WIP

lerobot/common/datasets/factory.py

@@ -123,6 +123,23 @@ def make_offline_buffer(
             stats["action", "min"] = torch.tensor([12.0, 25.0], dtype=torch.float32)
             stats["action", "max"] = torch.tensor([511.0, 511.0], dtype=torch.float32)
 
+        # if (
+        #     cfg.env.name == "aloha"
+        #     and cfg.env.task == "sim_transfer_cube"
+        #     and cfg.policy.name == "act"
+        # ):
+        #     import pickle
+
+        #     with open(
+        #         "/home/rcadene/code/act/tmp/2024_03_10_sim_transfer_cube_scripted/dataset_stats.pkl", "rb"
+        #     ) as file:
+        #         dataset_stats = pickle.load(file)
+
+        #     stats["action", "mean"] = torch.from_numpy(dataset_stats["action_mean"])
+        #     stats["action", "std"] = torch.from_numpy(dataset_stats["action_std"])
+        #     stats["observation", "state", "mean"] = torch.from_numpy(dataset_stats["qpos_mean"])
+        #     stats["observation", "state", "std"] = torch.from_numpy(dataset_stats["qpos_std"])
+
         # TODO(rcadene): remove this and put it in config. Ideally we want to reproduce SOTA results just with mean_std
         normalization_mode = "mean_std" if cfg.env.name == "aloha" else "min_max"
         transforms.append(NormalizeTransform(stats, in_keys, mode=normalization_mode))

lerobot/common/envs/simxarm/simxarm/tasks/base.py

@@ -138,6 +138,9 @@ class Base(robot_env.MujocoRobotEnv):
         # HACK
         self.model.vis.global_.offwidth = width
         self.model.vis.global_.offheight = height
+        if mode in self.mujoco_renderer._viewers:
+            self.mujoco_renderer._viewers.get(mode).viewport.width = width
+            self.mujoco_renderer._viewers.get(mode).viewport.height = height
         return self.mujoco_renderer.render(mode)
 
     def close(self):

lerobot/common/policies/factory.py

@@ -26,15 +26,22 @@ def make_policy(cfg):
         policy = ActionChunkingTransformerPolicy(
             cfg.policy, cfg.device, n_action_steps=cfg.n_action_steps + cfg.n_latency_steps
         )
+
+        # if cfg.env.name == "aloha" and cfg.env.task == "sim_transfer_cube_scripted":
+        #     import torch
+        #     state_dict = torch.load(
+        #         "/home/rcadene/code/act/tmp/2024_03_10_sim_transfer_cube_scripted/policy_best.ckpt"
+        #     )
+        #     policy.load_state_dict(state_dict)
     else:
         raise ValueError(cfg.policy.name)
 
     if cfg.policy.pretrained_model_path:
         # TODO(rcadene): hack for old pretrained models from fowm
         if cfg.policy.name == "tdmpc" and "fowm" in cfg.policy.pretrained_model_path:
-            if "offline" in cfg.pretrained_model_path:
+            if "offline" in cfg.policy.pretrained_model_path:
                 policy.step[0] = 25000
-            elif "final" in cfg.pretrained_model_path:
+            elif "final" in cfg.policy.pretrained_model_path:
                 policy.step[0] = 100000
             else:
                 raise NotImplementedError()

lerobot/common/policies/tdmpc/helper.py

@@ -152,7 +152,7 @@ def enc(cfg):
     if cfg.modality in {"pixels", "all"}:
         C = int(3 * cfg.frame_stack)  # noqa: N806
         pixels_enc_layers = [
-            NormalizeImg(),
+            nn.Identity(),  # NormalizeImg(),  # TODO(rcadene): we need to clean this
             nn.Conv2d(C, cfg.num_channels, 7, stride=2),
             nn.ReLU(),
             nn.Conv2d(cfg.num_channels, cfg.num_channels, 5, stride=2),

lerobot/scripts/eval.py

@@ -86,6 +86,8 @@ def eval_policy(
 
         def maybe_render_frame(env: EnvBase, _):
             if save_video or (return_first_video and i == 0):  # noqa: B023
+                # TODO(rcadene): set width and height from config or env
+                # ep_frames.append(env.render(width=384, height=384))  # noqa: B023
                 ep_frames.append(env.render())  # noqa: B023
 
         # Clear the policy's action queue before the start of a new rollout.
@@ -266,7 +268,8 @@ if __name__ == "__main__":
         )
         cfg = hydra.compose(Path(args.config).stem, args.overrides)
         # TODO(alexander-soare): Save and load stats in trained model directory.
-        stats_path = None
+        # stats_path = None
+        stats_path = "/home/rcadene/code/lerobot/outputs/pretrained_models/act_aloha_sim_transfer_cube_scripted/stats.pth"
     elif args.hub_id is not None:
         folder = Path(snapshot_download(args.hub_id, revision=args.revision))
         cfg = hydra.initialize(config_path=str(_relative_path_between(folder, Path(__file__).parent)))

lerobot/scripts/train.py

@@ -172,7 +172,8 @@ def train(cfg: dict, out_dir=None, job_name=None):
 
     # Note: this helper will be used in offline and online training loops.
     def _maybe_eval_and_maybe_save(step):
-        if step % cfg.eval_freq == 0:
+        # if step % cfg.eval_freq == 0:
+        if True:
             logging.info(f"Eval policy at step {step}")
             eval_info, first_video = eval_policy(
                 env,
@@ -187,8 +188,8 @@ def train(cfg: dict, out_dir=None, job_name=None):
             if cfg.wandb.enable:
                 logger.log_video(first_video, step, mode="eval")
             logging.info("Resume training")
-
-        if cfg.save_model and step % cfg.save_freq == 0:
+        if True:
+            # if cfg.save_model and step % cfg.save_freq == 0:
             logging.info(f"Checkpoint policy after step {step}")
             logger.save_model(policy, identifier=step)
             logging.info("Resume training")