configs/nano_action_causal_umi_sensoronly_noforce_v1.yaml
| 1 | # Exact training config of Slurm job 884599 (rebuttal_2) = runs/nano_action_causal_umi_sensoronly_noforce_v1 |
| 2 | # (motherboard; sensor-only: no view-action streams; no force: action_file = sensor_actions_fps15_noforce.pt). |
| 3 | # Only machine paths replaced by env vars. |
| 4 | # Needs ACTS_DATA_ROOT and ACTS_CKPT_ROOT (source env.sh). Launch: bash scripts/train.sh configs/nano_action_causal_umi_sensoronly_noforce_v1.yaml |
| 5 | run_name: nano_action_causal_umi_sensoronly_noforce_v1 |
| 6 | seed: 0 |
| 7 | out_dir: ${oc.env:ACTS_RUNS_ROOT,runs} |
| 8 | data: |
| 9 | latent_root: ${oc.env:ACTS_DATA_ROOT}/mixed_mb_latents |
| 10 | action_root: ${oc.env:ACTS_DATA_ROOT}/mixed_mb_actions |
| 11 | action_file: sensor_actions_fps15_noforce.pt |
| 12 | view_action_file: view_actions_fps15.pt |
| 13 | split_file: splits/mixed_mb_wrist_holdout.json |
| 14 | optional_streams: [] |
| 15 | action_norm: quantile |
| 16 | train_window_stride: 1 |
| 17 | val_window_stride: 50 |
| 18 | stats_window_stride: 8 |
| 19 | action_stats: null |
| 20 | num_workers: 4 |
| 21 | model: |
| 22 | language_model: nano |
| 23 | nano_checkpoint: ${oc.env:ACTS_CKPT_ROOT}/Cosmos3-Nano |
| 24 | strict_load: true |
| 25 | dtype: bfloat16 |
| 26 | attention_backend: flex |
| 27 | flex_preference: flex_triton |
| 28 | temporal_mode: causal |
| 29 | attention_topology: all_streams_causal |
| 30 | task_mode: action_full |
| 31 | active_streams: |
| 32 | - video |
| 33 | - wrist_left |
| 34 | - wrist_right |
| 35 | - tactile_left |
| 36 | - tactile_right |
| 37 | - action_left |
| 38 | - action_right |
| 39 | init_tactile_from_rgb: true |
| 40 | init_action_from_nano: |
| 41 | domain: umi |
| 42 | train_time_embedder: false |
| 43 | train_rgb_projections: false |
| 44 | lora: |
| 45 | enable: true |
| 46 | rank: 16 |
| 47 | alpha: 32 |
| 48 | targets: attn_ffn |
| 49 | training: |
| 50 | prediction_strategy: causal_teacher_forcing |
| 51 | init_from_checkpoint: null |
| 52 | missing_stream_training: false |
| 53 | wrist_dropout: |
| 54 | probability: 0.0 |
| 55 | mode: paired |
| 56 | scope: whole_clip |
| 57 | batching: group_by_active_streams |
| 58 | sampling: |
| 59 | mode: uniform |
| 60 | steps_per_epoch: null |
| 61 | val_hide_wrist: false |
| 62 | epochs: 20 |
| 63 | batch_size: 2 |
| 64 | lr: 5.0e-05 |
| 65 | weight_decay: 0.01 |
| 66 | warmup_steps: 100 |
| 67 | grad_clip: 1.0 |
| 68 | amp: bf16 |
| 69 | sigma_distribution: waver |
| 70 | sigma_shift: 3.0 |
| 71 | loss_weights: {} |
| 72 | log_every: 20 |
| 73 | ckpt_every_epochs: 1 |
| 74 | activation_checkpointing: true |
| 75 | inference: |
| 76 | prediction_strategy: causal_autoregressive |
| 77 | |