366eb0988cad154ad05d356b1ec950c8cb0206d6,paysage/fit.py,,persistent_contrastive_divergence,#Any#Any#Any#Any#,382

Before Change

    sampler.update_negative_state(steps)

    // compute the gradient
    return model.gradient(*sampler.get_states())

// alias
pcd = persistent_contrastive_divergence

After Change

    // compute the conditional sampling on all visible-side layers,
    // inclusive over hidden-side layers
    for i in range(1, model.num_layers - 1):
        clamped_layers = list(range(i))
        sampler.update_positive_state(steps, clamped=clamped_layers)
        sampler.update_negative_state(steps, clamped=clamped_layers)

    // make a mean field step to copmute the expectation on the last layer
    clamped_layers = list(range(model.num_layers - 1))
    grad_data_state = model.mean_field_iteration(1, sampler.pos_state, clamped=clamped_layers)
    grad_model_state = model.mean_field_iteration(1, sampler.neg_state, clamped=clamped_layers)

Instances