feat(policies): convert save_policy_to_safetensors with pipeline

tests/artifacts/policies/save_policy_to_safetensors.py

@@ -23,7 +23,8 @@ from lerobot.configs.default import DatasetConfig
 from lerobot.configs.train import TrainPipelineConfig
 from lerobot.datasets.factory import make_dataset
 from lerobot.optim.factory import make_optimizer_and_scheduler
-from lerobot.policies.factory import make_policy, make_policy_config
+from lerobot.policies.factory import make_policy, make_policy_config, make_processor
+from lerobot.processor import TransitionKey
 from lerobot.utils.random_utils import set_seed
 
 
@@ -37,7 +38,9 @@ def get_policy_stats(ds_repo_id: str, policy_name: str, policy_kwargs: dict):
     train_cfg.validate()  # Needed for auto-setting some parameters
 
     dataset = make_dataset(train_cfg)
+    dataset_stats = dataset.meta.stats
     policy = make_policy(train_cfg.policy, ds_meta=dataset.meta)
+    preprocessor, postprocessor = make_processor(train_cfg.policy, dataset_stats=dataset_stats)
     policy.train()
 
     optimizer, _ = make_optimizer_and_scheduler(train_cfg, policy)
@@ -49,7 +52,9 @@ def get_policy_stats(ds_repo_id: str, policy_name: str, policy_kwargs: dict):
     )
 
     batch = next(iter(dataloader))
+    batch = preprocessor(batch)
     loss, output_dict = policy.forward(batch)
+
     if output_dict is not None:
         output_dict = {k: v for k, v in output_dict.items() if isinstance(v, torch.Tensor)}
         output_dict["loss"] = loss
@@ -96,7 +101,12 @@ def get_policy_stats(ds_repo_id: str, policy_name: str, policy_kwargs: dict):
     else:
         actions_queue = train_cfg.policy.n_action_repeats
 
-    actions = {str(i): policy.select_action(obs).contiguous() for i in range(actions_queue)}
+    actions = {}
+    for i in range(actions_queue):
+        unnormalized_action = policy.select_action(obs).contiguous()
+        action_robot = postprocessor({TransitionKey.ACTION: unnormalized_action}).get(TransitionKey.ACTION)
+        actions[str(i)] = action_robot
+
     return output_dict, grad_stats, param_stats, actions