updating with adding masking in ACT - start adding some tests

examples/real_robot_example/README.md

@@ -1,5 +1,6 @@
 # Using `lerobot`  on a real world arm
 
+
 In this example, we'll be using `lerobot` on a real world arm to:
 - record a dataset in the `lerobot` format
 - (soon) train a policy on it
@@ -25,7 +26,9 @@ Follow these steps:
 - install `lerobot`
 - install the Dynamixel-sdk: ` pip install dynamixel-sdk`
 
-## 0 - record examples
+## Usage
+
+### 0 - record examples
 
 Run the `record_training_data.py` example, selecting the duration and number of episodes you want to record, e.g.
 ```
@@ -40,7 +43,7 @@ TODO:
 - being able to drop episodes
 - checking uploading to the hub
 
-## 1 - visualize the dataset
+### 1 - visualize the dataset
 
 Use the standard dataset visualization script pointing it to the right folder:
 ```
@@ -49,7 +52,7 @@ DATA_DIR='./data' python ../../lerobot/scripts/visualize_dataset.py \
     --episode-index 0
 ```
 
-## 2 - Train a policy
+### 2 - Train a policy
 
 From the example directory let's run this command to train a model using ACT
 
@@ -64,7 +67,7 @@ DATA_DIR='./data' python ../../lerobot/scripts/train.py \
     wandb.enable=false
 ```
 
-## 3 - Evaluate the policy in the real world
+### 3 - Evaluate the policy in the real world
 
 From the example directory let's run this command to evaluate our policy.
 The configuration for running the policy is in the checkpoint of the model.
@@ -75,3 +78,12 @@ python run_policy.py \
     -p ./outputs/train/blue_red_sort/checkpoints/last/pretrained_model/
     env.episode_length=1000
 ```
+
+
+## Convert a hdf5 dataset recorded with the original ACT repo
+
+You can convert a dataset from the raw data format of HDF5 files like in: https://github.com/tonyzhaozh/act with the following command:
+
+```
+python ./lerobot/scripts/push_dataset_to_hub.py
+```

examples/real_robot_example/gym_real_world/gym_environment.py

@@ -1,4 +1,5 @@
 import time
+from unittest.mock import MagicMock
 
 import cv2
 import gymnasium as gym
@@ -23,6 +24,14 @@ CAMERAS_PORTS = {
 LEADER_PORT = "/dev/ttyACM1"
 FOLLOWER_PORT = "/dev/ttyACM0"
 
+MockRobot = MagicMock()
+MockRobot.read_position = MagicMock()
+MockRobot.read_position.return_value = np.array([0.0, 1.0, 2.0, 3.0, 4.0, 5.0])
+
+MockCamera = MagicMock()
+MockCamera.isOpened = MagicMock(return_value=True)
+MockCamera.read = MagicMock(return_value=(True, np.zeros((480, 640, 3), dtype=np.uint8)))
+
 
 def capture_image(cam, cam_width, cam_height):
     # Capture a single frame
@@ -54,6 +63,7 @@ class RealEnv(gym.Env):
         trigger_torque=70,
         fps: int = FPS,
         fps_tolerance: float = 0.1,
+        mock: bool = False,
     ):
         self.num_joints = num_joints
         self.cameras_shapes = cameras_shapes
@@ -68,15 +78,15 @@ class RealEnv(gym.Env):
         self.fps_tolerance = fps_tolerance
 
         # Initialize the robot
-        self.follower = Robot(device_name=self.follower_port)
+        self.follower = Robot(device_name=self.follower_port) if not mock else MockRobot
         if self.record:
-            self.leader = Robot(device_name=self.leader_port)
+            self.leader = Robot(device_name=self.leader_port) if not mock else MockRobot
             self.leader.set_trigger_torque(trigger_torque)
 
         # Initialize the cameras - sorted by camera names
         self.cameras = {}
         for cn, p in sorted(self.cameras_ports.items()):
-            self.cameras[cn] = cv2.VideoCapture(p)
+            self.cameras[cn] = cv2.VideoCapture(p) if not mock else MockCamera
             if not self.cameras[cn].isOpened():
                 raise OSError(
                     f"Cannot open camera port {p} for {cn}."
@@ -118,7 +128,6 @@ class RealEnv(gym.Env):
 
         self._observation = {}
         self._terminated = False
-        self.starting_time = time.time()
         self.timestamps = []
 
     def _get_obs(self):
@@ -146,13 +155,8 @@ class RealEnv(gym.Env):
         if self.timestamps:
             # wait the right amount of time to stay at the desired fps
             time.sleep(max(0, 1 / self.fps - (time.time() - self.timestamps[-1])))
-            recording_time = time.time() - self.starting_time
-        else:
-            # it's the first step so we start the timer
-            self.starting_time = time.time()
-            recording_time = 0
 
-        self.timestamps.append(recording_time)
+        self.timestamps.append(time.time())
 
         # Get the observation
         self._get_obs()
@@ -165,13 +169,15 @@ class RealEnv(gym.Env):
 
         reward = 0
         terminated = truncated = self._terminated
-        info = {"timestamp": recording_time, "fps_error": False}
+        info = {"timestamp": self.timestamps[-1] - self.timestamps[0], "fps_error": False}
 
         # Check if we are able to keep up with the desired fps
-        if recording_time - self.timestamps[-1] > 1 / (self.fps - self.fps_tolerance):
+        if len(self.timestamps) > 1 and (self.timestamps[-1] - self.timestamps[-2]) > 1 / (
+            self.fps - self.fps_tolerance
+        ):
             print(
-                f"Error: recording time interval {recording_time - self.timestamps[-1]:.2f} is greater"
+                f"Error: recording fps {1 / (self.timestamps[-1] - self.timestamps[-2]):.5f} is lower"
-                f"than expected {1 / (self.fps - self.fps_tolerance):.2f}"
+                f" than min admited fps {(self.fps - self.fps_tolerance):.5f}"
                 f" at frame {len(self.timestamps)}"
             )
             info["fps_error"] = True

examples/real_robot_example/record_training_data.py

@@ -6,12 +6,14 @@ using a very simple gym environment (see in examples/real_robot_example/gym_real
 import argparse
 import copy
 import os
+from pathlib import Path
 
 import gym_real_world  # noqa: F401
 import gymnasium as gym
 import numpy as np
 import torch
 from datasets import Dataset, Features, Sequence, Value
+from omegaconf import OmegaConf
 from tqdm import tqdm
 
 from lerobot.common.datasets.compute_stats import compute_stats
@@ -30,17 +32,20 @@ parser.add_argument("--num-episodes", type=int, default=2)
 parser.add_argument("--num-frames", type=int, default=400)
 parser.add_argument("--num-workers", type=int, default=16)
 parser.add_argument("--keep-last", action="store_true")
+parser.add_argument("--data_dir", type=str, default=None)
 parser.add_argument("--push-to-hub", action="store_true")
 parser.add_argument("--fps", type=int, default=30, help="Frames per second of the recording.")
 parser.add_argument(
     "--fps_tolerance",
     type=float,
-    default=0.1,
+    default=0.5,
     help="Tolerance in fps for the recording before dropping episodes.",
 )
 parser.add_argument(
     "--revision", type=str, default=CODEBASE_VERSION, help="Codebase version used to generate the dataset."
 )
+parser.add_argument("--gym-config", type=str, default=None, help="Path to the gym config file.")
+parser.add_argument("--mock_robot", action="store_true")
 args = parser.parse_args()
 
 repo_id = args.repo_id
@@ -50,7 +55,7 @@ revision = args.revision
 fps = args.fps
 fps_tolerance = args.fps_tolerance
 
-out_data = DATA_DIR / repo_id
+out_data = DATA_DIR / repo_id if args.data_dir is None else Path(args.data_dir)
 
 # During data collection, frames are stored as png images in `images_dir`
 images_dir = out_data / "images"
@@ -58,6 +63,9 @@ images_dir = out_data / "images"
 videos_dir = out_data / "videos"
 meta_data_dir = out_data / "meta_data"
 
+gym_config = None
+if args.config is not None:
+    gym_config = OmegaConf.load(args.config)
 
 # Create image and video directories
 if not os.path.exists(images_dir):
@@ -68,7 +76,12 @@ if not os.path.exists(videos_dir):
 if __name__ == "__main__":
     # Create the gym environment - check the kwargs in gym_real_world/gym_environment.py
     gym_handle = "gym_real_world/RealEnv-v0"
-    env = gym.make(gym_handle, disable_env_checker=True, record=True, fps=fps, fps_tolerance=fps_tolerance)
+    gym_kwargs = {}
+    if gym_config is not None:
+        gym_kwargs = OmegaConf.to_container(gym_config.gym_kwargs)
+    env = gym.make(
+        gym_handle, disable_env_checker=True, record=True, fps=fps, fps_tolerance=fps_tolerance, mock=True
+    )
 
     ep_dicts = []
     episode_data_index = {"from": [], "to": []}

examples/real_robot_example/train_config/env/gym_real_world.yaml

@@ -10,3 +10,10 @@ env:
   fps: ${fps}
   episode_length: 200
   real_world: true
+  gym:
+    cameras_shapes:
+        images.high: [480, 640, 3]
+        images.low: [480, 640, 3]
+    cameras_ports:
+        images.high: /dev/video6
+        images.low: /dev/video0

examples/real_robot_example/train_config/env/gym_real_world_debug.yaml

@@ -0,0 +1,19 @@
+# @package _global_
+
+fps: 30
+
+env:
+  name: real_world
+  task: RealEnv-v0
+  state_dim: 6
+  action_dim: 6
+  fps: ${fps}
+  episode_length: 200
+  real_world: true
+  gym:
+    cameras_shapes:
+        images.top: [480, 640, 3]
+        images.front: [480, 640, 3]
+    cameras_ports:
+        images.top: /dev/video6
+        images.front: /dev/video0

examples/real_robot_example/train_config/policy/act_real_world_debug.yaml

@@ -0,0 +1,103 @@
+# @package _global_
+
+# Use `act_real.yaml` to train on real-world Aloha/Aloha2 datasets.
+# Compared to `act.yaml`, it contains 4 cameras (i.e. right_wrist, left_wrist, images,
+# front) instead of 1 camera (i.e. top). Also, `training.eval_freq` is set to -1. This config is used
+# to evaluate checkpoints at a certain frequency of training steps. When it is set to -1, it deactivates evaluation.
+# This is because real-world evaluation is done through [dora-lerobot](https://github.com/dora-rs/dora-lerobot).
+# Look at its README for more information on how to evaluate a checkpoint in the real-world.
+#
+# Example of usage for training:
+# ```bash
+# python lerobot/scripts/train.py \
+#   policy=act_real \
+#   env=aloha_real
+# ```
+
+seed: 1000
+dataset_repo_id: ???
+
+override_dataset_stats:
+  observation.images.top:
+    # stats from imagenet, since we use a pretrained vision model
+    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
+    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
+  observation.images.front:
+    # stats from imagenet, since we use a pretrained vision model
+    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
+    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
+
+training:
+  offline_steps: 1000
+  online_steps: 0
+  eval_freq: -1
+  save_freq: 1000
+  log_freq: 100
+  save_checkpoint: true
+
+  batch_size: 8
+  lr: 1e-5
+  lr_backbone: 1e-5
+  weight_decay: 1e-4
+  grad_clip_norm: 10
+  online_steps_between_rollouts: 1
+
+  delta_timestamps:
+    action: "[i / ${fps} for i in range(1, ${policy.chunk_size} + 1)]"
+
+eval:
+  n_episodes: 1
+  batch_size: 1
+
+# See `configuration_act.py` for more details.
+policy:
+  name: act
+
+  # Input / output structure.
+  n_obs_steps: 1
+  chunk_size: 100 # chunk_size
+  n_action_steps: 100
+
+  input_shapes:
+    # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
+    observation.images.top: [3, 480, 640]
+    observation.images.front: [3, 480, 640]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes:
+    observation.images.top: mean_std
+    observation.images.front: mean_std
+    observation.state: mean_std
+  output_normalization_modes:
+    action: mean_std
+
+  # Architecture.
+  # Vision backbone.
+  vision_backbone: resnet18
+  pretrained_backbone_weights: ResNet18_Weights.IMAGENET1K_V1
+  replace_final_stride_with_dilation: false
+  # Transformer layers.
+  pre_norm: false
+  dim_model: 512
+  n_heads: 8
+  dim_feedforward: 3200
+  feedforward_activation: relu
+  n_encoder_layers: 4
+  # Note: Although the original ACT implementation has 7 for `n_decoder_layers`, there is a bug in the code
+  # that means only the first layer is used. Here we match the original implementation by setting this to 1.
+  # See this issue https://github.com/tonyzhaozh/act/issues/25#issue-2258740521.
+  n_decoder_layers: 1
+  # VAE.
+  use_vae: true
+  latent_dim: 32
+  n_vae_encoder_layers: 4
+
+  # Inference.
+  temporal_ensemble_momentum: null
+
+  # Training and loss computation.
+  dropout: 0.1
+  kl_weight: 10.0

lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py

@@ -43,9 +43,6 @@ def get_cameras(hdf5_data):
 
 
 def check_format(raw_dir) -> bool:
-    # only frames from simulation are uncompressed
-    compressed_images = "sim" not in raw_dir.name
-
     hdf5_paths = list(raw_dir.glob("episode_*.hdf5"))
     assert len(hdf5_paths) != 0
     for hdf5_path in hdf5_paths:
@@ -62,17 +59,15 @@ def check_format(raw_dir) -> bool:
             for camera in get_cameras(data):
                 assert num_frames == data[f"/observations/images/{camera}"].shape[0]
 
-                if compressed_images:
+                # ndim 2 when image are compressed and 4 when uncompressed
-                    assert data[f"/observations/images/{camera}"].ndim == 2
+                assert data[f"/observations/images/{camera}"].ndim in [2, 4]
-                else:
+                if data[f"/observations/images/{camera}"].ndim == 4:
-                    assert data[f"/observations/images/{camera}"].ndim == 4
                     b, h, w, c = data[f"/observations/images/{camera}"].shape
                     assert c < h and c < w, f"Expect (h,w,c) image format but ({h=},{w=},{c=}) provided."
 
 
 def load_from_raw(raw_dir, out_dir, fps, video, debug):
     # only frames from simulation are uncompressed
-    compressed_images = "sim" not in raw_dir.name
 
     hdf5_files = list(raw_dir.glob("*.hdf5"))
     ep_dicts = []
@@ -99,7 +94,7 @@ def load_from_raw(raw_dir, out_dir, fps, video, debug):
             for camera in get_cameras(ep):
                 img_key = f"observation.images.{camera}"
 
-                if compressed_images:
+                if ep[f"/observations/images/{camera}"].ndim == 2:
                     import cv2
 
                     # load one compressed image after the other in RAM and uncompress

lerobot/common/policies/act/configuration_act.py

@@ -129,7 +129,9 @@ class ACTConfig:
     # Note: Although the original ACT implementation has 7 for `n_decoder_layers`, there is a bug in the code
     # that means only the first layer is used. Here we match the original implementation by setting this to 1.
     # See this issue https://github.com/tonyzhaozh/act/issues/25#issue-2258740521.
+    # As a consequence we also remove the final, unused layer normalization, by default
     n_decoder_layers: int = 1
+    decoder_norm: bool = False
     # VAE.
     use_vae: bool = True
     latent_dim: int = 32

lerobot/common/policies/act/modeling_act.py

@@ -315,8 +315,14 @@ class ACT(nn.Module):
             pos_embed = self.vae_encoder_pos_enc.clone().detach()  # (1, S+2, D)
 
             # Forward pass through VAE encoder to get the latent PDF parameters.
+            cls_joint_is_pad = torch.full((batch_size, 2), False).to(
+                batch["observation.state"].device
+            )  # False: not a padding
+            key_padding_mask = torch.cat([cls_joint_is_pad, batch["action_is_pad"]], axis=1)  # (bs, seq+1)
             cls_token_out = self.vae_encoder(
-                vae_encoder_input.permute(1, 0, 2), pos_embed=pos_embed.permute(1, 0, 2)
+                vae_encoder_input.permute(1, 0, 2),
+                pos_embed=pos_embed.permute(1, 0, 2),
+                key_padding_mask=key_padding_mask,
             )[0]  # select the class token, with shape (B, D)
             latent_pdf_params = self.vae_encoder_latent_output_proj(cls_token_out)
             mu = latent_pdf_params[:, : self.config.latent_dim]
@@ -402,9 +408,11 @@ class ACTEncoder(nn.Module):
         self.layers = nn.ModuleList([ACTEncoderLayer(config) for _ in range(config.n_encoder_layers)])
         self.norm = nn.LayerNorm(config.dim_model) if config.pre_norm else nn.Identity()
 
-    def forward(self, x: Tensor, pos_embed: Tensor | None = None) -> Tensor:
+    def forward(
+        self, x: Tensor, pos_embed: Tensor | None = None, key_padding_mask: Tensor | None = None
+    ) -> Tensor:
         for layer in self.layers:
-            x = layer(x, pos_embed=pos_embed)
+            x = layer(x, pos_embed=pos_embed, key_padding_mask=key_padding_mask)
         x = self.norm(x)
         return x
 
@@ -427,12 +435,14 @@ class ACTEncoderLayer(nn.Module):
         self.activation = get_activation_fn(config.feedforward_activation)
         self.pre_norm = config.pre_norm
 
-    def forward(self, x, pos_embed: Tensor | None = None) -> Tensor:
+    def forward(self, x, pos_embed: Tensor | None = None, key_padding_mask: Tensor | None = None) -> Tensor:
         skip = x
         if self.pre_norm:
             x = self.norm1(x)
         q = k = x if pos_embed is None else x + pos_embed
-        x = self.self_attn(q, k, value=x)[0]  # select just the output, not the attention weights
+        x = self.self_attn(q, k, value=x, key_padding_mask=key_padding_mask)[
+            0
+        ]  # select just the output, not the attention weights
         x = skip + self.dropout1(x)
         if self.pre_norm:
             skip = x
@@ -452,7 +462,10 @@ class ACTDecoder(nn.Module):
         """Convenience module for running multiple decoder layers followed by normalization."""
         super().__init__()
         self.layers = nn.ModuleList([ACTDecoderLayer(config) for _ in range(config.n_decoder_layers)])
-        self.norm = nn.LayerNorm(config.dim_model)
+        if config.decoder_norm:
+            self.norm = nn.LayerNorm(config.dim_model)
+        else:
+            self.norm = nn.Identity()
 
     def forward(
         self,
@@ -465,8 +478,7 @@ class ACTDecoder(nn.Module):
             x = layer(
                 x, encoder_out, decoder_pos_embed=decoder_pos_embed, encoder_pos_embed=encoder_pos_embed
             )
-        if self.norm is not None:
+        x = self.norm(x)
-            x = self.norm(x)
         return x
 
 

lerobot/configs/default.yaml

@@ -50,6 +50,8 @@ eval:
   batch_size: 1
   # `use_async_envs` specifies whether to use asynchronous environments (multiprocessing).
   use_async_envs: false
+  # Specify the number of episodes to render during evaluation.
+  max_episodes_rendered: 10
 
 wandb:
   enable: false

lerobot/scripts/eval.py

@@ -44,6 +44,7 @@ https://huggingface.co/lerobot/diffusion_pusht/tree/main.
 import argparse
 import json
 import logging
+import os
 import threading
 import time
 from contextlib import nullcontext
@@ -164,7 +165,10 @@ def rollout(
         # VectorEnv stores is_success in `info["final_info"][env_index]["is_success"]`. "final_info" isn't
         # available of none of the envs finished.
         if "final_info" in info:
-            successes = [i["is_success"] if i is not None else False for i in info["final_info"]]
+            successes = [
+                i["is_success"] if (i is not None and "is_success" in i) else False
+                for i in info["final_info"]
+            ]
         else:
             successes = [False] * env.num_envs
 
@@ -516,6 +520,7 @@ def eval(
     out_dir = (
         f"outputs/eval/{dt.now().strftime('%Y-%m-%d/%H-%M-%S')}_{hydra_cfg.env.name}_{hydra_cfg.policy.name}"
     )
+    os.makedirs(out_dir, exist_ok=True)
 
     if out_dir is None:
         raise NotImplementedError()
@@ -545,7 +550,7 @@ def eval(
             env,
             policy,
             hydra_cfg.eval.n_episodes,
-            max_episodes_rendered=10,
+            max_episodes_rendered=hydra_cfg.eval.max_episodes_rendered,
             video_dir=Path(out_dir) / "eval",
             start_seed=hydra_cfg.seed,
             enable_progbar=True,

tests/test_examples.py

@@ -29,8 +29,8 @@ def _find_and_replace(text: str, finds_and_replaces: list[tuple[str, str]]) -> s
     return text
 
 
-def _run_script(path):
+def _run_script(path, args=None):
-    subprocess.run([sys.executable, path], check=True)
+    subprocess.run([sys.executable, path] + args if args is not None else [], check=True)
 
 
 def _read_file(path):
@@ -126,3 +126,22 @@ def test_examples_basic2_basic3_advanced1():
     # Restore stdout to its original state
     sys.stdout = sys.__stdout__
     assert "Average loss on validation set" in printed_output
+
+
+def test_real_world_recording():
+    path = "examples/real_robot_example/record_training_data.py"
+    _run_script(
+        path,
+        [
+            "--data_dir",
+            "outputs/examples",
+            "--repo-id",
+            "real_world_debug",
+            "--num-episodes",
+            "2",
+            "--num-frames",
+            "10",
+            "--mock-robot",
+        ],
+    )
+    assert Path("outputs/examples/real_world_debug/video/episode_0.mp4").exists()