Use training.eval_freq=-1 for real-world data training + Add diffusion_real.yaml

lerobot/configs/policy/act_real.yaml

@@ -24,7 +24,7 @@ override_dataset_stats:
 training:
   offline_steps: 80000
   online_steps: 0
-  eval_freq: 99999999999999
+  eval_freq: -1
   save_freq: 10000
   log_freq: 100
   save_model: true

lerobot/configs/policy/diffusion_real.yaml

@@ -0,0 +1,114 @@
+# @package _global_
+
+# Defaults for training for the PushT dataset as per https://github.com/real-stanford/diffusion_policy.
+# Note: We do not track EMA model weights as we discovered it does not improve the results. See
+#       https://github.com/huggingface/lerobot/pull/134 for more details.
+
+seed: 100000
+dataset_repo_id: lerobot/pusht
+
+override_dataset_stats:
+  observation.images.cam_right_wrist:
+    # 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.cam_left_wrist:
+    # 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.cam_high:
+    # 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.cam_low:
+    # 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: 200000
+  online_steps: 0
+  eval_freq: -1
+  save_freq: 1000
+  log_freq: 100
+  save_model: true
+
+  batch_size: 64
+  grad_clip_norm: 10
+  lr: 1.0e-4
+  lr_scheduler: cosine
+  lr_warmup_steps: 500
+  adam_betas: [0.95, 0.999]
+  adam_eps: 1.0e-8
+  adam_weight_decay: 1.0e-6
+  online_steps_between_rollouts: 1
+
+  delta_timestamps:
+    observation.images.cam_right_wrist: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1)]"
+    observation.images.cam_left_wrist: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1)]"
+    observation.images.cam_high: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1)]"
+    observation.images.cam_low: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1)]"
+    observation.state: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1)]"
+    action: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1 - ${policy.n_obs_steps} + ${policy.horizon})]"
+
+eval:
+  n_episodes: 50
+  batch_size: 50
+
+policy:
+  name: diffusion
+
+  # Input / output structure.
+  n_obs_steps: 2
+  horizon: 16
+  n_action_steps: 8
+
+  input_shapes:
+    # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
+    observation.images.cam_right_wrist: [3, 480, 640]
+    observation.images.cam_left_wrist: [3, 480, 640]
+    observation.images.cam_high: [3, 480, 640]
+    observation.images.cam_low: [3, 480, 640]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes:
+    observation.images.cam_right_wrist: mean_std
+    observation.images.cam_left_wrist: mean_std
+    observation.images.cam_high: mean_std
+    observation.images.cam_low: mean_std
+    observation.state: mean_std
+  output_normalization_modes:
+    action: mean_std
+
+  # Architecture / modeling.
+  # Vision backbone.
+  vision_backbone: resnet18
+  crop_shape: null
+  crop_is_random: False
+  pretrained_backbone_weights: ResNet18_Weights.IMAGENET1K_V1
+  use_group_norm: False
+  spatial_softmax_num_keypoints: 32
+  # Unet.
+  down_dims: [512, 1024, 2048]
+  kernel_size: 5
+  n_groups: 8
+  diffusion_step_embed_dim: 128
+  use_film_scale_modulation: True
+  # Noise scheduler.
+  noise_scheduler_type: DDPM
+  num_train_timesteps: 100
+  beta_schedule: squaredcos_cap_v2
+  beta_start: 0.0001
+  beta_end: 0.02
+  prediction_type: epsilon # epsilon / sample
+  clip_sample: True
+  clip_sample_range: 1.0
+
+  # Inference
+  num_inference_steps: 100
+
+  # Loss computation
+  do_mask_loss_for_padding: false

lerobot/scripts/train.py

@@ -329,8 +329,12 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
     logging.info("make_dataset")
     offline_dataset = make_dataset(cfg)
 
-    logging.info("make_env")
+    # Create environment used for evaluating checkpoints during training on simulation data.
-    eval_env = make_env(cfg)
+    # On real-world data, no need to create an environment as evaluations are done outside train.py,
+    # using the eval.py instead, with gym_dora environment and dora-rs.
+    if cfg.training.eval_freq > 0:
+        logging.info("make_env")
+        eval_env = make_env(cfg)
 
     logging.info("make_policy")
     policy = make_policy(hydra_cfg=cfg, dataset_stats=offline_dataset.stats)
@@ -356,7 +360,7 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
 
     # Note: this helper will be used in offline and online training loops.
     def evaluate_and_checkpoint_if_needed(step):
-        if step % cfg.training.eval_freq == 0:
+        if cfg.training.eval_freq > 0 and step % cfg.training.eval_freq == 0:
             logging.info(f"Eval policy at step {step}")
             eval_info = eval_policy(
                 eval_env,
@@ -417,6 +421,13 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
 
         step += 1
 
+    logging.info("End of offline training")
+
+    if cfg.training.online_steps == 0:
+        if cfg.training.eval_freq > 0:
+            eval_env.close()
+        return
+
     # create an env dedicated to online episodes collection from policy rollout
     online_training_env = make_env(cfg, n_envs=1)
 
@@ -486,9 +497,10 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
             step += 1
             online_step += 1
 
+    logging.info("End of online training")
+
     eval_env.close()
     online_training_env.close()
-    logging.info("End of training")
 
 
 if __name__ == "__main__":