import torch

from pallet.models import PalletDetector


def test_p2_transfer_maps_only_semantically_matching_neck_blocks():
    torch.set_num_threads(1)
    source = PalletDetector(1, p2=False)
    target = PalletDetector(1, p2=True)
    source.yolo.model[17].conv.weight.data.fill_(0.17)
    source.yolo.model[19].cv2.conv.weight.data.fill_(0.19)
    source.yolo.model[20].conv.weight.data.fill_(0.20)
    source.yolo.model[22].cv2.conv.weight.data.fill_(0.22)
    source.yolo.model[23].cv3[0][0][0].conv.weight.data.fill_(0.23)
    new_p2 = target.yolo.model[19].cv2.conv.weight.detach().clone()
    target.load_pretrained(source.yolo)
    for index, expected in ((23, 0.17), (25, 0.19), (26, 0.20), (28, 0.22)):
        block = target.yolo.model[index]
        weight = block.conv.weight if hasattr(block, 'conv') else block.cv2.conv.weight
        assert torch.all(weight == expected)
    torch.testing.assert_close(target.yolo.model[19].cv2.conv.weight, new_p2)
    assert torch.all(target.yolo.model[29].cv3[1][0][0].conv.weight == 0.23)
    report = target.transfer_report
    assert report['transferred_numel'] > 0
    assert report['target_numel_coverage'] > 0
    assert any(row['source'].startswith('model.23.') for row in report['skipped'])
    assert 'model.19.cv2.conv.weight' in report['new_target']
