evolutionary-policy-optimization 0.0.11__tar.gz → 0.0.14__tar.gz

{evolutionary_policy_optimization-0.0.11 → evolutionary_policy_optimization-0.0.14}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: evolutionary-policy-optimization
-Version: 0.0.11
+Version: 0.0.14
 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-0.0.11 → evolutionary_policy_optimization-0.0.14}/evolutionary_policy_optimization/__init__.py

@@ -2,6 +2,7 @@ from evolutionary_policy_optimization.epo import (
     MLP,
     Actor,
     Critic,
+    create_agent,
     Agent,
     LatentGenePool
 )

{evolutionary_policy_optimization-0.0.11 → evolutionary_policy_optimization-0.0.14}/evolutionary_policy_optimization/epo.py

@@ -502,6 +502,38 @@ class LatentGenePool(Module):
 
 # agent contains the actor, critic, and the latent genetic pool
 
+def create_agent(
+    dim_state,
+    num_latents,
+    dim_latent,
+    actor_num_actions,
+    actor_dim_hiddens: int | tuple[int, ...],
+    critic_dim_hiddens: int | tuple[int, ...],
+    num_latent_sets = 1
+) -> Agent:
+
+    actor = Actor(
+        num_actions = actor_num_actions,
+        dim_state = dim_state,
+        dim_latent = dim_latent,
+        dim_hiddens = actor_dim_hiddens
+    )
+
+    critic = Critic(
+        dim_state = dim_state,
+        dim_latent = dim_latent,
+        dim_hiddens = critic_dim_hiddens
+    )
+
+    latent_gene_pool = LatentGenePool(
+        dim_state = dim_state,
+        num_latents = num_latents,
+        dim_latent = dim_latent,
+        num_latent_sets = num_latent_sets
+    )
+
+    return Agent(actor = actor, critic = critic, latent_gene_pool = latent_gene_pool)
+
 class Agent(Module):
     def __init__(
         self,
@@ -516,6 +548,28 @@ class Agent(Module):
 
         self.latent_gene_pool = latent_gene_pool
 
+    def get_actor_actions(
+        self,
+        state,
+        latent_id
+    ):
+        latent = self.latent_gene_pool(latent_id = latent_id, state = state)
+        return self.actor(state, latent)
+
+    def get_critic_values(
+        self,
+        state,
+        latent_id
+    ):
+        latent = self.latent_gene_pool(latent_id = latent_id, state = state)
+        return self.critic(state, latent)
+
+    def update_latent_gene_pool_(
+        self,
+        fitnesses
+    ):
+        return self.latent_gene_pool.genetic_algorithm_step(fitnesses)
+
     def forward(
         self,
         memories: list[Memory]
@@ -539,13 +593,10 @@ class EPO(Module):
 
     def __init__(
         self,
-        agent: Agent,
-        latent_gene_pool: LatentGenePool
+        agent: Agent
     ):
         super().__init__()
-
         self.agent = agent
-        self.latent_gene_pool = latent_gene_pool
 
     def forward(
         self,

{evolutionary_policy_optimization-0.0.11 → evolutionary_policy_optimization-0.0.14}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "evolutionary-policy-optimization"
-version = "0.0.11"
+version = "0.0.14"
 description = "EPO - Pytorch"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{evolutionary_policy_optimization-0.0.11 → evolutionary_policy_optimization-0.0.14}/tests/test_epo.py

@@ -36,3 +36,29 @@ def test_readme(
     fitness = torch.randn(128)
 
     latent_pool.genetic_algorithm_step(fitness) # update once
+
+
+def test_create_agent():
+    from evolutionary_policy_optimization import create_agent
+
+    agent = create_agent(
+        dim_state = 512,
+        num_latents = 128,
+        dim_latent = 32,
+        actor_num_actions = 5,
+        actor_dim_hiddens = (256, 128),
+        critic_dim_hiddens = (256, 128, 64)
+    )
+
+    state = torch.randn(1, 512)
+
+    actions = agent.get_actor_actions(state, latent_id = 3)
+    value = agent.get_critic_values(state, latent_id = 3)
+
+    # interact with environment and receive rewards, termination etc
+
+    # derive a fitness score for each gene / latent
+
+    fitness = torch.randn(128)
+
+    agent.update_latent_gene_pool_(fitness) # update once