neat-config

[NEAT]
pop_size              = 50
fitness_criterion     = max
fitness_threshold     = 10000.0
reset_on_extinction   = False

[OptimizerGenome]
activation_default      = none
activation_mutate_rate  = 2.5
aggregation_default     = sum
aggregation_mutate_rate = 2.5
attribute_add_prob      = 0.03
attribute_delete_prob   = 0.03

conn_add_prob         = 0.5
conn_delete_prob      = 0.05
node_add_prob         = 0.2
node_delete_prob      = 0.02

node_type_mutate_rate = 0.01

# TorchScript optimizers take (loss, prev_loss, named_parameters) and return
# a single dict of updated parameters, so every decoded graph must expose
# three input pins and one output pin.
num_inputs            = 3
num_outputs           = 1

compatibility_disjoint_coefficient = 1.0
compatibility_weight_coefficient   = 0.5

enabled_default     = True
enabled_mutate_rate = .01

[DefaultSpeciesSet]
compatibility_threshold = 2.5

[RelativeRankStagnation]
species_fitness_func = max
max_stagnation       = 25
species_elitism      = 5
rank_tolerance       = 1e-06

[GuidedReproduction]
elitism            = 5
survival_threshold = 0.2

[GuidedPopulation]
kl_partial_slice_ratio = 0.5
# kl_partial_slice_dims = 8
kl_partial_slice_start = 0
wl_kernel_loss_weight = 1.0
wl_kernel_iterations = 5
# NOTE: each phase here signifies what is trained not what is frozen
trainer_freeze_cycle = all
trainer_freeze_verbose = False
guided_debug_dump_enabled = False
guided_invalid_viz_enabled = True
guided_invalid_viz_limit = 10
teacher_force_epochs = 5
teacher_force_epochs_max = 15
teacher_force_weight = 2.0
teacher_force_weight_max = 4.0
teacher_force_verbose = True
trainer_dataset_max_samples = 256        # 0 disables subsampling
trainer_dataset_valid_ratio = 0.75       # only used when max_samples > 0
trainer_seed_generations = 20            # number of early generations that must keep seed graphs