backup wip

lerobot/common/logger.py

@@ -2,6 +2,7 @@ import logging
 import os
 from pathlib import Path
 
+from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
 from omegaconf import OmegaConf
 from termcolor import colored
 
@@ -70,20 +71,27 @@ class Logger:
     def save_model(self, policy, identifier):
         if self._save_model:
             self._model_dir.mkdir(parents=True, exist_ok=True)
-            fp = self._model_dir / f"{str(identifier)}.pt"
             # TODO(alexander-soare): This conditional branching is temporary while we add PyTorchModelHubMixin
-            # to all policies.
+            # to all policies. Once we're done, we should only use policy.save_pretrained.
             if hasattr(policy, "save"):
-                policy.save(fp)
+                model_path = self._model_dir / f"{str(identifier)}.pt"
+                policy.save(model_path)
             else:
-                policy.save_pretrained(fp)
+                save_dir = self._model_dir / str(identifier)
+                policy.save_pretrained(save_dir)
+                model_path = save_dir / SAFETENSORS_SINGLE_FILE
             if self._wandb and not self._disable_wandb_artifact:
                 # note wandb artifact does not accept ":" in its name
                 artifact = self._wandb.Artifact(
                     self._group.replace(":", "_") + "-" + str(self._seed) + "-" + str(identifier),
                     type="model",
                 )
-                artifact.add_file(fp)
+                # TODO(alexander-soare). See todo above. This conditional branching is temporary and we should
+                # only have the else path.
+                if hasattr(policy, "save"):
+                    artifact.add_file(model_path)
+                else:
+                    artifact.add_dir(save_dir)
                 self._wandb.log_artifact(artifact)
 
     def save_buffer(self, buffer, identifier):

lerobot/common/policies/act/modeling_act.py

@@ -24,7 +24,7 @@ from torchvision.ops.misc import FrozenBatchNorm2d
 from lerobot.common.policies.act.configuration_act import ActionChunkingTransformerConfig
 
 
-class ActPActionChunkingTransformerPolicyolicy(nn.Module, PyTorchModelHubMixin):
+class ActionChunkingTransformerPolicy(nn.Module, PyTorchModelHubMixin):
     """
     Action Chunking Transformer Policy as per Learning Fine-Grained Bimanual Manipulation with Low-Cost
     Hardware (paper: https://arxiv.org/abs/2304.13705, code: https://github.com/tonyzhaozh/act)
@@ -390,9 +390,6 @@ class ActPActionChunkingTransformerPolicyolicy(nn.Module, PyTorchModelHubMixin):
 
         return actions, (mu, log_sigma_x2)
 
-    def save(self, fp):
-        torch.save(self.state_dict(), fp)
-
     def load(self, fp):
         d = torch.load(fp)
         self.load_state_dict(d)

lerobot/common/policies/factory.py

@@ -1,11 +1,11 @@
 import inspect
 
-from omegaconf import OmegaConf
+from omegaconf import DictConfig, OmegaConf
 
 from lerobot.common.utils import get_safe_torch_device
 
 
-def make_policy(cfg):
+def make_policy(cfg: DictConfig):
     if cfg.policy.name == "tdmpc":
         from lerobot.common.policies.tdmpc.policy import TDMPCPolicy
 

lerobot/scripts/eval.py

@@ -42,6 +42,7 @@ import imageio
 import numpy as np
 import torch
 from huggingface_hub import snapshot_download
+from omegaconf import DictConfig
 
 from lerobot.common.datasets.factory import make_dataset
 from lerobot.common.envs.factory import make_env
@@ -290,7 +291,7 @@ def eval_policy(
     return info
 
 
-def eval(cfg: dict, out_dir=None, stats_path=None):
+def eval(cfg: DictConfig, out_dir=None, stats_path=None):
     if out_dir is None:
         raise NotImplementedError()