Merge pull request #80 from huggingface/alexander-soare/unify_policy_api

Unify policy API

examples/3_train_policy.py

@@ -54,7 +54,7 @@ done = False
 while not done:
     for batch in dataloader:
         batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
-        info = policy(batch)
+        info = policy.update(batch)
         if step % log_freq == 0:
             print(f"step: {step} loss: {info['loss']:.3f} update_time: {info['update_s']:.3f} (seconds)")
         step += 1

lerobot/common/policies/act/configuration_act.py

@@ -103,12 +103,21 @@ class ActionChunkingTransformerConfig:
     def __post_init__(self):
         """Input validation (not exhaustive)."""
         if not self.vision_backbone.startswith("resnet"):
-            raise ValueError("`vision_backbone` must be one of the ResNet variants.")
+            raise ValueError(
+                f"`vision_backbone` must be one of the ResNet variants. Got {self.vision_backbone}."
+            )
         if self.use_temporal_aggregation:
             raise NotImplementedError("Temporal aggregation is not yet implemented.")
         if self.n_action_steps > self.chunk_size:
             raise ValueError(
-                "The chunk size is the upper bound for the number of action steps per model invocation."
+                f"The chunk size is the upper bound for the number of action steps per model invocation. Got "
+                f"{self.n_action_steps} for `n_action_steps` and {self.chunk_size} for `chunk_size`."
+            )
+        if self.n_obs_steps != 1:
+            raise ValueError(
+                f"Multiple observation steps not handled yet. Got `nobs_steps={self.n_obs_steps}`"
             )
         if self.camera_names != ["top"]:
-            raise ValueError("For now, `camera_names` can only be ['top']")
+            raise ValueError(f"For now, `camera_names` can only be ['top']. Got {self.camera_names}.")
+        if len(set(self.camera_names)) != len(self.camera_names):
+            raise ValueError(f"`camera_names` should not have any repeated entries. Got {self.camera_names}.")

lerobot/common/policies/act/modeling_act.py

@@ -61,9 +61,6 @@ class ActionChunkingTransformerPolicy(nn.Module):
     """
 
     name = "act"
-    _multiple_obs_steps_not_handled_msg = (
-        "ActionChunkingTransformerPolicy does not handle multiple observation steps."
-    )
 
     def __init__(self, cfg: ActionChunkingTransformerConfig | None = None):
         """
@@ -74,8 +71,6 @@ class ActionChunkingTransformerPolicy(nn.Module):
         super().__init__()
         if cfg is None:
             cfg = ActionChunkingTransformerConfig()
-        if cfg.n_obs_steps != 1:
-            raise ValueError(self._multiple_obs_steps_not_handled_msg)
         self.cfg = cfg
 
         # BERT style VAE encoder with input [cls, *joint_space_configuration, *action_sequence].
@@ -173,73 +168,47 @@ class ActionChunkingTransformerPolicy(nn.Module):
         environment. It works by managing the actions in a queue and only calling `select_actions` when the
         queue is empty.
         """
+        self.eval()
         if len(self._action_queue) == 0:
-            # `select_actions` returns a (batch_size, n_action_steps, *) tensor, but the queue effectively has
+            # `_forward` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue effectively
-            # shape (n_action_steps, batch_size, *), hence the transpose.
+            # has shape (n_action_steps, batch_size, *), hence the transpose.
-            self._action_queue.extend(self.select_actions(batch).transpose(0, 1))
+            self._action_queue.extend(self._forward(batch)[0][: self.cfg.n_action_steps].transpose(0, 1))
         return self._action_queue.popleft()
 
-    @torch.no_grad
+    def forward(self, batch, **_) -> dict[str, Tensor]:
-    def select_actions(self, batch: dict[str, Tensor]) -> Tensor:
+        """Run the batch through the model and compute the loss for training or validation."""
-        """Use the action chunking transformer to generate a sequence of actions."""
+        actions_hat, (mu_hat, log_sigma_x2_hat) = self._forward(batch)
-        self.eval()
-        self._preprocess_batch(batch, add_obs_steps_dim=True)
 
-        action = self.forward(batch, return_loss=False)
+        l1_loss = (
+            F.l1_loss(batch["action"], actions_hat, reduction="none") * ~batch["action_is_pad"].unsqueeze(-1)
+        ).mean()
 
-        return action[: self.cfg.n_action_steps]
+        loss_dict = {"l1_loss": l1_loss}
+        if self.cfg.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
+            # KL-divergence per batch element, then take the mean over the batch.
+            # (See App. B of https://arxiv.org/abs/1312.6114 for more details).
+            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["loss"] = l1_loss + mean_kld * self.cfg.kl_weight
+        else:
+            loss_dict["loss"] = l1_loss
 
-    def __call__(self, *args, **kwargs) -> dict:
+        return loss_dict
-        # TODO(alexander-soare): Temporary bridge until we know what to do about the `update` method.
-        return self.update(*args, **kwargs)
-
-    def _preprocess_batch(
-        self, batch: dict[str, Tensor], add_obs_steps_dim: bool = False
-    ) -> dict[str, Tensor]:
-        """
-        This function expects `batch` to have (at least):
-        {
-            "observation.state": (B, 1, J) OR (B, J) tensor of robot states (joint configuration).
-            "observation.images.top": (B, 1, C, H, W) OR (B, C, H, W) tensor of images.
-        }
-        """
-        if add_obs_steps_dim:
-            # Add a dimension for the observations steps. Since n_obs_steps > 1 is not supported right now,
-            # this just amounts to an unsqueeze.
-            for k in batch:
-                if k.startswith("observation."):
-                    batch[k] = batch[k].unsqueeze(1)
-
-        if batch["observation.state"].shape[1] != 1:
-            raise ValueError(self._multiple_obs_steps_not_handled_msg)
-        batch["observation.state"] = batch["observation.state"].squeeze(1)
-        # TODO(alexander-soare): generalize this to multiple images.
-        assert (
-            sum(k.startswith("observation.images.") and not k.endswith("is_pad") for k in batch) == 1
-        ), "ACT only handles one image for now."
-        # Note: no squeeze is required for "observation.images.top" because then we'd have to unsqueeze to get
-        # the image index dimension.
 
     def update(self, batch, **_) -> dict:
         """Run the model in train mode, compute the loss, and do an optimization step."""
         start_time = time.time()
-        self._preprocess_batch(batch)
-
         self.train()
-
+        loss_dict = self.forward(batch)
-        num_slices = self.cfg.batch_size
+        loss = loss_dict["loss"]
-        batch_size = self.cfg.chunk_size * num_slices
-
-        assert batch_size % self.cfg.chunk_size == 0
-        assert batch_size % num_slices == 0
-
-        loss = self.forward(batch, return_loss=True)["loss"]
         loss.backward()
 
         grad_norm = torch.nn.utils.clip_grad_norm_(
-            self.parameters(),
+            self.parameters(), self.cfg.grad_clip_norm, error_if_nonfinite=False
-            self.cfg.grad_clip_norm,
-            error_if_nonfinite=False,
         )
 
         self.optimizer.step()
@@ -254,68 +223,64 @@ class ActionChunkingTransformerPolicy(nn.Module):
 
         return info
 
-    def forward(self, batch: dict[str, Tensor], return_loss: bool = False) -> dict | Tensor:
+    def _stack_images(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        """A forward pass through the DNN part of this policy with optional loss computation."""
+        """Stacks all the images in a batch and puts them in a new key: "observation.images".
-        images = self.image_normalizer(batch["observation.images.top"])
 
-        if return_loss:  # training time
+        This function expects `batch` to have (at least):
-            actions_hat, (mu_hat, log_sigma_x2_hat) = self._forward(
+        {
-                batch["observation.state"], images, batch["action"]
+            "observation.state": (B, state_dim) batch of robot states.
-            )
+            "observation.images.{name}": (B, C, H, W) tensor of images.
-
+        }
-            l1_loss = (
-                F.l1_loss(batch["action"], actions_hat, reduction="none")
-                * ~batch["action_is_pad"].unsqueeze(-1)
-            ).mean()
-
-            loss_dict = {}
-            loss_dict["l1"] = l1_loss
-            if self.cfg.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
-                # KL-divergence per batch element, then take the mean over the batch.
-                # (See App. B of https://arxiv.org/abs/1312.6114 for more details).
-                mean_kld = (
-                    (-0.5 * (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())).sum(-1).mean()
-                )
-                loss_dict["kl"] = mean_kld
-                loss_dict["loss"] = loss_dict["l1"] + loss_dict["kl"] * self.cfg.kl_weight
-            else:
-                loss_dict["loss"] = loss_dict["l1"]
-            return loss_dict
-        else:
-            action, _ = self._forward(batch["observation.state"], images)
-            return action
-
-    def _forward(
-        self, robot_state: Tensor, image: Tensor, actions: Tensor | None = None
-    ) -> tuple[Tensor, tuple[Tensor | None, Tensor | None]]:
         """
-        Args:
+        # Check that there is only one image.
-            robot_state: (B, J) batch of robot joint configurations.
+        # TODO(alexander-soare): generalize this to multiple images.
-            image: (B, N, C, H, W) batch of N camera frames.
+        provided_cameras = {k.rsplit(".", 1)[-1] for k in batch if k.startswith("observation.images.")}
-            actions: (B, S, A) batch of actions from the target dataset which must be provided if the
+        if len(missing := set(self.cfg.camera_names).difference(provided_cameras)) > 0:
-                VAE is enabled and the model is in training mode.
+            raise ValueError(
+                f"The following camera images are missing from the provided batch: {missing}. Check the "
+                "configuration parameter: `camera_names`."
+            )
+        # Stack images in the order dictated by the camera names.
+        batch["observation.images"] = torch.stack(
+            [batch[f"observation.images.{name}"] for name in self.cfg.camera_names],
+            dim=-4,
+        )
+
+    def _forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, tuple[Tensor, Tensor] | tuple[None, None]]:
+        """A forward pass through the Action Chunking Transformer (with optional VAE encoder).
+
+        `batch` should have the following structure:
+
+        {
+            "observation.state": (B, state_dim) batch of robot states.
+            "observation.images": (B, n_cameras, C, H, W) batch of images.
+            "action" (optional, only if training with VAE): (B, chunk_size, action dim) batch of actions.
+        }
+
         Returns:
-            (B, S, A) batch of action sequences
+            (B, chunk_size, action_dim) batch of action sequences
             Tuple containing the latent PDF's parameters (mean, log(σ²)) both as (B, L) tensors where L is the
             latent dimension.
         """
         if self.cfg.use_vae and self.training:
             assert (
-                actions is not None
+                "action" in batch
             ), "actions must be provided when using the variational objective in training mode."
 
-        batch_size = robot_state.shape[0]
+        self._stack_images(batch)
+
+        batch_size = batch["observation.state"].shape[0]
 
         # Prepare the latent for input to the transformer encoder.
-        if self.cfg.use_vae and actions is not None:
+        if self.cfg.use_vae and "action" in batch:
             # Prepare the input to the VAE encoder: [cls, *joint_space_configuration, *action_sequence].
             cls_embed = einops.repeat(
                 self.vae_encoder_cls_embed.weight, "1 d -> b 1 d", b=batch_size
             )  # (B, 1, D)
-            robot_state_embed = self.vae_encoder_robot_state_input_proj(robot_state).unsqueeze(1)  # (B, 1, D)
+            robot_state_embed = self.vae_encoder_robot_state_input_proj(batch["observation.state"]).unsqueeze(
-            action_embed = self.vae_encoder_action_input_proj(actions)  # (B, S, D)
+                1
+            )  # (B, 1, D)
+            action_embed = self.vae_encoder_action_input_proj(batch["action"])  # (B, S, D)
             vae_encoder_input = torch.cat([cls_embed, robot_state_embed, action_embed], axis=1)  # (B, S+2, D)
 
             # Prepare fixed positional embedding.
@@ -337,15 +302,16 @@ class ActionChunkingTransformerPolicy(nn.Module):
             # When not using the VAE encoder, we set the latent to be all zeros.
             mu = log_sigma_x2 = None
             latent_sample = torch.zeros([batch_size, self.latent_dim], dtype=torch.float32).to(
-                robot_state.device
+                batch["observation.state"].device
             )
 
         # Prepare all other transformer encoder inputs.
         # Camera observation features and positional embeddings.
         all_cam_features = []
         all_cam_pos_embeds = []
-        for cam_id, _ in enumerate(self.cfg.camera_names):
+        images = self.image_normalizer(batch["observation.images"])
-            cam_features = self.backbone(image[:, cam_id])["feature_map"]
+        for cam_index in range(len(self.cfg.camera_names)):
+            cam_features = self.backbone(images[:, cam_index])["feature_map"]
             cam_pos_embed = self.encoder_cam_feat_pos_embed(cam_features).to(dtype=cam_features.dtype)
             cam_features = self.encoder_img_feat_input_proj(cam_features)  # (B, C, h, w)
             all_cam_features.append(cam_features)
@@ -355,7 +321,7 @@ class ActionChunkingTransformerPolicy(nn.Module):
         cam_pos_embed = torch.cat(all_cam_pos_embeds, axis=3)
 
         # Get positional embeddings for robot state and latent.
-        robot_state_embed = self.encoder_robot_state_input_proj(robot_state)
+        robot_state_embed = self.encoder_robot_state_input_proj(batch["observation.state"])
         latent_embed = self.encoder_latent_input_proj(latent_sample)
 
         # Stack encoder input and positional embeddings moving to (S, B, C).

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -43,12 +43,12 @@ class DiffusionPolicy(nn.Module):
     name = "diffusion"
 
     def __init__(self, cfg: DiffusionConfig | None, lr_scheduler_num_training_steps: int = 0):
-        super().__init__()
         """
         Args:
             cfg: Policy configuration class instance or None, in which case the default instantiation of the
                  configuration class is used.
         """
+        super().__init__()
         # TODO(alexander-soare): LR scheduler will be removed.
         assert lr_scheduler_num_training_steps > 0
         if cfg is None:
@@ -140,12 +140,18 @@ class DiffusionPolicy(nn.Module):
         action = self._queues["action"].popleft()
         return action
 
-    def forward(self, batch, **_):
+    def forward(self, batch: dict[str, Tensor], **_) -> dict[str, Tensor]:
+        """Run the batch through the model and compute the loss for training or validation."""
+        loss = self.diffusion.compute_loss(batch)
+        return {"loss": loss}
+
+    def update(self, batch: dict[str, Tensor], **_) -> dict:
+        """Run the model in train mode, compute the loss, and do an optimization step."""
         start_time = time.time()
 
         self.diffusion.train()
 
-        loss = self.diffusion.compute_loss(batch)
+        loss = self.forward(batch)["loss"]
         loss.backward()
 
         grad_norm = torch.nn.utils.clip_grad_norm_(

lerobot/common/policies/policy_protocol.py

@@ -0,0 +1,45 @@
+"""A protocol that all policies should follow.
+
+This provides a mechanism for type-hinting and isinstance checks without requiring the policies classes
+subclass a base class.
+
+The protocol structure, method signatures, and docstrings should be used by developers as a reference for
+how to implement new policies.
+"""
+
+from typing import Protocol, runtime_checkable
+
+from torch import Tensor
+
+
+@runtime_checkable
+class Policy(Protocol):
+    """The required interface for implementing a policy."""
+
+    name: str
+
+    def reset(self):
+        """To be called whenever the environment is reset.
+
+        Does things like clearing caches.
+        """
+
+    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.
+        """
+
+    def select_action(self, batch: dict[str, Tensor]):
+        """Return one action to run in the environment (potentially in batch mode).
+
+        When the model uses a history of observations, or outputs a sequence of actions, this method deals
+        with caching.
+        """
+
+    def update(self, batch):
+        """Does compute_loss then an optimization step.
+
+        TODO(alexander-soare): We will move the optimization step back into the training loop, so this will
+        disappear.
+        """

lerobot/common/policies/tdmpc/policy.py

@@ -330,6 +330,10 @@ class TDMPCPolicy(nn.Module):
         return td_target
 
     def forward(self, batch, step):
+        # TODO(alexander-soare): Refactor TDMPC and make it comply with the policy interface documentation.
+        raise NotImplementedError()
+
+    def update(self, batch, step):
         """Main update function. Corresponds to one iteration of the model learning."""
         start_time = time.time()
 

lerobot/configs/policy/act.yaml

@@ -67,6 +67,4 @@ policy:
   utd: 1
 
   delta_timestamps:
-    observation.images.top: "[i / ${fps} for i in range(1 - ${n_obs_steps}, 1)]"
-    observation.state: "[i / ${fps} for i in range(1 - ${n_obs_steps}, 1)]"
     action: "[i / ${fps} for i in range(${policy.chunk_size})]"

lerobot/scripts/train.py

@@ -257,7 +257,7 @@ def train(cfg: dict, out_dir=None, job_name=None):
         for key in batch:
             batch[key] = batch[key].to(cfg.device, non_blocking=True)
 
-        train_info = policy(batch, step=step)
+        train_info = policy.update(batch, step=step)
 
         # TODO(rcadene): is it ok if step_t=0 = 0 and not 1 as previously done?
         if step % cfg.log_freq == 0:
@@ -318,7 +318,7 @@ def train(cfg: dict, out_dir=None, job_name=None):
             for key in batch:
                 batch[key] = batch[key].to(cfg.device, non_blocking=True)
 
-            train_info = policy(batch, step)
+            train_info = policy.update(batch, step)
 
             if step % cfg.log_freq == 0:
                 log_train_info(logger, train_info, step, cfg, online_dataset, is_offline)

tests/test_policies.py

@@ -4,11 +4,13 @@ import torch
 from lerobot.common.datasets.utils import cycle
 from lerobot.common.envs.utils import postprocess_action, preprocess_observation
 from lerobot.common.policies.factory import make_policy
+from lerobot.common.policies.policy_protocol import Policy
 from lerobot.common.envs.factory import make_env
 from lerobot.common.datasets.factory import make_dataset
 from lerobot.common.utils import init_hydra_config
 from .utils import DEVICE, DEFAULT_CONFIG_PATH
 
+
 @pytest.mark.parametrize(
     "env_name,policy_name,extra_overrides",
     [
@@ -27,6 +29,7 @@ def test_policy(env_name, policy_name, extra_overrides):
     """
     Tests:
         - Making the policy object.
+        - Checking that the policy follows the correct protocol.
         - Updating the policy.
         - Using the policy to select actions at inference time.
         - Test the action can be applied to the policy
@@ -38,10 +41,14 @@ def test_policy(env_name, policy_name, extra_overrides):
             f"policy={policy_name}",
             f"device={DEVICE}",
         ]
-        + extra_overrides
+        + extra_overrides,
     )
     # Check that we can make the policy object.
     policy = make_policy(cfg)
+    # Check that the policy follows the required protocol.
+    assert isinstance(
+        policy, Policy
+    ), f"The policy does not follow the required protocol. Please see {Policy.__module__}.{Policy.__name__}."
     # Check that we run select_actions and get the appropriate output.
     dataset = make_dataset(cfg)
     env = make_env(cfg, num_parallel_envs=2)
@@ -62,7 +69,7 @@ def test_policy(env_name, policy_name, extra_overrides):
         batch[key] = batch[key].to(DEVICE, non_blocking=True)
 
     # Test updating the policy
-    policy(batch, step=0)
+    policy.update(batch, step=0)
 
     # reset the policy and environment
     policy.reset()
@@ -83,4 +90,3 @@ def test_policy(env_name, policy_name, extra_overrides):
 
     # Test step through policy
     env.step(action)
-