evolutionary-policy-optimization 0.2.9__tar.gz → 0.2.11__tar.gz

{evolutionary_policy_optimization-0.2.9 → evolutionary_policy_optimization-0.2.11}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: evolutionary-policy-optimization
-Version: 0.2.9
+Version: 0.2.11
 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-0.2.9 → evolutionary_policy_optimization-0.2.11}/README.md

@@ -221,4 +221,15 @@ 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-0.2.9 → evolutionary_policy_optimization-0.2.11}/evolutionary_policy_optimization/distributed.py

@@ -61,6 +61,8 @@ def has_only_one_value(t):
     return (t == t[0]).all()
 
 def all_gather_variable_dim(t, dim = 0, sizes = None):
+    device = t.device
+
     if not exists(sizes):
         sizes = gather_sizes(t, dim = dim)
 
@@ -77,7 +79,7 @@ def all_gather_variable_dim(t, dim = 0, sizes = None):
     gathered_tensors = torch.cat(gathered_tensors, dim = dim)
     seq = torch.arange(max_size, device = device)
 
-    mask = einx.less('j i -> (i j)', seq, sizes)
+    mask = einx.less('j, i -> (i j)', seq, sizes)
     seq = torch.arange(mask.shape[-1], device = device)
     indices = seq[mask]
 

{evolutionary_policy_optimization-0.2.9 → evolutionary_policy_optimization-0.2.11}/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-0.2.9 → evolutionary_policy_optimization-0.2.11}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "evolutionary-policy-optimization"
-version = "0.2.9"
+version = "0.2.11"
 description = "EPO - Pytorch"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{evolutionary_policy_optimization-0.2.9 → evolutionary_policy_optimization-0.2.11}/tests/test_epo.py

@@ -58,7 +58,8 @@ def test_create_agent(
         actor_dim = 256,
         actor_mlp_depth = 2,
         critic_dim = 256,
-        critic_mlp_depth = 4
+        critic_mlp_depth = 4,
+        wrap_with_accelerate = False
     )
 
     state = torch.randn(2, 512)