agilerl 2.5.0.dev2__tar.gz → 2.5.0.dev4__tar.gz

{agilerl-2.5.0.dev2 → agilerl-2.5.0.dev4}/.pre-commit-config.yaml

@@ -31,7 +31,7 @@ repos:
                 - --skip=*.css,*.js,*.map,*.scss,*.svg
                 - --ignore-words-list=magent,pres,roate
     - repo: https://github.com/astral-sh/ruff-pre-commit
-      rev: v0.15.1
+      rev: v0.15.2
       hooks:
           - id: ruff
             name: Ruff Linter
@@ -46,6 +46,6 @@ repos:
 
     - repo: https://github.com/astral-sh/uv-pre-commit
       # uv version.
-      rev: 0.10.3
+      rev: 0.10.4
       hooks:
           - id: uv-lock

{agilerl-2.5.0.dev2 → agilerl-2.5.0.dev4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agilerl
-Version: 2.5.0.dev2
+Version: 2.5.0.dev4
 Summary: AgileRL is a deep reinforcement learning library focused on improving RL development through RLOps.
 Author-email: Nick Ustaran-Anderegg <dev@agilerl.com>
 License-Expression: Apache-2.0

{agilerl-2.5.0.dev2 → agilerl-2.5.0.dev4}/agilerl/algorithms/core/base.py

@@ -77,6 +77,7 @@ from agilerl.utils.algo_utils import (
     chkpt_attribute_to_device,
     clone_llm,
     create_warmup_cosine_scheduler,
+    filter_init_dict,
     get_input_size_from_space,
     get_output_size_from_space,
     isroutine,
@@ -1094,22 +1095,20 @@ class EvolvableAlgorithm(ABC, metaclass=RegistryMeta):
             )
             if isinstance(module_cls, dict):
                 for agent_id, mod_cls in module_cls.items():
-                    d = init_dict[agent_id]
+                    d = filter_init_dict(init_dict[agent_id], mod_cls)
                     d["device"] = device
                     mod: EvolvableModule = mod_cls(**d)
                     loaded_modules[name][agent_id] = mod
             else:
+                init_dict = filter_init_dict(init_dict, module_cls)
                 init_dict["device"] = device
                 module = module_cls(**init_dict)
                 loaded_modules[name] = module
 
         # Reconstruct the algorithm
-        constructor_params = inspect.signature(cls.__init__).parameters.keys()
         checkpoint["accelerator"] = accelerator
         checkpoint["device"] = device
-        class_init_dict = {
-            k: v for k, v in checkpoint.items() if k in constructor_params
-        }
+        class_init_dict = filter_init_dict(checkpoint, cls)
         self = cls(**class_init_dict)
         registry: MutationRegistry = checkpoint["registry"]
         self.registry = registry

{agilerl-2.5.0.dev2 → agilerl-2.5.0.dev4}/agilerl/algorithms/maddpg.py

@@ -299,7 +299,7 @@ class MADDPG(MultiAgentRLAlgorithm):
             )
             max_latent_dim = max(
                 [
-                    agent_configs[agent_id].get("max_latent_dim", 128)
+                    agent_configs[agent_id].get("max_latent_dim", 1024)
                     for agent_id in self.agent_ids
                 ],
             )

{agilerl-2.5.0.dev2 → agilerl-2.5.0.dev4}/agilerl/algorithms/matd3.py

@@ -337,7 +337,7 @@ class MATD3(MultiAgentRLAlgorithm):
             )
             max_latent_dim = max(
                 [
-                    agent_configs[agent_id].get("max_latent_dim", 128)
+                    agent_configs[agent_id].get("max_latent_dim", 1024)
                     for agent_id in self.agent_ids
                 ],
             )

{agilerl-2.5.0.dev2 → agilerl-2.5.0.dev4}/agilerl/utils/algo_utils.py

@@ -452,6 +452,21 @@ def chkpt_attribute_to_device(
     return chkpt_dict
 
 
+def filter_init_dict(init_dict: dict[str, Any], cls: type) -> dict[str, Any]:
+    """Filter the init dict to only include parameters that are valid for the given class.
+
+    :param init_dict: Initialization dictionary
+    :type init_dict: dict[str, Any]
+    :param cls: Class to filter the init dict for
+    :type cls: type
+
+    :return: Filtered initialization dictionary
+    :rtype: dict[str, Any]
+    """
+    init_params = inspect.signature(cls.__init__).parameters.keys()
+    return {k: v for k, v in init_dict.items() if k in init_params}
+
+
 def key_in_nested_dict(nested_dict: dict[str, Any], target: str) -> bool:
     """Determine if key is in nested dictionary.
 
@@ -584,6 +599,8 @@ def format_shared_critic_encoder(encoder_configs: NetConfigType) -> dict[str, An
         if encoder_key == "mlp_config":
             encoder_config[encoder_key] = config
             encoder_config["latent_dim"] = config.get("hidden_size", [32])[-1]
+            encoder_config["min_latent_dim"] = config.get("min_mlp_nodes", 8)
+            encoder_config["max_latent_dim"] = config.get("max_mlp_nodes", 1024)
         else:
             encoder_config["init_dicts"][encoder_key] = config
 

{agilerl-2.5.0.dev2 → agilerl-2.5.0.dev4}/configs/training/multi_agent/maddpg.yaml

@@ -71,7 +71,8 @@ MUTATION_PARAMS:
 
 NET_CONFIG:
     latent_dim: 64
-
+    min_latent_dim: 8
+    max_latent_dim: 1024
     encoder_config:
         hidden_size: [64]
         min_mlp_nodes: 64

{agilerl-2.5.0.dev2 → agilerl-2.5.0.dev4}/pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "agilerl"
 
-version = "2.5.0.dev2"
+version = "2.5.0.dev4"
 description = "AgileRL is a deep reinforcement learning library focused on improving RL development through RLOps."
 authors = [{ name = "Nick Ustaran-Anderegg", email = "dev@agilerl.com" }]
 license = "Apache-2.0"

{agilerl-2.5.0.dev2 → agilerl-2.5.0.dev4}/uv.lock

@@ -52,7 +52,7 @@ wheels = [
 
 [[package]]
 name = "agilerl"
-version = "2.5.0.dev3"
+version = "2.5.0.dev4"
 source = { editable = "." }
 dependencies = [
     { name = "accelerate" },