evolutionary-policy-optimization 0.1.7__tar.gz → 0.1.8__tar.gz

{evolutionary_policy_optimization-0.1.7 → evolutionary_policy_optimization-0.1.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: evolutionary-policy-optimization
-Version: 0.1.7
+Version: 0.1.8
 Summary: EPO - Pytorch
 Project-URL: Homepage, https://pypi.org/project/evolutionary-policy-optimization/
 Project-URL: Repository, https://github.com/lucidrains/evolutionary-policy-optimization
@@ -196,4 +196,14 @@ agent.load('./agent.pt')
 }
 ```
 
+```bibtex
+@article{Doerr2017FastGA,
+    title   = {Fast genetic algorithms},
+    author  = {Benjamin Doerr and Huu Phuoc Le and R{\'e}gis Makhmara and Ta Duy Nguyen},
+    journal = {Proceedings of the Genetic and Evolutionary Computation Conference},
+    year    = {2017},
+    url     = {https://api.semanticscholar.org/CorpusID:16196841}
+}
+```
+
 *Evolution is cleverer than you are.* - Leslie Orgel

{evolutionary_policy_optimization-0.1.7 → evolutionary_policy_optimization-0.1.8}/README.md

@@ -143,4 +143,14 @@ agent.load('./agent.pt')
 }
 ```
 
+```bibtex
+@article{Doerr2017FastGA,
+    title   = {Fast genetic algorithms},
+    author  = {Benjamin Doerr and Huu Phuoc Le and R{\'e}gis Makhmara and Ta Duy Nguyen},
+    journal = {Proceedings of the Genetic and Evolutionary Computation Conference},
+    year    = {2017},
+    url     = {https://api.semanticscholar.org/CorpusID:16196841}
+}
+```
+
 *Evolution is cleverer than you are.* - Leslie Orgel

{evolutionary_policy_optimization-0.1.7 → evolutionary_policy_optimization-0.1.8}/evolutionary_policy_optimization/epo.py

@@ -207,13 +207,60 @@ def mutation(
 ):
     mutations = torch.randn_like(latents)
 
-    mutated = latents + mutations * mutation_strength
+    if is_tensor(mutation_strength):
+        mutations = einx.multiply('b, b ...', mutation_strength, mutations)
+    else:
+        mutations *= mutation_strength
+
+    mutated = latents + mutations
 
     if not l2norm_output:
         return mutated
 
     return l2norm(mutated)
 
+# drawing mutation strengths from power law distribution
+# proposed by https://arxiv.org/abs/1703.03334
+
+class PowerLawDist(Module):
+    def __init__(
+        self,
+        values: Tensor | list[float] | None = None,
+        bins = None,
+        beta = 1.5,
+    ):
+        super().__init__()
+        assert beta > 1.
+
+        assert exists(bins) or exists(values)
+
+        if exists(values):
+            if not is_tensor(values):
+                values = tensor(values)
+
+            assert values.ndim == 1
+            bins = values.shape[0]
+
+        self.beta = beta
+
+        cdf = torch.linspace(1, bins, bins).pow(-beta).cumsum(dim = -1)
+        cdf = cdf / cdf[-1]
+
+        self.register_buffer('cdf', cdf)
+        self.register_buffer('values', values)
+
+    def forward(self, shape):
+        device = self.cdf.device
+
+        uniform = torch.rand(shape, device = device)
+
+        sampled = torch.searchsorted(self.cdf, uniform)
+
+        if not exists(self.values):
+            return sampled
+
+        return self.values[sampled]
+
 # simple MLP networks, but with latent variables
 # the latent variables are the "genes" with the rest of the network as the scaffold for "gene expression" - as suggested in the paper
 
@@ -441,6 +488,8 @@ class LatentGenePool(Module):
         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)
         mutation_strength = 1.,          # factor to multiply to gaussian noise as mutation to latents
+        fast_genetic_algorithm = False,
+        fast_ga_values = torch.linspace(1, 5, 10),
         should_run_genetic_algorithm: Module | None = None, # eq (3) in paper
         default_should_run_ga_gamma = 1.5,
         migrate_every = 100,                 # how many steps before a migration between islands
@@ -488,6 +537,8 @@ class LatentGenePool(Module):
         self.crossover_random  = crossover_random
 
         self.mutation_strength = mutation_strength
+        self.mutation_strength_sampler = PowerLawDist(fast_ga_values) if fast_genetic_algorithm else None
+
         self.num_elites = int(frac_elitism * latents_per_island)
         self.has_elites = self.num_elites > 0
 
@@ -656,9 +707,14 @@ class LatentGenePool(Module):
         if self.has_elites:
             genes, elites = genes[:, :-self.num_elites], genes[:, -self.num_elites:]
 
-        # 5. mutate with gaussian noise - todo: add drawing the mutation rate from exponential distribution, from the fast genetic algorithms paper from 2017
+        # 5. mutate with gaussian noise
+
+        if exists(self.mutation_strength_sampler):
+            mutation_strength = self.mutation_strength_sampler(genes.shape[:1])
+        else:
+            mutation_strength = self.mutation_strength
 
-        genes = mutation(genes, mutation_strength = self.mutation_strength)
+        genes = mutation(genes, mutation_strength = mutation_strength)
 
         # 6. maybe migration
 

{evolutionary_policy_optimization-0.1.7 → evolutionary_policy_optimization-0.1.8}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "evolutionary-policy-optimization"
-version = "0.1.7"
+version = "0.1.8"
 description = "EPO - Pytorch"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{evolutionary_policy_optimization-0.1.7 → evolutionary_policy_optimization-0.1.8}/tests/test_epo.py

@@ -9,15 +9,18 @@ from evolutionary_policy_optimization import (
 
 @pytest.mark.parametrize('latent_ids', (2, (2, 4)))
 @pytest.mark.parametrize('num_islands', (1, 4))
+@pytest.mark.parametrize('sampled_mutation_strengths', (False, True))
 def test_readme(
     latent_ids,
-    num_islands
+    num_islands,
+    sampled_mutation_strengths
 ):
 
     latent_pool = LatentGenePool(
         num_latents = 128,
         dim_latent = 32,        
         num_islands = num_islands,
+        fast_genetic_algorithm = sampled_mutation_strengths
     )
 
     state = torch.randn(2, 512)