Add exact runtime config for piper color sorting (config.yaml)
Browse files
configs/piper-color-sorting/config.yaml
ADDED
|
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
action_dim: 7
|
| 2 |
+
action_model_type: DiT-L
|
| 3 |
+
consolidate_type: tome
|
| 4 |
+
data_root_dir: /training-results/haodong/Real/memoryvla_rlds
|
| 5 |
+
dataloader_type: stream
|
| 6 |
+
ema_decay: 0.999
|
| 7 |
+
fusion_type: gate
|
| 8 |
+
future_action_window_size: 15
|
| 9 |
+
group_size: 16
|
| 10 |
+
hf_token: HF_TOKEN
|
| 11 |
+
image_aug: false
|
| 12 |
+
image_aug_mode: spatial
|
| 13 |
+
is_resume: false
|
| 14 |
+
load_all_data_for_training: true
|
| 15 |
+
mem_length: 256
|
| 16 |
+
per_token_size: 256
|
| 17 |
+
pretrained_checkpoint: /training-results/haodong/MemoryVLA/code/pretrained/CogACT-Large/checkpoints/CogACT-Large.pt
|
| 18 |
+
repeated_diffusion_steps: 4
|
| 19 |
+
resume_epoch: 0
|
| 20 |
+
resume_step: 0
|
| 21 |
+
retain_optimizer_checkpoints: 2
|
| 22 |
+
retrieval_layers: 2
|
| 23 |
+
run_id: piper-color-sorting-frozen16-ema-real21-20260823-053817
|
| 24 |
+
run_id_note: null
|
| 25 |
+
run_root_dir: /training-results/haodong/MemoryVLA/runs/piper-color-sorting
|
| 26 |
+
save_interval: 2000
|
| 27 |
+
save_on_terminate: true
|
| 28 |
+
seed: 42
|
| 29 |
+
trackers:
|
| 30 |
+
- jsonl
|
| 31 |
+
- wandb
|
| 32 |
+
update_fused: false
|
| 33 |
+
use_ema: true
|
| 34 |
+
use_timestep_pe: true
|
| 35 |
+
vla:
|
| 36 |
+
base_vlm: prism-dinosiglip-224px+7b
|
| 37 |
+
data_mix: custom_finetuning
|
| 38 |
+
enable_gradient_checkpointing: true
|
| 39 |
+
enable_mixed_precision_training: true
|
| 40 |
+
epochs: 100
|
| 41 |
+
expected_world_size: 16
|
| 42 |
+
freeze_llm_backbone: true
|
| 43 |
+
freeze_vision_backbone: true
|
| 44 |
+
global_batch_size: 32
|
| 45 |
+
learning_rate: 2.0e-05
|
| 46 |
+
lr_scheduler_type: linear-warmup+cosine-decay
|
| 47 |
+
max_grad_norm: 1.0
|
| 48 |
+
max_steps: 20000
|
| 49 |
+
per_device_batch_size: 2
|
| 50 |
+
reduce_in_full_precision: true
|
| 51 |
+
shuffle_buffer_size: 64
|
| 52 |
+
train_strategy: fsdp-full-shard
|
| 53 |
+
type: prism-dinosiglip-224px+oxe+diffusion
|
| 54 |
+
unfreeze_last_llm_layer: false
|
| 55 |
+
vla_id: prism-dinosiglip-224px+oxe+diffusion
|
| 56 |
+
warmup_ratio: 0.03
|
| 57 |
+
weight_decay: 0.0
|
| 58 |
+
wandb_entity: spikingtransformer
|
| 59 |
+
wandb_project: Memory World Model
|