Upload config.yaml with huggingface_hub
Browse files- config.yaml +67 -0
config.yaml
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# @package _global_
|
| 2 |
+
#
|
| 3 |
+
# Stage B on LIBERO-90, energy score -- the distributional objective.
|
| 4 |
+
#
|
| 5 |
+
# bash scripts/train_zero2.sh 6 task=libero90_opsd_distill_energy # resume=$STORE/checkpoints/fastwam_release/libero_uncond_2cam224.pt # opsd.teacher_ckpt=<Stage A checkpoint>
|
| 6 |
+
#
|
| 7 |
+
# Confirm "teacher loaded: N adapter + M provider tensors" in the first log
|
| 8 |
+
# lines. Its absence means the teacher is untrained and the run is void.
|
| 9 |
+
#
|
| 10 |
+
# Each student sample costs one extra sampler rollout, so `energy_samples=4`
|
| 11 |
+
# makes a step roughly four times the cost of the squared-error arm;
|
| 12 |
+
# `num_sigma_points` is dropped to 1 to claw part of that back. The top point
|
| 13 |
+
# alone carries most of the discretization mass.
|
| 14 |
+
|
| 15 |
+
defaults:
|
| 16 |
+
- override /data: libero90
|
| 17 |
+
- override /model: fastwam
|
| 18 |
+
- _self_
|
| 19 |
+
|
| 20 |
+
batch_size: 8
|
| 21 |
+
num_workers: 8
|
| 22 |
+
|
| 23 |
+
model:
|
| 24 |
+
mot_checkpoint_mixed_attn: true
|
| 25 |
+
# Every arm starts from the released LIBERO checkpoint, passed as `resume`;
|
| 26 |
+
# loading the generic ActionDiT backbone first would overwrite it.
|
| 27 |
+
skip_dit_load_from_pretrain: true
|
| 28 |
+
action_dit_pretrained_path: null
|
| 29 |
+
|
| 30 |
+
opsd:
|
| 31 |
+
enabled: true
|
| 32 |
+
mode: distill
|
| 33 |
+
phi_provider: future_latent
|
| 34 |
+
phi_dim: 512
|
| 35 |
+
phi_tokens: 8
|
| 36 |
+
adapter_layers: [4, 9, 14, 19, 24, 29]
|
| 37 |
+
adapter_bottleneck: 512
|
| 38 |
+
alpha_init: 1.0
|
| 39 |
+
zero_init_up: true
|
| 40 |
+
action_loss_only: true
|
| 41 |
+
teacher_ckpt: null
|
| 42 |
+
loss_type: energy
|
| 43 |
+
energy_samples: 4
|
| 44 |
+
lambda_flow: 1.0
|
| 45 |
+
lambda_bc: 1.0
|
| 46 |
+
distill_inference_steps: 10
|
| 47 |
+
num_sigma_points: 1
|
| 48 |
+
loss_weighting: uniform
|
| 49 |
+
# student = on-policy states from the student's own sampler (TOPS-WAM);
|
| 50 |
+
# demo = the demonstration's action noised to the same sigma points (off-policy ablation)
|
| 51 |
+
state_source: student
|
| 52 |
+
|
| 53 |
+
frozen_experts: [video]
|
| 54 |
+
save_trainable_only: false
|
| 55 |
+
save_training_state: false
|
| 56 |
+
|
| 57 |
+
lr_scheduler_type: "cosine"
|
| 58 |
+
learning_rate: 1e-5
|
| 59 |
+
num_epochs: 100
|
| 60 |
+
max_steps: 500
|
| 61 |
+
log_every: 10
|
| 62 |
+
save_every: 50
|
| 63 |
+
eval_every: 50
|
| 64 |
+
|
| 65 |
+
gradient_accumulation_steps: 4
|
| 66 |
+
weight_decay: 1e-2
|
| 67 |
+
resume: null
|