Enable logging all the information returned by the forward methods of policies (#151)

lerobot/common/logger.py

@@ -114,6 +114,11 @@ class Logger:
         assert mode in {"train", "eval"}
         if self._wandb is not None:
             for k, v in d.items():
+                if not isinstance(v, (int, float, str)):
+                    logging.warning(
+                        f'WandB logging of key "{k}" was ignored as its type is not handled by this wrapper.'
+                    )
+                    continue
                 self._wandb.log({f"{mode}/{k}": v}, step=step)
 
     def log_video(self, video_path: str, step: int, mode: str = "train"):

lerobot/common/policies/act/modeling_act.py

@@ -101,7 +101,7 @@ class ACTPolicy(nn.Module, PyTorchModelHubMixin):
             F.l1_loss(batch["action"], actions_hat, reduction="none") * ~batch["action_is_pad"].unsqueeze(-1)
         ).mean()
 
-        loss_dict = {"l1_loss": l1_loss}
+        loss_dict = {"l1_loss": l1_loss.item()}
         if self.config.use_vae:
             # Calculate Dₖₗ(latent_pdf || standard_normal). Note: After computing the KL-divergence for
             # each dimension independently, we sum over the latent dimension to get the total
@@ -110,7 +110,7 @@ class ACTPolicy(nn.Module, PyTorchModelHubMixin):
             mean_kld = (
                 (-0.5 * (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())).sum(-1).mean()
             )
-            loss_dict["kld_loss"] = mean_kld
+            loss_dict["kld_loss"] = mean_kld.item()
             loss_dict["loss"] = l1_loss + mean_kld * self.config.kl_weight
         else:
             loss_dict["loss"] = l1_loss

lerobot/common/policies/policy_protocol.py

@@ -38,7 +38,8 @@ class Policy(Protocol):
     def forward(self, batch: dict[str, Tensor]) -> dict:
         """Run the batch through the model and compute the loss for training or validation.
 
-        Returns a dictionary with "loss" and maybe other information.
+        Returns a dictionary with "loss" and potentially other information. Apart from "loss" which is a Tensor, all
+        other items should be logging-friendly, native Python types.
         """
 
     def select_action(self, batch: dict[str, Tensor]):