evolutionary-policy-optimization 0.2.8__py3-none-any.whl → 0.2.10__py3-none-any.whl

evolutionary_policy_optimization/epo.py

@@ -167,7 +167,8 @@ def shrink_and_perturb_(
 
     assert 0. <= shrink_factor <= 1.
 
-    maybe_sync_seed()
+    device = next(module.parameters()).device
+    maybe_sync_seed(device)
 
     for p in module.parameters():
         noise = torch.randn_like(p.data)
@@ -854,10 +855,11 @@ class LatentGenePool(Module):
         inplace = True,
         migrate = None # trigger a migration in the setting of multiple islands, the loop outside will need to have some `migrate_every` hyperparameter
     ):
-        maybe_sync_seed()
 
         device = self.latents.device
 
+        maybe_sync_seed(device)
+
         if not divisible_by(self.step.item(), self.apply_genetic_algorithm_every):
             self.advance_step_()
             return

evolutionary_policy_optimization/experimental.py

@@ -2,6 +2,7 @@ from random import uniform
 from copy import deepcopy
 
 import torch
+from torch import Tensor
 import torch.nn.functional as F
 from torch.func import vmap, functional_call
 from torch.nn import Module, ParameterList
@@ -14,7 +15,26 @@ def exists(v):
 def l2norm(t, dim = -1):
     return F.normalize(t, dim = dim)
 
-def crossover_weights(w1, w2):
+def shrink_and_perturb_(
+    t: Tensor,
+    shrink_factor = 0.4,
+    perturb_factor = 0.1
+):
+    # Shrink & Perturb
+    # Ash et al. https://arxiv.org/abs/1910.08475
+    # Applied to PBT NAS here https://arxiv.org/abs/2307.15621 - (0.4, 0.1)
+
+    assert 0. <= shrink_factor <= 1.
+    noise = torch.randn_like(t)
+    t.mul_(1. - shrink_factor).add_(noise * perturb_factor)
+    return t
+
+def crossover_weights(
+    w1, w2,
+    shrink_perturb = False,
+    shrink_factor = 0.4,
+    perturb_factor = 0.1
+):
     assert w2.shape == w2.shape
 
     no_batch = w1.ndim == 2
@@ -53,6 +73,9 @@ def crossover_weights(w1, w2):
     if no_batch:
         out = rearrange(out, '1 ... -> ...')
 
+    if shrink_perturb:
+        shrink_and_perturb_(out, shrink_factor = shrink_factor, perturb_factor = perturb_factor)
+
     return out
 
 def mutate_weight(

{evolutionary_policy_optimization-0.2.8.dist-info → evolutionary_policy_optimization-0.2.10.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: evolutionary-policy-optimization
-Version: 0.2.8
+Version: 0.2.10
 Summary: EPO - Pytorch
 Project-URL: Homepage, https://pypi.org/project/evolutionary-policy-optimization/
 Project-URL: Repository, https://github.com/lucidrains/evolutionary-policy-optimization
@@ -278,4 +278,15 @@ That's it
 }
 ```
 
+```bibtex
+@article{Chebykin2023ShrinkPerturbIA,
+    title   = {Shrink-Perturb Improves Architecture Mixing during Population Based Training for Neural Architecture Search},
+    author  = {Alexander Chebykin and Arkadiy Dushatskiy and Tanja Alderliesten and Peter A. N. Bosman},
+    journal = {ArXiv},
+    year    = {2023},
+    volume  = {abs/2307.15621},
+    url     = {https://api.semanticscholar.org/CorpusID:260316291}
+}
+```
+
 *Evolution is cleverer than you are.* - Leslie Orgel

evolutionary_policy_optimization-0.2.10.dist-info/RECORD

@@ -0,0 +1,10 @@
+evolutionary_policy_optimization/__init__.py,sha256=NyiYDYU7DlpmOTM7xiBQET3r1WwX0ebrgMCBLSQrW3c,288
+evolutionary_policy_optimization/distributed.py,sha256=MxyxqxANAuOm8GYb0Yu09EHd_aVLhK2uwgrfuVWciPU,2342
+evolutionary_policy_optimization/env_wrappers.py,sha256=bDL06o9_b1iW6k3fw2xifnOnYlzs643tdW6Yv2gsIdw,803
+evolutionary_policy_optimization/epo.py,sha256=adUHDtgrXnSOoPLnbPy6xxGl6QLYxbN1mB_sl2KPwgI,53135
+evolutionary_policy_optimization/experimental.py,sha256=QZG0__wwFqHN_LJK7e-mHxlIL1mwjlvG6o6bcOpeAKg,6166
+evolutionary_policy_optimization/mock_env.py,sha256=TLyyRm6tOD0Kdn9QqJJQriaSnsR-YmNQHo4OohmZFG4,1410
+evolutionary_policy_optimization-0.2.10.dist-info/METADATA,sha256=dLg2z2BFihNUd0dirhDcabw9yNj0V_2v42Tk_Pg-EwY,9601
+evolutionary_policy_optimization-0.2.10.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+evolutionary_policy_optimization-0.2.10.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+evolutionary_policy_optimization-0.2.10.dist-info/RECORD,,

evolutionary_policy_optimization-0.2.8.dist-info/RECORD

@@ -1,10 +0,0 @@
-evolutionary_policy_optimization/__init__.py,sha256=NyiYDYU7DlpmOTM7xiBQET3r1WwX0ebrgMCBLSQrW3c,288
-evolutionary_policy_optimization/distributed.py,sha256=MxyxqxANAuOm8GYb0Yu09EHd_aVLhK2uwgrfuVWciPU,2342
-evolutionary_policy_optimization/env_wrappers.py,sha256=bDL06o9_b1iW6k3fw2xifnOnYlzs643tdW6Yv2gsIdw,803
-evolutionary_policy_optimization/epo.py,sha256=San0DS_Z0Ueze6Rz1BKap7SKFww1bx3U0WansTRprdo,53076
-evolutionary_policy_optimization/experimental.py,sha256=7LOrMIaU4fr2Vme1ZpHNIvlvFEIdWj0-uemhQoNJcPQ,5549
-evolutionary_policy_optimization/mock_env.py,sha256=TLyyRm6tOD0Kdn9QqJJQriaSnsR-YmNQHo4OohmZFG4,1410
-evolutionary_policy_optimization-0.2.8.dist-info/METADATA,sha256=PcsHBD4zX1EsV-kWFJIR5fsLxX6hXFPOF4FJyfaI9oU,9171
-evolutionary_policy_optimization-0.2.8.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-evolutionary_policy_optimization-0.2.8.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-evolutionary_policy_optimization-0.2.8.dist-info/RECORD,,