Views
No views yet
1pip install huggingface_hub
2huggingface-cli download denkiwakame/Qwen3.5-2B-FFT-LAP-NOMASK-UR5e --local-dir ./Qwen3.5-2B-FFT-LAP-NOMASK-UR5e1import pickle
2import torch
3from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2_5_VLProcessor
4
5ckpt_dir = "./Qwen3.5-2B-FFT-LAP-NOMASK-UR5e"
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("./Qwen3.5-2B-FFT-LAP-NOMASK-UR5e", 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/Qwen3.5-2B-FFT-LAP-NOMASK-UR5e --revision step-18000 --local-dir ./Qwen3.5-2B-FFT-LAP-NOMASK-UR5e-step-18000
3
4# Or load directly via transformers
5Qwen2_5_VLForConditionalGeneration.from_pretrained(
6 "denkiwakame/Qwen3.5-2B-FFT-LAP-NOMASK-UR5e",
7 revision="step-18000",
8 subfolder="model_final",
9)Qwen/Qwen3.5-2B1DATALOADER:
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: lap_qwen3_5_2b_fft_ur5e_cluttered_pick_3obj_120
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 - 2000
22 - 4000
23 - 6000
24 save_ckp: 0
25 save_last_ckpt: true
26 test_eval_freq: 1
27 val_eval_freq: 1
28LR_SCHED:
29 lr_clip: 1.0e-08
30 lr_decay_factor: 0.5
31 lr_patience: 4
32MODEL:
33 QWEN:
34 action_mask_aug_per: 0.0
35 action_type: original
36 add_vision_id: true
37 attention_dropout: 0.0
38 enable_thinking: true
39 grad_checkpoint: false
40 history: 1
41 horizon: 8
42 lap_action_is_absolute: true
43 lap_emit_holds: false
44 lap_rotation_precision: 1
45 lap_sum_decimal: 1f
46 lora_config: default
47 lora_rank: 8
48 num_bins_actions: 1000
49 num_cam: 2
50 original_action_dim: 7
51 qwen_model_id: Qwen/Qwen3.5-2B
52 reasoning: true
53 rgb_img_size:
54 - 224
55 - 224
56 rgb_input: true
57 tiled_rgb_imgs: true
58 use_flash_attention_2: true
59 use_lora: false
60 use_qlora: false
61TRAIN:
62 clip_grad_norm: 0.0
63 l2: 1.0e-10
64 lr: 1.0e-05
65 num_epochs: 100
66 num_iters: 20000
67 save_iter_ckp: 2500
68WANDB:
69 enable: true
70 entity: ''
71 log_interval: 50
72 mode: online
73 project: vla0
74 resume_id: ''
75 run_name: ''
76 tags: ''
77
78| 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 |