evolutionary-policy-optimization 0.0.50__py3-none-any.whl → 0.0.52__py3-none-any.whl

evolutionary_policy_optimization/epo.py

@@ -48,6 +48,9 @@ def divisible_by(num, den):
 def l2norm(t):
     return F.normalize(t, p = 2, dim = -1)
 
+def batch_randperm(shape, device):
+    return torch.randn(shape, device = device).argsort(dim = -1)
+
 def log(t, eps = 1e-20):
     return t.clamp(min = eps).log()
 
@@ -393,7 +396,6 @@ class LatentGenePool(Module):
 
         latents_per_island = num_latents // num_islands
         self.num_natural_selected = int(frac_natural_selected * latents_per_island)
-
         self.num_tournament_participants = int(frac_tournaments * self.num_natural_selected)
 
         self.crossover_random  = crossover_random
@@ -530,7 +532,9 @@ class LatentGenePool(Module):
 
         # 2. for finding pairs of parents to replete gene pool, we will go with the popular tournament strategy
 
-        rand_tournament_gene_ids = torch.randn((islands, pop_size_per_island - self.num_natural_selected, tournament_participants), device = device).argsort(dim = -1)
+        tournament_shape = (islands, pop_size_per_island - self.num_natural_selected, self.num_natural_selected) # (island, num children needed, natural selected population to be bred)
+
+        rand_tournament_gene_ids = batch_randperm(tournament_shape, device)[..., :tournament_participants]
         rand_tournament_gene_ids_for_gather = rearrange(rand_tournament_gene_ids, 'i p t -> i (p t)')
 
         participant_fitness = fitness.gather(1, rand_tournament_gene_ids_for_gather)
@@ -901,7 +905,7 @@ class Agent(Module):
 
                 if self.has_diversity_loss:
                     diversity = self.latent_gene_pool.get_distance()
-                    diversity_loss = diversity.mul(-1).exp().mean()
+                    diversity_loss = (-diversity).tril(-1).exp().mean()
 
                     (diversity_loss * self.diversity_aux_loss_weight).backward()
 

{evolutionary_policy_optimization-0.0.50.dist-info → evolutionary_policy_optimization-0.0.52.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: evolutionary-policy-optimization
-Version: 0.0.50
+Version: 0.0.52
 Summary: EPO - Pytorch
 Project-URL: Homepage, https://pypi.org/project/evolutionary-policy-optimization/
 Project-URL: Repository, https://github.com/lucidrains/evolutionary-policy-optimization

{evolutionary_policy_optimization-0.0.50.dist-info → evolutionary_policy_optimization-0.0.52.dist-info}/RECORD

@@ -1,8 +1,8 @@
 evolutionary_policy_optimization/__init__.py,sha256=0q0aBuFgWi06MLMD8FiHzBYQ3_W4LYWrwmCtF3u5H2A,201
-evolutionary_policy_optimization/epo.py,sha256=nIZzjz_3WUvHYPwNH9zoKM3fyRJd3MOd6UL7ooSTQV4,33445
+evolutionary_policy_optimization/epo.py,sha256=7iGmye36JsQ1jUm_UCCrSdvC67u70EPTAp3nCbKGjO8,33675
 evolutionary_policy_optimization/experimental.py,sha256=9FrJGviLESlYysHI3i83efT9g2ZB9ha4u3K9HXN98_w,1100
 evolutionary_policy_optimization/mock_env.py,sha256=202KJ5g57wQvOzhGYzgHfBa7Y2do5uuDvl5kFg5o73g,934
-evolutionary_policy_optimization-0.0.50.dist-info/METADATA,sha256=iCB34UsXdosJV9q8-qXJpJfI8OdNNXB5o8MhmKy82zY,6213
-evolutionary_policy_optimization-0.0.50.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-evolutionary_policy_optimization-0.0.50.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-evolutionary_policy_optimization-0.0.50.dist-info/RECORD,,
+evolutionary_policy_optimization-0.0.52.dist-info/METADATA,sha256=dET2Z11ktfJYei-xFfv7SjGKTglco6dDvELIzKrLsAQ,6213
+evolutionary_policy_optimization-0.0.52.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+evolutionary_policy_optimization-0.0.52.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+evolutionary_policy_optimization-0.0.52.dist-info/RECORD,,