Add save inference

lerobot/common/policies/act/modeling_act.py

@@ -155,21 +155,29 @@ class ACTPolicy(PreTrainedPolicy):
         batch = self.normalize_targets(batch)
         actions_hat, (mu_hat, log_sigma_x2_hat) = self.model(batch)
 
-        l1_loss = (
-            F.l1_loss(batch["action"], actions_hat, reduction="none") * ~batch["action_is_pad"].unsqueeze(-1)
-        ).mean()
+        l1_loss = F.l1_loss(batch["action"], actions_hat, reduction="none")
+        l1_loss *= ~batch["action_is_pad"].unsqueeze(-1)
 
-        loss_dict = {"l1_loss": l1_loss.item()}
+        bsize, seqlen, num_motors = l1_loss.shape
+        loss_dict = {
+            "l1_loss": l1_loss.mean().item(),
+            "l1_loss_per_item": l1_loss.view(bsize, seqlen * num_motors).mean(dim=1),
+        }
         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 = (
-                (-0.5 * (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())).sum(-1).mean()
-            )
+            mean_kld = (-0.5 * (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())).sum(-1)
+            loss_dict["kld_loss_per_item"] = mean_kld
+
+            mean_kld = mean_kld.mean()
             loss_dict["kld_loss"] = mean_kld.item()
+
             loss = l1_loss + mean_kld * self.config.kl_weight
+            loss_dict["loss_per_item"] = (
+                loss_dict["l1_loss_per_item"] + loss_dict["kld_loss_per_item"] * self.config.kl_weight
+            )
         else:
             loss = l1_loss
 

lerobot/scripts/save_inference.py

@@ -0,0 +1,159 @@
+import logging
+import shutil
+import tempfile
+from contextlib import nullcontext
+from dataclasses import asdict, dataclass
+from pathlib import Path
+from pprint import pformat
+
+import torch
+import tqdm
+
+from lerobot.common.datasets.factory import make_dataset
+from lerobot.common.policies.factory import make_policy
+from lerobot.common.utils.utils import (
+    auto_select_torch_device,
+    init_logging,
+    is_amp_available,
+    is_torch_device_available,
+)
+from lerobot.configs import parser
+from lerobot.configs.default import DatasetConfig
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.train import TrainPipelineConfig
+
+
+@dataclass
+class SaveInferenceConfig:
+    dataset: DatasetConfig
+    # Delete the output directory if it exists already.
+    force_override: bool = False
+
+    batch_size: int = 16
+    num_workers: int = 4
+
+    policy: PreTrainedConfig | None = None
+    # TODO(rcadene, aliberts): By default, use device and use_amp values from policy checkpoint.
+    device: str | None = None  # cuda | cpu | mps
+    # `use_amp` determines whether to use Automatic Mixed Precision (AMP) for training and evaluation. With AMP,
+    # automatic gradient scaling is used.
+    use_amp: bool | None = None
+
+    output_dir: str | Path | None = None
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
+        return ["policy"]
+
+    def __post_init__(self):
+        # HACK: We parse again the cli args here to get the pretrained path if there was one.
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+
+            # When no device or use_amp are given, use the one from training config.
+            if self.device is None or self.use_amp is None:
+                train_cfg = TrainPipelineConfig.from_pretrained(policy_path)
+                if self.device is None:
+                    self.device = train_cfg.device
+                if self.use_amp is None:
+                    self.use_amp = train_cfg.use_amp
+
+            # Automatically switch to available device if necessary
+            if not is_torch_device_available(self.device):
+                auto_device = auto_select_torch_device()
+                logging.warning(f"Device '{self.device}' is not available. Switching to '{auto_device}'.")
+                self.device = auto_device
+
+            # Automatically deactivate AMP if necessary
+            if self.use_amp and not is_amp_available(self.device):
+                logging.warning(
+                    f"Automatic Mixed Precision (amp) is not available on device '{self.device}'. Deactivating AMP."
+                )
+                self.use_amp = False
+
+
+@parser.wrap()
+def save_inference(cfg: SaveInferenceConfig):
+    init_logging()
+    logging.info(pformat(asdict(cfg)))
+
+    dataset = make_dataset(cfg)
+    policy = make_policy(cfg.policy, cfg.device, ds_meta=dataset.meta)
+
+    output_dir = cfg.output_dir
+    if output_dir is None:
+        # Create a temporary directory that will be automatically cleaned up
+        output_dir = tempfile.mkdtemp(prefix="lerobot_save_inference_")
+
+    elif Path(output_dir).exists():
+        if cfg.force_override:
+            shutil.rmtree(cfg.output_dir)
+        else:
+            raise NotImplementedError(f"Output directory already exists: {cfg.output_dir}")
+
+    output_dir = Path(output_dir)
+
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=cfg.num_workers,
+        batch_size=cfg.batch_size,
+        pin_memory=cfg.device != "cpu",
+        drop_last=False,
+    )
+
+    policy.train()
+
+    episode_indices = []
+    frame_indices = []
+    feats = {}
+
+    for batch in tqdm.tqdm(dataloader):
+        for key in batch:
+            if isinstance(batch[key], torch.Tensor):
+                batch[key] = batch[key].to(cfg.device, non_blocking=True)
+
+        with torch.no_grad(), torch.autocast(device_type=cfg.device) if cfg.use_amp else nullcontext():
+            _, output_dict = policy.forward(batch)
+
+        bsize = batch["episode_index"].shape[0]
+        episode_indices.append(batch["episode_index"])
+        frame_indices.append(batch["frame_index"])
+
+        for key in output_dict:
+            if "loss_per_item" not in key:
+                continue
+
+            if key not in feats:
+                feats[key] = []
+
+            if not (output_dict[key].ndim == 1 and output_dict[key].shape[0] == bsize):
+                raise ValueError(output_dict[key].shape)
+
+            feats[key].append(output_dict[key])
+
+    episode_indices = torch.cat(episode_indices)
+    frame_indices = torch.cat(frame_indices)
+
+    for key in feats:
+        feats[key] = torch.cat(feats[key])
+
+    # Find unique episode indices
+    unique_episodes = torch.unique(episode_indices)
+
+    for episode in unique_episodes:
+        ep_feats = {}
+        for key in feats:
+            ep_feats[key] = feats[key][episode_indices == episode].data.cpu()
+
+        output_dir.mkdir(parents=True, exist_ok=True)
+        torch.save(ep_feats, output_dir / f"output_features_episode_{episode}.pth")
+
+    print(f"Features can be found in: {output_dir}")
+
+
+if __name__ == "__main__":
+    save_inference()

lerobot/scripts/visualize_dataset_html.py

@@ -65,6 +65,7 @@ from pathlib import Path
 import numpy as np
 import pandas as pd
 import requests
+import torch
 from flask import Flask, redirect, render_template, request, url_for
 
 from lerobot import available_datasets
@@ -80,6 +81,7 @@ def run_server(
     port: str,
     static_folder: Path,
     template_folder: Path,
+    inference_dir: Path | None,
 ):
     app = Flask(__name__, static_folder=static_folder.resolve(), template_folder=template_folder.resolve())
     app.config["SEND_FILE_MAX_AGE_DEFAULT"] = 0  # specifying not to cache
@@ -139,7 +141,14 @@ def run_server(
         )
 
     @app.route("/<string:dataset_namespace>/<string:dataset_name>/episode_<int:episode_id>")
-    def show_episode(dataset_namespace, dataset_name, episode_id, dataset=dataset, episodes=episodes):
+    def show_episode(
+        dataset_namespace,
+        dataset_name,
+        episode_id,
+        dataset=dataset,
+        episodes=episodes,
+        inference_dir=inference_dir,
+    ):
         repo_id = f"{dataset_namespace}/{dataset_name}"
         try:
             if dataset is None:
@@ -158,7 +167,7 @@ def run_server(
             if major_version < 2:
                 return "Make sure to convert your LeRobotDataset to v2 & above."
 
-        episode_data_csv_str, columns = get_episode_data(dataset, episode_id)
+        episode_data_csv_str, columns = get_episode_data(dataset, episode_id, inference_dir)
         dataset_info = {
             "repo_id": f"{dataset_namespace}/{dataset_name}",
             "num_samples": dataset.num_frames
@@ -228,7 +237,9 @@ def get_ep_csv_fname(episode_id: int):
     return ep_csv_fname
 
 
-def get_episode_data(dataset: LeRobotDataset | IterableNamespace, episode_index):
+def get_episode_data(
+    dataset: LeRobotDataset | IterableNamespace, episode_index, inference_dir: Path | None = None
+):
     """Get a csv str containing timeseries data of an episode (e.g. state and action).
     This file will be loaded by Dygraph javascript to plot data in real time."""
     columns = []
@@ -279,7 +290,15 @@ def get_episode_data(dataset: LeRobotDataset | IterableNamespace, episode_index)
             np.expand_dims(data["timestamp"], axis=1),
             *[np.vstack(data[col]) for col in selected_columns[1:]],
         )
-    ).tolist()
+    )
+
+    if inference_dir is not None:
+        feats = torch.load(inference_dir / f"output_features_episode_{episode_index}.pth")
+        for key in feats:
+            header.append(key.replace("loss_per_item", "loss"))
+            rows = np.concatenate([rows, feats[key][:, None]], axis=1)
+
+    rows = rows.tolist()
 
     # Convert data to CSV string
     csv_buffer = StringIO()
@@ -332,6 +351,7 @@ def visualize_dataset_html(
     host: str = "127.0.0.1",
     port: int = 9090,
     force_override: bool = False,
+    inference_dir: Path | None = None,
 ) -> Path | None:
     init_logging()
 
@@ -372,7 +392,7 @@ def visualize_dataset_html(
                 ln_videos_dir.symlink_to((dataset.root / "videos").resolve())
 
         if serve:
-            run_server(dataset, episodes, host, port, static_dir, template_dir)
+            run_server(dataset, episodes, host, port, static_dir, template_dir, inference_dir)
 
 
 def main():
@@ -439,6 +459,12 @@ def main():
         default=0,
         help="Delete the output directory if it exists already.",
     )
+    parser.add_argument(
+        "--inference-dir",
+        type=Path,
+        default=None,
+        help="",
+    )
 
     args = parser.parse_args()
     kwargs = vars(args)