tests/test_train_collate.py

from __future__ import annotations

import torch

from pino.pimt_model import DEFAULT_EMBEDDING_DIM, objective_targets_to_pyramid
from pino.train import pad_trajectory_collate


def _item(target_obj: torch.Tensor, *, n_mol: int = 2, t_steps: int = 49) -> dict:
    return {
        "tokens": torch.ones(n_mol, DEFAULT_EMBEDDING_DIM),
        "physics": torch.ones(t_steps, n_mol, 2),
        "target_obj": target_obj,
        "target_sub": torch.tensor([0.25, 0.25, 0.25, 0.25, 0.0, 0.5, 0.5]),
    }


def test_objective_targets_to_pyramid_keeps_native_pyramid() -> None:
    targets = torch.rand(3, 138)

    normalized = objective_targets_to_pyramid(targets)

    assert normalized.shape == (3, 138)
    assert torch.equal(normalized, targets)


def test_pad_trajectory_collate_normalizes_mixed_objective_target_shapes() -> None:
    pyramid = torch.zeros(3, 138)
    pyramid[0, 1] = 1.0
    trajectory = torch.zeros(49, 138)
    trajectory[0, 2] = 0.4
    trajectory[20, 3] = 0.7
    trajectory[-1, 4] = 1.0

    batch = pad_trajectory_collate([_item(pyramid), _item(trajectory, n_mol=3)])

    assert batch["target_obj"].shape == (2, 3, 138)
    assert torch.equal(batch["target_obj"][0], pyramid)
    assert batch["target_obj"][1, 0, 2] == 0.4
    assert batch["target_obj"][1, 1, 3] == 0.7
    assert batch["target_obj"][1, 2, 4] == 1.0


def test_pad_trajectory_collate_normalizes_mixed_objective_target_shapes_when_legacy_first() -> None:
    trajectory = torch.zeros(49, 138)
    pyramid = torch.zeros(3, 138)

    batch = pad_trajectory_collate([_item(trajectory), _item(pyramid)])

    assert batch["target_obj"].shape == (2, 3, 138)