Views
No views yet
1pip install huggingface_hub
2huggingface-cli download denkiwakame/Qwen2.5-VL-3B-FFT-VLA0-UR5e-4GPU --local-dir ./Qwen2.5-VL-3B-FFT-VLA0-UR5e-4GPU1import pickle
2import torch
3from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2_5_VLProcessor
4
5ckpt_dir = "./Qwen2.5-VL-3B-FFT-VLA0-UR5e-4GPU"
6
7model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
8 f"{ckpt_dir}/model_final", torch_dtype=torch.bfloat16, device_map="auto",
9)
10processor = Qwen2_5_VLProcessor.from_pretrained(f"{ckpt_dir}/model_final")
11
12# Load dataset stats (required for action denormalization)
13with open(f"{ckpt_dir}/dataset_stats.pkl", "rb") as f:
14 dataset_stats = pickle.load(f)1from rv_train.train import get_pretrained_model
2
3model, cfg = get_pretrained_model("./Qwen2.5-VL-3B-FFT-VLA0-UR5e-4GPU", device=0)
4model.eval()dataset_stats.pkl1import pickle
2
3with open("dataset_stats.pkl", "rb") as f:
4 stats = pickle.load(f)
5# stats contains mean/std for action dimensionsmain holds the recommended/final weights.
Earlier training-step snapshots are published as branches named step-<global_step> (e.g., step-17000, step-18000).
Load any of them by passing revision=:1# Download a specific revision
2huggingface-cli download denkiwakame/Qwen2.5-VL-3B-FFT-VLA0-UR5e-4GPU --revision step-18000 --local-dir ./Qwen2.5-VL-3B-FFT-VLA0-UR5e-4GPU-step-18000
3
4# Or load directly via transformers
5Qwen2_5_VLForConditionalGeneration.from_pretrained(
6 "denkiwakame/Qwen2.5-VL-3B-FFT-VLA0-UR5e-4GPU",
7 revision="step-18000",
8 subfolder="model_final",
9)Qwen/Qwen2.5-VL-3B-Instruct1DATALOADER:
2 ROBOVERSE:
3 cfg_opts: IMAGE.crop_img:0.9:IMAGE.img_size:224:IMAGE.cam_list:('3p1','wrist_right1')
4 cfg_path: libs/RoboVerse/roboverse/configs/ur5e_cluttered_pick_3obj_120.yaml
5 batch_size: 16
6 num_workers: 8
7EXP:
8 AMP: true
9 DATASET: roboverse
10 EXP_ID: vla0_qwen2_5_3b_ur5e_4gpu
11 LOSS: {}
12 LR_SCHED: none
13 MODEL: qwen
14 OPTIMIZER: adamw
15 SEED: 0
16EXP_EXTRA:
17 no_test: true
18 no_track: true
19 no_val: true
20 save_at_steps:
21 - 10000
22 - 16000
23 save_ckp: 0
24 save_last_ckpt: true
25 test_eval_freq: 1
26 val_eval_freq: 1
27LR_SCHED:
28 lr_clip: 1.0e-08
29 lr_decay_factor: 0.5
30 lr_patience: 4
31MODEL:
32 QWEN:
33 action_mask_aug_per: 0.4
34 action_type: original
35 add_vision_id: true
36 attention_dropout: 0.0
37 enable_thinking: false
38 grad_checkpoint: false
39 history: 1
40 horizon: 8
41 # lap_action_is_absolute: false
42 # lap_emit_holds: false
43 # lap_rotation_precision: 10
44 # lap_sum_decimal: 0f
45 lora_config: default
46 lora_rank: 8
47 num_bins_actions: 1000
48 num_cam: 2
49 original_action_dim: 7
50 qwen_model_id: Qwen/Qwen2.5-VL-3B-Instruct
51 reasoning: false
52 rgb_img_size:
53 - 224
54 - 224
55 rgb_input: true
56 tiled_rgb_imgs: true
57 use_flash_attention_2: true
58 use_lora: false
59 use_qlora: false
60TRAIN:
61 clip_grad_norm: 0.0
62 l2: 1.0e-10
63 lr: 1.0e-05
64 num_epochs: 100
65 num_iters: 16000
66 save_iter_ckp: 2500
67WANDB:
68 enable: true
69 entity: ''
70 log_interval: 50
71 mode: online
72 project: vla0
73 resume_id: ''
74 run_name: ''
75 tags: ''
76
77| File | Description |
|---|---|
model_final/model-*.safetensors | Full model weights |
model_final/config.json | Model configuration |
model_final/tokenizer.json | Tokenizer |
dataset_stats.pkl | Action normalization statistics (required for inference) |
config.yaml | Training configuration |