experiments.equilibrium_propagation package

Submodules

experiments.equilibrium_propagation.models module

Adapted from https://github.com/Laborieux-Axel/Equilibrium-Propagation/blob/master/model_utils.py

class experiments.equilibrium_propagation.models.P_MLP(archi, activation=<built-in method tanh of type object>)

Bases: Module

Layered equilibrium-propagation multilayer perceptron.

Parameters:

• archi – Sequence of layer widths including input and output sizes.

• activation – Activation used for hidden-state updates.

Phi(x, y, neurons, beta, criterion)

Compute the primitive energy for the current network state.

Parameters:

• x – Input batch.

• y – Target labels.

• neurons – Current layer-state tensors.

• beta – Nudging coefficient for the loss term.

• criterion – Loss function used when beta is non-zero.

Returns:

Primitive value per sample.

compute_syn_grads(x, y, neurons_1, neurons_2, betas, criterion)

Compute synaptic gradients from two EP steady states.

Parameters:

• x – Input batch.

• y – Target labels.

• neurons_1 – First steady-state neurons.

• neurons_2 – Second steady-state neurons.

• betas – Pair of nudging coefficients.

• criterion – Loss function used by the primitive.

forward(x, y, neurons, T, beta=0.0, criterion=MSELoss())

Run fixed-point dynamics for T steps.

Parameters:

• x – Input batch.

• y – Target labels.

• neurons – Initial layer-state tensors.

• T – Number of dynamics steps.

• beta – Nudging coefficient.

• criterion – Loss function used by the primitive.

Returns:

Updated neuron states.

init_neurons(mbs, device)

Initialize all non-input layer states to zero.

Parameters:

• mbs – Minibatch size.

• device – Torch device for the state tensors.

Returns:

List of initialized neuron tensors.

class experiments.equilibrium_propagation.models.RON(archi, device, activation=<built-in method tanh of type object>, tau=1, epsilon_min=0, epsilon_max=1, gamma_min=0, gamma_max=1, learn_oscillators=True)

Bases: Module

Equilibrium-propagation Random Oscillator Network.

Parameters:

• archi – Sequence of layer widths including input and output sizes.

• device – Torch device for parameters and states.

• activation – Activation used by the state updates.

• tau – Oscillator integration scale.

• epsilon_min – Minimum initial epsilon value.

• epsilon_max – Maximum initial epsilon value.

• gamma_min – Minimum initial gamma value.

• gamma_max – Maximum initial gamma value.

• learn_oscillators – If True, learn gamma and epsilon.

Phi(x, y, neuronsz, neuronsy, beta, criterion)

Compute the full RON primitive for gradient updates.

Parameters:

• x – Input batch.

• y – Target labels.

• neuronsz – Current z state tensors.

• neuronsy – Current y state tensors.

• beta – Nudging coefficient.

• criterion – Loss function used by the primitive.

Returns:

Primitive value per sample.

Phi_statey(neuronsz, neuronsy)

Compute the primitive used for the y oscillator update.

Parameters:

• neuronsz – Current z state tensors.

• neuronsy – Current y state tensors.

Returns:

Primitive value per sample.

Phi_statez(x, y, neuronsy, beta, criterion)

Compute the primitive used for the z oscillator update.

Parameters:

• x – Input batch.

• y – Target labels.

• neuronsy – Current y state tensors.

• beta – Nudging coefficient.

• criterion – Loss function used by the primitive.

Returns:

Primitive value per sample.

compute_syn_grads(x, y, neurons_1, neurons_2, betas, criterion)

Compute synaptic gradients from two RON EP steady states.

Parameters:

• x – Input batch.

• y – Target labels.

• neurons_1 – First (z, y) steady state.

• neurons_2 – Second (z, y) steady state.

• betas – Pair of nudging coefficients.

• criterion – Loss function used by the primitive.

forward(x, y, neuronsz, neuronsy, T, beta=0.0, criterion=MSELoss())

Run RON fixed-point dynamics for T steps.

Parameters:

• x – Input batch.

• y – Target labels.

• neuronsz – Initial z state tensors.

• neuronsy – Initial y state tensors.

• T – Number of dynamics steps.

• beta – Nudging coefficient.

• criterion – Loss function used by the primitive.

Returns:

Tuple (neuronsz, neuronsy) after the updates.

init_neurons(mbs, device)

Initialize RON z and y state tensors.

Parameters:

• mbs – Minibatch size.

• device – Torch device for the state tensors.

Returns:

Tuple (neuronsz, neuronsy).

experiments.equilibrium_propagation.models.copy(neurons)

Clone neuron state tensors while preserving gradient tracking.

Parameters:

neurons – Iterable of state tensors.

Returns:

List of detached tensor copies with requires_grad enabled.

experiments.equilibrium_propagation.models.ctrd_hard_sig(x)

Apply a centered hard-sigmoid-like activation.

Parameters:

x – Input tensor.

Returns:

Activated tensor centered around zero.

experiments.equilibrium_propagation.models.evaluate(model, loader, T, device, ron=False)

Evaluate classification accuracy for fixed-point dynamics.

Parameters:

• model – EP model to evaluate.

• loader – Data loader yielding evaluation batches.

• T – Number of dynamics steps.

• device – Torch device for batches and states.

• ron – If True, use two-state RON dynamics.

Returns:

Accuracy in [0, 1].

experiments.equilibrium_propagation.models.evaluate_TS(model, loader, T, device, ron=False)

Evaluate classification accuracy on time-series data.

Parameters:

• model – EP model to evaluate.

• loader – Data loader yielding sequence batches.

• T – Number of dynamics steps per time step.

• device – Torch device for batches and states.

• ron – If True, use two-state RON dynamics.

Returns:

Accuracy in [0, 1].

experiments.equilibrium_propagation.models.hard_sigmoid(x)

Apply a hard-sigmoid activation on [0, 1].

Parameters:

x – Input tensor.

Returns:

Activated tensor.

experiments.equilibrium_propagation.models.make_pools(letters)

Build pooling layers from a compact letter specification.

Parameters:

letters – String where m means max pool, a average pool, and i identity.

Returns:

List of Torch pooling/identity modules.

experiments.equilibrium_propagation.models.my_hard_sig(x)

Apply the alternate hard-sigmoid activation used in experiments.

Parameters:

x – Input tensor.

Returns:

Activated tensor.

experiments.equilibrium_propagation.models.my_init(scale)

Create a scaled initializer for convolutional and linear layers.

Parameters:

scale – Multiplicative factor applied after Kaiming/uniform initialization.

Returns:

Initializer function suitable for Module.apply.

experiments.equilibrium_propagation.models.my_sigmoid(x)

Apply the shifted sigmoid activation used by the EP examples.

Parameters:

x – Input tensor.

Returns:

Activated tensor.

experiments.equilibrium_propagation.models.train_epoch(model, optimizer, epoch_number, train_loader, T1, T2, betas, device, criterion, alg='EP', random_sign=False, thirdphase=False, cep_debug=False, ron=False, id=None)

Train one epoch with EP, CEP, or BPTT dynamics.

Parameters:

• model – EP model to train.

• optimizer – Torch optimizer.

• epoch_number – Current epoch index used for logging.

• train_loader – Loader yielding training batches.

• T1 – Number of first-phase dynamics steps.

• T2 – Number of second-phase dynamics steps.

• betas – Pair of nudging coefficients.

• device – Torch device for batches and states.

• criterion – Loss function used by the primitive.

• alg – Training algorithm, such as EP, CEP, or BPTT.

• random_sign – If True, randomly flip the second beta sign.

• thirdphase – If True, run the third-phase correction.

• cep_debug – If True, apply CEP debug scaling.

• ron – If True, use two-state RON dynamics.

• id – Optional trial id that suppresses progress printing.

experiments.equilibrium_propagation.models.train_epoch_TS(model, optimizer, epoch_number, train_loader, T1, T2, betas, device, criterion, reset_factor=0.0, id=None, ron=False)

Train one epoch on time-series data with per-timestep updates.

Parameters:

• model – EP model to train.

• optimizer – Torch optimizer.

• epoch_number – Current epoch index used for logging.

• train_loader – Loader yielding sequence batches.

• T1 – Number of first-phase dynamics steps per time step.

• T2 – Number of second-phase dynamics steps per time step.

• betas – Pair of nudging coefficients.

• device – Torch device for batches and states.

• criterion – Loss function used by the primitive.

• reset_factor – Scale applied to carried state between time steps.

• id – Optional trial id that suppresses progress printing.

• ron – If True, use two-state RON dynamics.

experiments.equilibrium_propagation.models.visualize_convergence(model, loader, T_ep, device, ron=False, name=None)

Plot fixed-point convergence for non-time-series dynamics.

Parameters:

• model – EP model to evaluate.

• loader – Data loader used to fetch one batch.

• T_ep – Number of single-step EP iterations to visualize.

• device – Torch device for batches and states.

• ron – If True, use two-state RON dynamics.

• name – Optional plot title.

Returns:

Mean L2 differences between consecutive states.

experiments.equilibrium_propagation.models.visualize_convergence_TS(model, loader, T_ep, device, ron=False, name=None)

Plot fixed-point convergence for time-series dynamics.

Parameters:

• model – EP model to evaluate.

• loader – Data loader used to fetch one sequence batch.

• T_ep – Number of single-step EP iterations per time step.

• device – Torch device for batches and states.

• ron – If True, use two-state RON dynamics.

• name – Optional plot title.

Returns:

Mean L2 differences across all time steps and EP iterations.

experiments.equilibrium_propagation.pendigits_dataset module

class experiments.equilibrium_propagation.pendigits_dataset.PenDigitsDataset(ts_file)

Bases: Dataset

Torch dataset wrapper for PenDigits .ts files.

Parameters:

ts_file – Path to an aeon-compatible PenDigits time-series file.

Module contents

Equilibrium-propagation experiment models and datasets.