameters:

  • evaluate_timesteps (int)

  • dt (float)

  • batch_size (int)

  • rng_seed (int)

  • kernel_profiling (bool)

  • communicator (Communicator)

build_neuron_model(pop, model, compile_state)

Apply compiler-specific processing to the base neuron model returned by ml_genn.neurons.Neuron.get_model(). If modifications are made, this should be done to a (deep) copy.

Parameters:

  • pop – Population neuron model is associated with

  • model – Base neuron model

  • compile_state – Compiler-specific state created by pre_compile().

create_compiled_network(genn_model, neuron_populations, connection_populations, compile_state)

Perform any final compiler-specific modifications to compiled GeNNModel and return ml_genn.compilers.CompiledNetwork derived object.

Parameters:

  • genn_model – GeNNModel with all neuron and synapse groups added

  • neuron_populations – dictionary mapping ml_genn.Population objects to GeNN NeuronGroup objects they have been compiled into

  • connection_populations – dictionary mapping ml_genn.Connection objects to GeNN SynapseGroup objects they have been compiled into

  • compile_state – Compiler-specific state created by pre_compile().

ml_genn_tf.converters.spike_norm.spike_normalise(net, net_inputs, net_outputs, norm_data, evaluate_timesteps, num_batches=None, dt=1.0, batch_size=1, rng_seed=0, kernel_profiling=False, prefer_in_memory_connect=True, reset_time_between_batches=True, **genn_kwargs)
Parameters:

  • evaluate_timesteps (int)

  • num_batches (int | None)

  • dt (float)

  • batch_size (int)

  • rng_seed (int)

  • kernel_profiling (bool)