Add diffusion, vqbet, act bigg

lerobot/configs/policy/act_koch_real.yaml

@@ -42,7 +42,7 @@ training:
   online_steps_between_rollouts: 1
 
   delta_timestamps:
-    action: "[i / ${fps} for i in range(1, ${policy.chunk_size}+1)]"
+    action: "[i / ${fps} + 1 / ${fps} for i in range(${policy.chunk_size})]"
 
 eval:
   n_episodes: 50

lerobot/configs/policy/act_koch_real_bigger.yaml

@@ -0,0 +1,102 @@
+# @package _global_
+
+# Use `act_koch_real.yaml` to train on real-world datasets collected on Alexander Koch's robots.
+# Compared to `act.yaml`, it contains 2 cameras (i.e. laptop, phone) 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 our `control_robot.py` script.
+# Look at the documentation in header of `control_robot.py` for more information on how to collect data , train and evaluate a policy.
+#
+# Example of usage for training:
+# ```bash
+# python lerobot/scripts/train.py \
+#   policy=act_koch_real \
+#   env=koch_real
+# ```
+
+seed: 1000
+dataset_repo_id: lerobot/koch_pick_place_lego
+
+override_dataset_stats:
+  observation.images.laptop:
+    # 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.phone:
+    # 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: 40000
+  online_steps: 0
+  eval_freq: -1
+  save_freq: 5000
+  log_freq: 10
+  save_checkpoint: true
+
+  batch_size: 64
+  lr: 1e-4
+  lr_backbone: 1e-4
+  weight_decay: 1e-4
+  grad_clip_norm: 10
+  online_steps_between_rollouts: 1
+
+  delta_timestamps:
+    action: "[i / ${fps} + 1 / ${fps} for i in range(${policy.chunk_size})]"
+
+eval:
+  n_episodes: 50
+  batch_size: 50
+
+# See `configuration_act.py` for more details.
+policy:
+  name: act
+
+  # Input / output structure.
+  n_obs_steps: 1
+  chunk_size: 100
+  n_action_steps: 100
+
+  input_shapes:
+    # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
+    observation.images.laptop: [3, 480, 640]
+    observation.images.phone: [3, 480, 640]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes:
+    observation.images.laptop: mean_std
+    observation.images.phone: 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: 8
+  # 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/configs/policy/diffusion_koch_real.yaml

@@ -0,0 +1,102 @@
+# @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/koch_pick_place_lego
+
+override_dataset_stats:
+  observation.images.laptop:
+    # 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.phone:
+    # 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: 40000
+  online_steps: 0
+  eval_freq: -1
+  save_freq: 5000
+  log_freq: 100
+  save_checkpoint: 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} + 1 / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1 - ${policy.n_obs_steps} + ${policy.horizon})]"
+
+eval:
+  n_episodes: 50
+  batch_size: 50
+
+policy:
+  name: diffusion
+
+  # Input / output structure.
+  n_obs_steps: 1
+  horizon: 100
+  n_action_steps: 100
+
+  input_shapes:
+    # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
+    observation.images.laptop: [3, 480, 640]
+    observation.images.phone: [3, 480, 640]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes:
+    observation.images.laptop: mean_std
+    observation.images.phone: 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: 512
+  # 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/configs/policy/vqbet_koch_real.yaml

@@ -16,17 +16,17 @@ override_dataset_stats:
     std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
 
 training:
-  offline_steps: 80000
+  offline_steps: 40000
   online_steps: 0
   eval_freq: -1
-  save_freq: 10000
+  save_freq: 5000
   save_checkpoint: true
 
-  batch_size: 8
+  batch_size: 64
   grad_clip_norm: 10
   lr: 1.0e-4
   lr_scheduler: cosine
-  lr_warmup_steps: 2000
+  lr_warmup_steps: 500
   adam_betas: [0.95, 0.999]
   adam_eps: 1.0e-8
   adam_weight_decay: 1.0e-6
@@ -34,7 +34,7 @@ training:
 
   # VQ-BeT specific
   vqvae_lr: 1.0e-3
-  n_vqvae_training_steps: 20000
+  n_vqvae_training_steps: 500
   bet_weight_decay: 2e-4
   bet_learning_rate: 5.5e-5
   bet_betas: [0.9, 0.999]
@@ -43,7 +43,7 @@ training:
     observation.images.laptop: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1)]"
     observation.images.phone: "[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}, ${policy.n_action_pred_token} + ${policy.action_chunk_size} - 1)]"
+    action: "[i / ${fps} + 1 / ${fps} for i in range(1 - ${policy.n_obs_steps}, ${policy.n_action_pred_token} + ${policy.action_chunk_size} - 1)]"
 
 eval:
   n_episodes: 50
@@ -53,9 +53,9 @@ policy:
   name: vqbet
 
   # Input / output structure.
-  n_obs_steps: 5
+  n_obs_steps: 1
   n_action_pred_token: 7
-  action_chunk_size: 5
+  action_chunk_size: 100
 
   input_shapes:
     # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?