RAQEEB-CCTV / code /test_raqeeb_trainer.py
swedishtech's picture
RAQEEB-CCTV v1.0.0
d0f2b17 verified
Raw History Blame Contribute Delete
3.04 kB
"""Regression checks for the RAQEEB partial-label training logic (CPU only)."""
import sys
import unittest
from pathlib import Path
import torch
sys.path.insert(0, str(Path(__file__).resolve().parent))
import raqeeb_trainer as rt # noqa: E402
class MaskedBCETest(unittest.TestCase):
def test_negatives_ignored_for_unannotated_classes_positives_kept(self):
bce = rt.MaskedBCE()
bce.mask = torch.tensor([[1., 0., 0.], [1., 1., 1.]]) # image0 annotates class 0 only
pred = torch.zeros(2, 4, 3)
tgt = torch.zeros(2, 4, 3)
tgt[0, 0, 1] = 0.8 # a positive of a non-full (partial) class must still count
loss = bce(pred, tgt)
self.assertTrue(torch.all(loss[0, :, 0] > 0)) # full class: all anchors count
self.assertGreater(loss[0, 0, 1].item(), 0) # partial positive counts
self.assertEqual(loss[0, 1:, 1].abs().sum().item(), 0) # partial negatives ignored
self.assertEqual(loss[0, :, 2].abs().sum().item(), 0) # unannotated class ignored
self.assertTrue(torch.all(loss[1] > 0)) # fully annotated image unchanged
def test_no_mask_is_plain_bce(self):
bce = rt.MaskedBCE()
p, t = torch.randn(2, 5, 3), torch.rand(2, 5, 3)
ref = torch.nn.functional.binary_cross_entropy_with_logits(p, t, reduction='none')
self.assertTrue(torch.allclose(bce(p, t), ref))
class CollateTest(unittest.TestCase):
def test_cls_mask_stacked(self):
b = [dict(img=torch.zeros(3, 8, 8, dtype=torch.uint8), cls=torch.zeros(1, 1), bboxes=torch.zeros(1, 4),
masks=torch.zeros(1, 2, 2), batch_idx=torch.zeros(1), cls_mask=torch.ones(11) * i,
im_file=f'x{i}') for i in range(3)]
out = rt.RaqeebDataset.collate_fn(b)
self.assertEqual(tuple(out['cls_mask'].shape), (3, 11))
self.assertEqual(out['cls_mask'][2, 0].item(), 2)
self.assertEqual(out['img'].shape[0], 3)
class MosaicSwapTest(unittest.TestCase):
def test_swap(self):
from ultralytics.data.augment import Compose, Mosaic
m = Mosaic.__new__(Mosaic)
c = Compose([Compose([m])])
rt._swap_mosaic(c)
self.assertIsInstance(m, rt.GroupMosaic)
class ValidatorFilterTest(unittest.TestCase):
def test_predictions_of_unannotated_classes_dropped(self):
v = rt.RaqeebSegValidator.__new__(rt.RaqeebSegValidator)
seen = {}
rt.SegmentationValidator.update_metrics = lambda self, preds, batch: seen.setdefault('p', preds)
preds = [dict(cls=torch.tensor([0., 1., 2.]), conf=torch.tensor([.9, .8, .7]),
bboxes=torch.zeros(3, 4), masks=torch.zeros(3, 4, 4))]
batch = dict(cls_mask=torch.tensor([[0., 1., 0.] + [0.] * 8]))
v.update_metrics(preds, batch)
p = seen['p'][0]
self.assertEqual(p['cls'].tolist(), [1.0])
self.assertEqual(tuple(p['masks'].shape), (1, 4, 4))
if __name__ == '__main__':
unittest.main(verbosity=2)