configs/nano_action_causal_rope_v1.yaml
1.8 KB · 67 lines · yaml Raw
1 # Exact training config of runs/nano_action_causal_rope_v1 (only machine paths replaced by env vars).
2 # Needs ACTS_DATA_ROOT and ACTS_CKPT_ROOT (source env.sh). Launch: bash scripts/train.sh configs/nano_action_causal_rope_v1.yaml
3 run_name: nano_action_causal_rope_v1
4 seed: 0
5 out_dir: ${oc.env:ACTS_RUNS_ROOT,runs}
6 data:
7 latent_root: ${oc.env:ACTS_DATA_ROOT}/mixed_rope_latents
8 action_root: ${oc.env:ACTS_DATA_ROOT}/mixed_rope_actions
9 action_file: sensor_actions_with_dforce_fps15.pt
10 view_action_file: view_actions_fps15.pt
11 split_file: splits/mixed_rope_wrist_holdout.json
12 optional_streams: []
13 action_norm: quantile
14 train_window_stride: 1
15 val_window_stride: 50
16 stats_window_stride: 8
17 action_stats: null
18 num_workers: 4
19 model:
20 language_model: nano
21 nano_checkpoint: ${oc.env:ACTS_CKPT_ROOT}/Cosmos3-Nano
22 strict_load: true
23 dtype: bfloat16
24 attention_backend: flex
25 flex_preference: flex_triton
26 temporal_mode: causal
27 attention_topology: all_streams_causal
28 task_mode: action_full
29 init_tactile_from_rgb: true
30 init_action_from_nano:
31 domain: umi
32 train_time_embedder: false
33 train_rgb_projections: false
34 lora:
35 enable: true
36 rank: 16
37 alpha: 32
38 targets: attn_ffn
39 training:
40 prediction_strategy: causal_teacher_forcing
41 init_from_checkpoint: null
42 missing_stream_training: false
43 wrist_dropout:
44 probability: 0.0
45 mode: paired
46 scope: whole_clip
47 batching: group_by_active_streams
48 sampling:
49 mode: uniform
50 steps_per_epoch: null
51 val_hide_wrist: false
52 epochs: 20
53 batch_size: 2
54 lr: 5.0e-05
55 weight_decay: 0.01
56 warmup_steps: 100
57 grad_clip: 1.0
58 amp: bf16
59 sigma_distribution: waver
60 sigma_shift: 3.0
61 loss_weights: {}
62 log_every: 20
63 ckpt_every_epochs: 1
64 activation_checkpointing: true
65 inference:
66 prediction_strategy: causal_autoregressive
67