pi06/example_usage.py at main · elias0819/pi06 · 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
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
"""
Example usage of Recap training.

This script demonstrates how to use the Recap implementation
with a Lerobot v3 dataset.
"""

import torch
from pi06 import (
    VLAModel,
    ValueFunction,
    TokenizerWrapper,
    RecapTrainer,
    create_dataloader,
)


def main():
    # Configuration
    device = "cuda" if torch.cuda.is_available() else "cpu"
    action_dim = 7  # Adjust based on your robot
    dataset_path = "path/to/lerobot/dataset"

    # Create tokenizer
    tokenizer = TokenizerWrapper(
        text_tokenizer_name="bert-base-uncased",
        vision_processor_name="openai/clip-vit-base-patch32",
    )

    # You can change tokenizers dynamically:
    # tokenizer.set_text_tokenizer("distilbert-base-uncased")
    # tokenizer.set_vision_processor("openai/clip-vit-large-patch14")

    # Create VLA model
    model = VLAModel(
        action_dim=action_dim,
        vision_model_name="openai/clip-vit-base-patch32",
        text_model_name="bert-base-uncased",
        hidden_dim=512,
        freeze_vision=False,
        freeze_text=False,
        use_advantage_conditioning=True,
        use_quality_conditioning=True,
    )

    # Create value function
    value_function = ValueFunction(
        input_dim=512,  # Should match model.hidden_dim
        hidden_dims=[512, 256],
        activation="gelu",
        output_type="scalar",
    )

    # Create trainer
    trainer = RecapTrainer(
        model=model,
        value_function=value_function,
        tokenizer=tokenizer,
        device=device,
        learning_rate=1e-4,
        value_lr=1e-4,
        weight_decay=0.0,
        use_wandb=True,
        wandb_project="recap",
        wandb_name="example_run",
    )

    # Stage 1: Train with demonstrations
    print("Training with demonstrations...")
    demo_dataloader = create_dataloader(
        dataset_path=dataset_path,
        episode_type="demo",
        batch_size=1,
        shuffle=True,
        num_workers=0,
    )

    trainer.train_demonstrations(
        dataloader=demo_dataloader,
        num_epochs=10,
        log_every=100,
    )

    # Stage 2: Train with corrections
    print("Training with corrections...")
    correction_dataloader = create_dataloader(
        dataset_path=dataset_path,
        episode_type="correction",
        batch_size=1,
        shuffle=True,
        num_workers=0,
    )

    trainer.train_corrections(
        dataloader=correction_dataloader,
        num_epochs=5,
        log_every=100,
    )

    # Stage 3: Train with autonomous experience
    print("Training with autonomous experience...")
    autonomous_dataloader = create_dataloader(
        dataset_path=dataset_path,
        episode_type="autonomous",
        batch_size=1,
        shuffle=True,
        num_workers=0,
    )

    trainer.train_autonomous(
        dataloader=autonomous_dataloader,
        num_epochs=20,
        log_every=100,
        gamma=0.99,
        lambda_=0.95,
        value_loss_weight=0.5,
        policy_loss_weight=1.0,
        entropy_weight=0.01,
    )

    # Save final checkpoint
    trainer.save_checkpoint("checkpoint_final.pt")
    print("Training complete! Checkpoint saved.")


if __name__ == "__main__":
    main()