ml_genn.losses package

Loss functions are used to compute the quantity that a model should seek to minimize during training. For supervised learning tasks, these are based on some labels and the prediction obtained from a model using a ml_genn.readouts.Readout

class ml_genn.losses.Loss

Bases: ABC

Base class for all loss functions

abstract add_to_neuron(model, shape, batch_size, example_timesteps)

Modify a neuron model, adding any additional state and functionality required to implement this loss function.

Parameters:

• model (NeuronModel) – Neuron model to modify (in place)

• shape – Shape of population loss is calculated for

• batch_size (int) – Batch size of model used with loss function

• example_timesteps (int) – How many timestamps each example will be presented to the network for

abstract set_target(genn_pop, y_true, shape, batch_size, example_timesteps)

Write the current target output value to the compiled neuron group.

Parameters:

• genn_pop – GeNN NeuronGroup object population with loss function has been compiled into

• y_true – ‘true’ values provided to compiled network train method

• shape – Shape of population loss is calculated for

• batch_size (int) – Batch size of model used with loss function

• example_timesteps (int) – How many timestamps each example will be presented to the network for

class ml_genn.losses.MeanSquareError

Bases: Loss

Computes the mean squared error between labels and prediction

add_to_neuron(model, shape, batch_size, example_timesteps)

Modify a neuron model, adding any additional state and functionality required to implement this loss function.

Parameters:

• model (NeuronModel) – Neuron model to modify (in place)

• shape – Shape of population loss is calculated for

• batch_size (int) – Batch size of model used with loss function

• example_timesteps (int) – How many timestamps each example will be presented to the network for

set_target(genn_pop, y_true, shape, batch_size, example_timesteps)

Write the current target output value to the compiled neuron group.

Parameters:

• genn_pop – GeNN NeuronGroup object population with loss function has been compiled into

• y_true – ‘true’ values provided to compiled network train method

• shape – Shape of population loss is calculated for

• batch_size (int) – Batch size of model used with loss function

• example_timesteps (int) – How many timestamps each example will be presented to the network for

class ml_genn.losses.SparseCategoricalCrossentropy

Bases: Loss

Computes the crossentropy between labels and prediction when there are two or more label classes, specified as integers.

add_to_neuron(model, shape, batch_size, example_timesteps)

Modify a neuron model, adding any additional state and functionality required to implement this loss function.

Parameters:

• model (NeuronModel) – Neuron model to modify (in place)

• shape – Shape of population loss is calculated for

• batch_size (int) – Batch size of model used with loss function

• example_timesteps (int) – How many timestamps each example will be presented to the network for

set_target(genn_pop, y_true, shape, batch_size, example_timesteps)

Write the current target output value to the compiled neuron group.

Parameters:

• genn_pop – GeNN NeuronGroup object population with loss function has been compiled into

• y_true – ‘true’ values provided to compiled network train method

• shape – Shape of population loss is calculated for

• batch_size (int) – Batch size of model used with loss function

• example_timesteps (int) – How many timestamps each example will be presented to the network for