tony/tensorneat-mend
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

try to accelerate the speed of speciate

Browse Source
This commit is contained in:
wls2002
2023-05-08 18:41:19 +08:00
parent 8653f49826
commit ee6bb01eff
4 changed files with 12 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
algorithms/neat/pipeline.py
View File

@@ -2,6 +2,7 @@ from typing import List, Union, Tuple, Callable
import time
import jax
import jax.numpy as jnp
import numpy as np
from .species import SpeciesController
@@ -104,7 +105,6 @@ class Pipeline:
lpc = self.pop_connections[crossover_pair[:, 1]] # loser pop connections
npn, npc = self.crossover_func(crossover_rand_keys, wpn, wpc, lpn,
lpc) # new pop nodes, new pop connections
npn, npc = jax.device_get(npn), jax.device_get(npc)
# mutate
mutate_rand_keys = jax.random.split(k2, self.pop_size)
@@ -113,11 +113,8 @@ class Pipeline:
m_npn, m_npc = self.mutate_func(mutate_rand_keys, npn, npc, new_node_keys) # mutate_new_pop_nodes
# elitism don't mutate
# (pop_size, ) to (pop_size, 1, 1)
m_npn, m_npc = jax.device_get(m_npn), jax.device_get(m_npc)
npn, npc, m_npn, m_npc = jax.device_get([npn, npc, m_npn, m_npc])
self.pop_nodes = np.where(elitism_mask[:, None, None], npn, m_npn)
# (pop_size, ) to (pop_size, 1, 1, 1)
self.pop_connections = np.where(elitism_mask[:, None, None, None], npc, m_npc)
def expand(self):

12
algorithms/neat/species.py
View File

@@ -76,11 +76,13 @@ class SpeciesController:
new_representatives = {}
new_members = {}
for sid, species in self.species.items():
# calculate the distance between the representative and the population
r_nodes, r_connections = species.representative
distances = o2m_distance(r_nodes, r_connections, pop_nodes, pop_connections)
distances = jax.device_get(distances)
total_distances = jax.device_get([
o2m_distance(*self.species[sid].representative, pop_nodes, pop_connections)
for sid in previous_species_list
])
for i, sid in enumerate(previous_species_list):
distances = total_distances[i]
min_idx = find_min_with_mask(distances, unspeciated) # find the min un-specified distance
new_representatives[sid] = min_idx