sgd_optimiser_type

sgd_optimiser_type(
   learning_rate=0.01,
   momentum=0.0,
   nesterov=.false.,
   num_params=1,
   regulariser=None,
   clip_dict,
   lr_decay
)

The sgd_optimiser_type derived type provides a data structure that contains all optimisation/learning parameters for a network model. Most simply, it defines the learning rate for the model.

This type provides an implementation of the momentum-based Stochastic Gradient Descent (SGD) method.

Arguments

• learning_rate: A real scalar. The rate of learning applied to the weights.
• momentum: A real scalar. The momentum parameter used for momentum-based learning.
• nesterov: A boolean whether to use Nesterov Accelerated Gradient (NAG).
• num_params: An integer scalar. The number of learnable parameters in the model.
• regulariser: A derived type extended from the base_regulariser_type derived type.
• clip_dict: A derived data type defining weight clipping parameters. These nested parameters can be set using the set_clip procedure.
• lr_decay: A derived type extended from the base_lr_decay_type derived type.