WIP

lerobot/common/policies/act/modeling_act.py

@@ -140,25 +140,25 @@ class ACTPolicy(
         batch = self.normalize_targets(batch)
         actions_hat, (mu_hat, log_sigma_x2_hat) = self.model(batch)
 
-        l1_loss = (
+        bsize = actions_hat.shape[0]
-            F.l1_loss(batch["action"], actions_hat, reduction="none") * ~batch["action_is_pad"].unsqueeze(-1)
+        l1_loss = F.l1_loss(batch["action"], actions_hat, reduction="none")
-        ).mean()
+        l1_loss = l1_loss * ~batch["action_is_pad"].unsqueeze(-1)
+        l1_loss = l1_loss.view(bsize, -1).mean(dim=1)
 
-        loss_dict = {"l1_loss": l1_loss.item()}
+        out_dict = {}
+        out_dict["l1_loss"] = l1_loss
         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
             # 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 = (
+            kld_loss = (-0.5 * (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())).sum(-1)
-                (-0.5 * (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())).sum(-1).mean()
+            out_dict["loss"] = l1_loss + kld_loss * self.config.kl_weight
-            )
-            loss_dict["kld_loss"] = mean_kld.item()
-            loss_dict["loss"] = l1_loss + mean_kld * self.config.kl_weight
         else:
-            loss_dict["loss"] = l1_loss
+            out_dict["loss"] = l1_loss
 
-        return loss_dict
+        out_dict["action"] = self.unnormalize_outputs({"action": actions_hat})["action"]
+        return out_dict
 
 
 class ACTTemporalEnsembler:

lerobot/scripts/visualize_dataset.py

@@ -268,10 +268,11 @@ def main():
     args = parser.parse_args()
     kwargs = vars(args)
     repo_id = kwargs.pop("repo_id")
-    root = kwargs.pop("root")
+    # root = kwargs.pop("root")
 
     logging.info("Loading dataset")
-    dataset = LeRobotDataset(repo_id, root=root, local_files_only=True)
+
+    dataset = LeRobotDataset(repo_id)
 
     visualize_dataset(dataset, **vars(args))
 

lerobot/scripts/visualize_dataset_html.py

@@ -55,13 +55,30 @@ python lerobot/scripts/visualize_dataset_html.py \
 import argparse
 import logging
 import shutil
+import warnings
 from pathlib import Path
 
+import torch
 import tqdm
 from flask import Flask, redirect, render_template, url_for
 
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.common.utils.utils import init_logging
+from lerobot.common.policies.factory import make_policy
+from lerobot.common.utils.utils import init_hydra_config, init_logging
+from lerobot.scripts.eval import get_pretrained_policy_path
+
+
+class EpisodeSampler(torch.utils.data.Sampler):
+    def __init__(self, dataset, episode_index):
+        from_idx = dataset.episode_data_index["from"][episode_index].item()
+        to_idx = dataset.episode_data_index["to"][episode_index].item()
+        self.frame_ids = range(from_idx, to_idx)
+
+    def __iter__(self):
+        return iter(self.frame_ids)
+
+    def __len__(self):
+        return len(self.frame_ids)
 
 
 def run_server(
@@ -119,14 +136,95 @@ def run_server(
     app.run(host=host, port=port)
 
 
+def run_inference(
+    dataset, episode_index, policy, policy_method="select_action", num_workers=4, batch_size=32, device="mps"
+):
+    if policy_method not in ["select_action", "forward"]:
+        raise ValueError(
+            f"`policy_method` is expected to be 'select_action' or 'forward', but '{policy_method}' is provided instead."
+        )
+
+    policy.eval()
+    policy.to(device)
+
+    logging.info("Loading dataloader")
+    episode_sampler = EpisodeSampler(dataset, episode_index)
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=num_workers,
+        # When using `select_action`, we set batch size 1 so that we feed 1 frame at a time, in a continuous fashion.
+        batch_size=1 if policy_method == "select_action" else batch_size,
+        sampler=episode_sampler,
+        drop_last=False,
+    )
+
+    warned_ndim_eq_0 = False
+    warned_ndim_gt_2 = False
+
+    logging.info("Running inference")
+    inference_results = {}
+    for batch in tqdm.tqdm(dataloader, total=len(dataloader)):
+        batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
+        with torch.inference_mode():
+            if policy_method == "select_action":
+                gt_action = batch.pop("action")
+                output_dict = {"action": policy.select_action(batch)}
+                batch["action"] = gt_action
+            elif policy_method == "forward":
+                output_dict = policy.forward(batch)
+                # TODO(rcadene): Save and display all predicted actions at a given timestamp
+                # Save predicted action for the next timestamp only
+                output_dict["action"] = output_dict["action"][:, 0, :]
+
+        for key in output_dict:
+            if output_dict[key].ndim == 0:
+                if not warned_ndim_eq_0:
+                    warnings.warn(
+                        f"Ignore output key '{key}'. Its value is a scalar instead of a vector. It might have been aggregated over the batch dimension (e.g. `loss.mean()`).",
+                        stacklevel=1,
+                    )
+                    warned_ndim_eq_0 = True
+                continue
+
+            if output_dict[key].ndim > 2:
+                if not warned_ndim_gt_2:
+                    warnings.warn(
+                        f"Ignore output key '{key}'. Its value is a tensor of {output_dict[key].ndim} dimensions instead of a vector.",
+                        stacklevel=1,
+                    )
+                    warned_ndim_gt_2 = True
+                continue
+
+            if key not in inference_results:
+                inference_results[key] = []
+            inference_results[key].append(output_dict[key].to("cpu"))
+
+    for key in inference_results:
+        inference_results[key] = torch.cat(inference_results[key])
+
+    return inference_results
+
+
 def get_ep_csv_fname(episode_id: int):
     ep_csv_fname = f"episode_{episode_id}.csv"
     return ep_csv_fname
 
 
-def write_episode_data_csv(output_dir, file_name, episode_index, dataset):
+def write_episode_data_csv(output_dir, file_name, episode_index, dataset, policy=None):
     """Write a csv file containg timeseries data of an episode (e.g. state and action).
     This file will be loaded by Dygraph javascript to plot data in real time."""
+
+    if policy is not None:
+        inference_results = run_inference(
+            dataset,
+            episode_index,
+            policy,
+            policy_method="select_action",
+            # num_workers=hydra_cfg.training.num_workers,
+            # batch_size=hydra_cfg.training.batch_size,
+            # device=hydra_cfg.device,
+        )
+
     from_idx = dataset.episode_data_index["from"][episode_index]
     to_idx = dataset.episode_data_index["to"][episode_index]
 
@@ -141,21 +239,26 @@ def write_episode_data_csv(output_dir, file_name, episode_index, dataset):
     if has_action:
         dim_action = dataset.meta.shapes["action"][0]
         header += [f"action_{i}" for i in range(dim_action)]
+    if policy is not None:
+        dim_action = dataset.meta.shapes["action"][0]
+        header += [f"pred_action_{i}" for i in range(dim_action)]
 
     columns = ["timestamp"]
     if has_state:
         columns += ["observation.state"]
     if has_action:
         columns += ["action"]
+    data = dataset.hf_dataset.select_columns(columns)
 
     rows = []
-    data = dataset.hf_dataset.select_columns(columns)
     for i in range(from_idx, to_idx):
         row = [data[i]["timestamp"].item()]
         if has_state:
             row += data[i]["observation.state"].tolist()
         if has_action:
             row += data[i]["action"].tolist()
+        if policy is not None:
+            row += inference_results["action"][i].tolist()
         rows.append(row)
 
     output_dir.mkdir(parents=True, exist_ok=True)
@@ -183,6 +286,9 @@ def visualize_dataset_html(
     host: str = "127.0.0.1",
     port: int = 9090,
     force_override: bool = False,
+    policy_method: str = "select_action",
+    pretrained_policy_name_or_path: str | None = None,
+    policy_overrides: list[str] | None = None,
 ) -> Path | None:
     init_logging()
 
@@ -214,12 +320,28 @@ def visualize_dataset_html(
     if episodes is None:
         episodes = list(range(dataset.num_episodes))
 
+    pretrained_policy_name_or_path = "aliberts/act_reachy_test_model"
+
+    policy = None
+    if pretrained_policy_name_or_path is not None:
+        logging.info("Loading policy")
+        pretrained_policy_path = get_pretrained_policy_path(pretrained_policy_name_or_path)
+
+        hydra_cfg = init_hydra_config(pretrained_policy_path / "config.yaml", overrides=["device=mps"])
+        # dataset = make_dataset(hydra_cfg)
+        policy = make_policy(hydra_cfg, pretrained_policy_name_or_path=pretrained_policy_path)
+
+        if policy_method == "select_action":
+            # Do not load previous observations or future actions, to simulate that the observations come from
+            # an environment.
+            dataset.delta_timestamps = None
+
     logging.info("Writing CSV files")
     for episode_index in tqdm.tqdm(episodes):
         # write states and actions in a csv (it can be slow for big datasets)
         ep_csv_fname = get_ep_csv_fname(episode_index)
         # TODO(rcadene): speedup script by loading directly from dataset, pyarrow, parquet, safetensors?
-        write_episode_data_csv(static_dir, ep_csv_fname, episode_index, dataset)
+        write_episode_data_csv(static_dir, ep_csv_fname, episode_index, dataset, policy=policy)
 
     if serve:
         run_server(dataset, episodes, host, port, static_dir, template_dir)
@@ -281,8 +403,8 @@ def main():
     args = parser.parse_args()
     kwargs = vars(args)
     repo_id = kwargs.pop("repo_id")
-    root = kwargs.pop("root")
+    # root = kwargs.pop("root")
-    dataset = LeRobotDataset(repo_id, root=root, local_files_only=True)
+    dataset = LeRobotDataset(repo_id)
     visualize_dataset_html(dataset, **kwargs)