evolutionary-policy-optimization 0.0.46__py3-none-any.whl → 0.0.48__py3-none-any.whl

evolutionary_policy_optimization/epo.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 from functools import partial, wraps
 from pathlib import Path
 from collections import namedtuple
+from random import randrange
 
 import torch
 from torch import nn, cat, stack, is_tensor, tensor
@@ -356,9 +357,10 @@ class LatentGenePool(Module):
         frac_natural_selected = 0.25,    # number of least fit genes to remove from the pool
         frac_elitism = 0.1,              # frac of population to preserve from being noised
         frac_migrate = 0.1,              # frac of population, excluding elites, that migrate between islands randomly. will use a designated set migration pattern (since for some reason using random it seems to be worse for me)
+        migrate_every = 100,             # how many steps before a migration between islands
         mutation_strength = 1.,          # factor to multiply to gaussian noise as mutation to latents
         should_run_genetic_algorithm: Module | None = None, # eq (3) in paper
-        default_should_run_ga_gamma = 1.5
+        default_should_run_ga_gamma = 1.5,
     ):
         super().__init__()
 
@@ -408,6 +410,13 @@ class LatentGenePool(Module):
 
         self.should_run_genetic_algorithm = should_run_genetic_algorithm
 
+        self.can_migrate = num_islands > 1
+        self.migrate_every = migrate_every
+        self.register_buffer('step', tensor(0))
+
+    def advance_step_(self):
+        self.step.add_(1)
+
     def firefly_step(
         self,
         fitness,
@@ -459,7 +468,7 @@ class LatentGenePool(Module):
         self,
         fitness, # Float['p'],
         inplace = True,
-        migrate = False # trigger a migration in the setting of multiple islands, the loop outside will need to have some `migrate_every` hyperparameter
+        migrate = None # trigger a migration in the setting of multiple islands, the loop outside will need to have some `migrate_every` hyperparameter
     ):
         device = self.latents.device
 
@@ -546,8 +555,9 @@ class LatentGenePool(Module):
 
         # 6. maybe migration
 
+        migrate = self.can_migrate and default(migrate, divisible_by(self.step.item(), self.migrate_every))
+
         if migrate:
-            assert self.num_islands > 1
             randperm = torch.randn(genes.shape[:-1], device = device).argsort(dim = -1)
 
             migrate_mask = randperm < self.num_migrate
@@ -580,6 +590,8 @@ class LatentGenePool(Module):
 
         self.latents.copy_(genes)
 
+        self.advance_step_()
+
     def forward(
         self,
         *args,
@@ -989,7 +1001,8 @@ class EPO(Module):
     @torch.no_grad()
     def forward(
         self,
-        env
+        env,
+        fix_seed_across_latents = True
     ) -> MemoriesAndCumulativeRewards:
 
         self.agent.eval()
@@ -1002,12 +1015,22 @@ class EPO(Module):
 
         for episode_id in tqdm(range(self.episodes_per_latent), desc = 'episode'):
 
+            # maybe fix seed for environment across all latents
+
+            env_reset_kwargs = dict()
+
+            if fix_seed_across_latents:
+                seed = randrange(int(1e6))
+                env_reset_kwargs = dict(seed = seed)
+
+            # for each latent (on a single machine for now)
+
             for latent_id in tqdm(range(self.num_latents), desc = 'latent'):
                 time = 0
 
                 # initial state
 
-                state = env.reset()
+                state = env.reset(**env_reset_kwargs)
 
                 # get latent from pool
 

evolutionary_policy_optimization/mock_env.py

@@ -25,7 +25,8 @@ class Env(Module):
         return self.dummy.device
 
     def reset(
-        self
+        self,
+        seed
     ):
         state = randn(self.state_shape, device = self.device)
         return state

{evolutionary_policy_optimization-0.0.46.dist-info → evolutionary_policy_optimization-0.0.48.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: evolutionary-policy-optimization
-Version: 0.0.46
+Version: 0.0.48
 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-0.0.48.dist-info/RECORD

@@ -0,0 +1,8 @@
+evolutionary_policy_optimization/__init__.py,sha256=0q0aBuFgWi06MLMD8FiHzBYQ3_W4LYWrwmCtF3u5H2A,201
+evolutionary_policy_optimization/epo.py,sha256=FkliOiKdmUvKuwFqb1_A-ddahnOqjTR8Djx_I6UZAlU,32625
+evolutionary_policy_optimization/experimental.py,sha256=9FrJGviLESlYysHI3i83efT9g2ZB9ha4u3K9HXN98_w,1100
+evolutionary_policy_optimization/mock_env.py,sha256=202KJ5g57wQvOzhGYzgHfBa7Y2do5uuDvl5kFg5o73g,934
+evolutionary_policy_optimization-0.0.48.dist-info/METADATA,sha256=GpuUVs0VO2ydhU3X4-A_cA_xNmvdtYaAM8tb_VKneBo,6213
+evolutionary_policy_optimization-0.0.48.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+evolutionary_policy_optimization-0.0.48.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+evolutionary_policy_optimization-0.0.48.dist-info/RECORD,,

evolutionary_policy_optimization-0.0.46.dist-info/RECORD

@@ -1,8 +0,0 @@
-evolutionary_policy_optimization/__init__.py,sha256=0q0aBuFgWi06MLMD8FiHzBYQ3_W4LYWrwmCtF3u5H2A,201
-evolutionary_policy_optimization/epo.py,sha256=SAhWgRY8uPQEKFg1_nz1mvh8A6S_sHwnDykhd0F5xEI,31853
-evolutionary_policy_optimization/experimental.py,sha256=9FrJGviLESlYysHI3i83efT9g2ZB9ha4u3K9HXN98_w,1100
-evolutionary_policy_optimization/mock_env.py,sha256=6AIc4mwL_C6JkAxwESJgCLxXHMzCAu2FcffVg3HkSm0,920
-evolutionary_policy_optimization-0.0.46.dist-info/METADATA,sha256=xP2kdKo52-X4Z5XXTPpW0M_NFI0spuigeL7fvqFlsRM,6213
-evolutionary_policy_optimization-0.0.46.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-evolutionary_policy_optimization-0.0.46.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-evolutionary_policy_optimization-0.0.46.dist-info/RECORD,,