File size: 2,745 Bytes
06f9561
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import numpy as np

from psi.config.transform import SeedPosttrainRepackTransform


def _sample(mask: np.ndarray) -> dict[str, np.ndarray | str]:
    actions = np.arange(4 * 36, dtype=np.float32).reshape(4, 36)
    states = (actions + 0.25).astype(np.float32)
    return {
        "actions": actions,
        "states": states,
        "action_mask": mask,
        "instruction": "synthetic",
    }


def test_textop_preserves_frame_mask_from_1d_input() -> None:
    transform = SeedPosttrainRepackTransform(textop_actions=True)
    output = transform(_sample(np.asarray([1, 1, 0, 0], dtype=np.float32)))
    mask = output["actions_mask"]
    assert mask.shape == (4, 39)
    np.testing.assert_array_equal(mask[:, 0], np.asarray([1, 1, 0, 0], dtype=np.float32))
    np.testing.assert_array_equal(mask[0], np.ones(39, dtype=np.float32))
    np.testing.assert_array_equal(mask[2], np.zeros(39, dtype=np.float32))


def test_textop_reduces_per_dimension_mask_to_frame_validity() -> None:
    mask = np.ones((4, 36), dtype=np.float32)
    mask[1, :] = 0.0
    mask[3, 0] = 0.0  # at least one original action dimension remains valid
    output = SeedPosttrainRepackTransform(textop_actions=True)(_sample(mask))
    output_mask = output["actions_mask"]
    np.testing.assert_array_equal(output_mask[0], np.ones(39, dtype=np.float32))
    np.testing.assert_array_equal(output_mask[1], np.zeros(39, dtype=np.float32))
    np.testing.assert_array_equal(output_mask[3], np.ones(39, dtype=np.float32))


def test_textop_rejects_mask_with_wrong_frame_count() -> None:
    transform = SeedPosttrainRepackTransform(textop_actions=True)
    try:
        transform(_sample(np.ones((3, 36), dtype=np.float32)))
    except ValueError as exc:
        assert "does not match actions" in str(exc)
    else:
        raise AssertionError("TextOp must reject an action mask with wrong frame count")


def test_textop_repack_builds_two_temporal_context_tokens() -> None:
    sample = _sample(np.ones(4, dtype=np.float32))
    sample["smoothness_context_actions"] = np.asarray(
        sample["actions"][:2],
        dtype=np.float32,
    )
    sample["smoothness_context_reference_states"] = np.asarray(
        sample["states"][:1],
        dtype=np.float32,
    )
    output = SeedPosttrainRepackTransform(textop_actions=True)(sample)
    assert output["smoothness_context_actions"].shape == (2, 39)
    assert np.isfinite(output["smoothness_context_actions"]).all()


if __name__ == "__main__":
    tests = [
        value
        for name, value in sorted(globals().items())
        if name.startswith("test_") and callable(value)
    ]
    for test in tests:
        test()
    print(f"transform_seed_test: {len(tests)} passed")