AgiBot-World/go1/configs/go1_sft_libero.py at main · OpenDriveLab/AgiBot-World · GitHub

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
61
62
63
64
65
66
67
68
69
70
71
72
import os
from dataclasses import dataclass, field
from typing import List, Optional

from transformers import TrainingArguments

from go1.configs.go1_base_cfg import BaseDatasetArguments, BaseModelArguments, BaseSpaceArguments
from go1.tools.env_parse import get_bool_env

RUNNAME = os.environ.get("RUNNAME")
DEBUG_MODE = get_bool_env("DEBUG_MODE")


@dataclass
class DatasetArguments(BaseDatasetArguments):
    dataset_type: Optional[str] = field(default="lerobot")
    data_root_dir: Optional[List[str]] = field(
        default_factory=lambda: [
            "/path/to/your/libero/dataset",
        ],
    )
    transforms: Optional[List[str]] = field(default_factory=lambda: [dict(type="Normalize")])


@dataclass
class GOModelArguments(BaseModelArguments):
    model_name_or_path: str = field(default="agibot-world/GO-1")
    freeze_llm: bool = field(default=False if not DEBUG_MODE else True)
    freeze_backbone: bool = field(default=False if not DEBUG_MODE else True)
    freeze_mlp: bool = field(default=False if not DEBUG_MODE else True)
    action_chunk_size: int = field(default=10)
    latent_planning: bool = field(default=True)
    freeze_latent_planner: bool = field(default=False)


@dataclass
class GOTrainingArguments(TrainingArguments):
    output_dir: str = field(default=f"experiment/{RUNNAME}")
    overwrite_output_dir: bool = field(default=True)
    dataloader_num_workers: int = field(default=20 if not DEBUG_MODE else 0)
    bf16: bool = field(default=True)
    num_train_epochs: float = field(default=100.0)
    per_device_train_batch_size: int = field(default=16 if not DEBUG_MODE else 2)
    gradient_accumulation_steps: int = field(default=1)
    learning_rate: float = field(default=2e-5)
    weight_decay: float = field(default=0.01)
    lr_scheduler_type: str = field(default="cosine")
    warmup_steps: int = field(default=1000)
    do_train: bool = field(default=True)
    deepspeed: str = field(default="go1/zero_stage1_config.json")

    save_strategy: str = field(default="steps")
    save_steps: int = field(default=10000)
    save_total_limit: int = field(default=100)
    logging_steps: int = field(default=10)
    report_to: str = field(default="tensorboard")


@dataclass
class SpaceArguments(BaseSpaceArguments):
    state_dim: int = field(default=8)
    action_dim: int = field(default=7)
    space_repack: dict = field(
        default_factory=lambda: {
            "state": "state",
            "action": "actions",
            "cam_head_color": "image",
            "cam_hand_left_color": "wrist_image",
            "final_prompt": "task",
        }
    )
    ctrl_freq: int = field(default=10)