File size: 1,632 Bytes
2680c4d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 | action_dim: 7
action_model_type: DiT-L
consolidate_type: tome
data_root_dir: /training-results/haodong/Real/memoryvla_rlds
dataloader_type: stream
ema_decay: 0.999
fusion_type: gate
future_action_window_size: 15
group_size: 16
hf_token: HF_TOKEN
image_aug: false
image_aug_mode: spatial
is_resume: false
load_all_data_for_training: true
mem_length: 256
per_token_size: 256
pretrained_checkpoint: /training-results/haodong/MemoryVLA/code/pretrained/CogACT-Large/checkpoints/CogACT-Large.pt
repeated_diffusion_steps: 4
resume_epoch: 0
resume_step: 0
retain_optimizer_checkpoints: 2
retrieval_layers: 2
run_id: piper-color-sorting-frozen16-ema-real21-20260823-053817
run_id_note: null
run_root_dir: /training-results/haodong/MemoryVLA/runs/piper-color-sorting
save_interval: 2000
save_on_terminate: true
seed: 42
trackers:
- jsonl
- wandb
update_fused: false
use_ema: true
use_timestep_pe: true
vla:
base_vlm: prism-dinosiglip-224px+7b
data_mix: custom_finetuning
enable_gradient_checkpointing: true
enable_mixed_precision_training: true
epochs: 100
expected_world_size: 16
freeze_llm_backbone: true
freeze_vision_backbone: true
global_batch_size: 32
learning_rate: 2.0e-05
lr_scheduler_type: linear-warmup+cosine-decay
max_grad_norm: 1.0
max_steps: 20000
per_device_batch_size: 2
reduce_in_full_precision: true
shuffle_buffer_size: 64
train_strategy: fsdp-full-shard
type: prism-dinosiglip-224px+oxe+diffusion
unfreeze_last_llm_layer: false
vla_id: prism-dinosiglip-224px+oxe+diffusion
warmup_ratio: 0.03
weight_decay: 0.0
wandb_entity: spikingtransformer
wandb_project: Memory World Model
|