def test_dataset():
data_path = 'test_data'
data = Data(data_path, mode='train')
from auto_driving.config import Config
config = Config()
config.size = (400, 512)
sizes = [(2710, 3384), config.size]
scale = FrameTransform(config)
transforms = [None, scale.transform]
for transform, expected_size in zip(transforms, sizes):
dataset = FrameDataset(data, config, transform=transform)
sample = dataset[0]
assert len(dataset) == 8
assert 'image' in sample
assert 'mask' in sample
assert 'bboxes' in sample
assert 'classes' in sample
assert sample['image'].type() == torch.FloatTensor(0).type()
assert sample['mask'].type() == torch.FloatTensor(0).type()
assert sample['bboxes'].type() == torch.FloatTensor(0).type()
assert sample['classes'].type() == torch.LongTensor(0).type()
assert sample['image'].shape[0] == 3
assert sample['image'].shape[1:] == expected_size
assert sample['mask'].shape[1] == 1
assert sample['mask'].shape[2:] == expected_size
def dataset_t():
data_path = 'test_data'
data = Data(data_path, mode='train')
config = Config()
config.size = (400, 512)
scale = FrameTransform(config)
dataset = FrameDataset(data, config, transform=scale.transform)
return dataset
def test_evaluate_img_cat(masks1, masks2):
config = Config()
dt_masks, dt_cats = masks1
gt_masks, gt_cats = masks2
scores = np.array([0.8, 0.7, 0.6, 0.5, 0.4] + [0.0] * (len(dt_cats) - 5))
evaluator = Eval(config)
evaluator.params.catIds = [0, 1, 2, 3]
dtm_1, scores_1, G_1 = evaluator.evaluate_img_cat(dt_masks[dt_cats == 1],
gt_masks[gt_cats == 1],
scores[dt_cats == 1])
assert dtm_1.shape[1] == 1
assert dtm_1[0, 0] == 1
for i in range(1, dtm_1.shape[0]):
assert dtm_1[i, 0] == 0
assert G_1 == 1
dtm_2, scores_2, G_2 = evaluator.evaluate_img_cat(dt_masks[dt_cats == 2],
gt_masks[gt_cats == 2],
scores[dt_cats == 2])
assert G_2 == 3
assert dtm_2[3, 0] == 1
assert dtm_2[4, 0] == 0
assert dtm_2[6, 1] == 1
assert dtm_2[7, 1] == 0
assert dtm_2[0, 2] == 0
dtm_3, scores_3, G_3 = evaluator.evaluate_img_cat(dt_masks[dt_cats == 3],
gt_masks[gt_cats == 3],
scores[dt_cats == 3])
assert G_3 == 0
def test_accumulate():
config = Config()
scores = np.array([0.9, 0.8, 0.7, 0.6, 0.5, 0.4])
evaluator = Eval(config)
evaluator.params.catIds = [0, 1, 2]
evaluator.params.iouThrs = [0.5, 0.75, 0.95]
matches = []
dtm = np.array([
[1, 1, 0, 1, 0, 0],
[0, 1, 0, 1, 0, 0],
[0, 1, 0, 0, 0, 0],
],
dtype=np.float64)
gtN = 4
matches.append({'dtMatches': dtm, 'dtScores': scores, 'gtN': gtN})
matches.append(None)
evaluator.eval_res.append(matches)
evaluator.accumulate()
pr = np.array([
[1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.75, 0.75, 0.0, 0.0, 0.0],
[0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.0, 0.0, 0.0, 0.0, 0.0],
[0.5, 0.5, 0.5, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0],
])
pr = np.stack([pr, -np.ones_like(pr)], axis=2)
assert np.allclose(pr, evaluator.precision)
def test_transform():
from auto_driving.config import Config
config = Config()
config.size = (400, 512)
transform = FrameTransform(config)
dtypes = [np.int8, np.int32, np.float32, np.float64]
for dtype in dtypes:
sample = {}
sample['image'] = np.ones((2710, 3384, 3), dtype=dtype)
sample['mask'] = np.ones((10, 2710, 3384), dtype=dtype)
newsample = transform.transform(sample)
assert sample != newsample
assert sample['image'] != newsample['image']
assert sample['mask'] != newsample['mask']
assert newsample['image'].shape == (config.size[0], config.size[1], 3)
assert newsample['mask'].shape == (sample['mask'].shape[0],
config.size[0], config.size[1])
assert newsample['image'].dtype == dtype
assert newsample['mask'].dtype == dtype
assert newsample['image'].max() == 1.0
assert newsample['mask'].max() == 1.0
def test_evaluate_img(masks1, masks2):
config = Config()
dt_masks, dt_cats = masks1
gt_masks, gt_cats = masks2
scores = np.array([0.8, 0.7, 0.6, 0.5, 0.4] + [0.0] * (len(dt_cats) - 5))
evaluator = Eval(config)
evaluator.params.catIds = [0, 1, 2, 3]
evaluator.evaluate_img(dt_masks, gt_masks, dt_cats, gt_cats, scores)
assert len(evaluator.eval_res) == 1
assert len(evaluator.eval_res[0]) == len(evaluator.params.catIds) - 1
assert evaluator.eval_res[0][-1] is None
def test_overlap_det():
gt_masks = np.array([[[0, 0, 0], [0, 1, 1], [0, 1, 1]]], dtype=np.bool)
dt_masks = np.array(
[[[0, 0, 0], [0, 1, 1], [0, 1, 1]], [[0, 0, 0], [0, 0, 1], [0, 1, 1]],
[[0, 0, 0], [0, 1, 0], [0, 0, 1]]],
dtype=np.bool)
scores = np.array([0.6, 0.7, 0.8])
config = Config()
evaluator = Eval(config)
dtm, scores, G = evaluator.evaluate_img_cat(dt_masks=dt_masks,
gt_masks=gt_masks,
scores=scores)
thrs = evaluator.params.iouThrs
assert np.allclose(dtm[:, 0][thrs <= 0.5], 1)
assert np.allclose(dtm[:, 1:][thrs <= 0.5], 0)
assert np.allclose(dtm[:, 0][thrs > 0.5], 0)
assert np.allclose(dtm[:, 1][np.logical_and(thrs > 0.5, thrs <= 3 / 4)], 1)
assert np.allclose(dtm[:, 2][np.logical_and(thrs > 0.5, thrs <= 3 / 4)], 0)
assert np.allclose(dtm[:, 2][thrs > 3 / 4], 1)