def test_match_neat():
with open("tests/test-genome.pkl", "rb") as f:
genome = pickle.load(f)
# use tanh since neat sets output nodes with no inputs to 0
# (sigmoid would output 0.5 for us)
def neat_tanh_activation(z):
return float(torch.tanh(2.5 * torch.tensor(z, dtype=torch.float64)))
for node in genome.nodes.values():
node.response = 0.5
config = neat.Config(
neat.DefaultGenome,
neat.DefaultReproduction,
neat.DefaultSpeciesSet,
neat.DefaultStagnation,
"tests/test-config.cfg",
)
for _ in range(500):
genome.mutate(config.genome_config)
# print(genome)
neat_net = neat.nn.RecurrentNetwork.create(genome, config)
for i, (node, _activation, aggregation, bias, response,
links) in enumerate(neat_net.node_evals):
neat_net.node_evals[i] = (
node,
neat_tanh_activation,
aggregation,
bias,
response,
links,
)
torch_net = RecurrentNet.create(genome,
config,
activation=tanh_activation,
prune_empty=True)
for _ in range(5):
inputs = np.random.randn(12)
# print(inputs)
neat_result = neat_net.activate(inputs)
torch_result = torch_net.activate([inputs])[0].numpy()
assert np.allclose(neat_result, torch_result, atol=1e-8)
def make_net(genome, config, bs):
# TODO: get number of agents directly instead of magic number
return RecurrentNet.create(genome, config, bs * 2)
def make_net(genome, config, bs):
return RecurrentNet.create(genome, config, bs)
def __init__(self, genome, config, batch_size=64):
self._batch_size = batch_size
self._net = RecurrentNet.create(genome, config, batch_size)