def test_label_invariance(self, label0: int, label1: int,
shuffle_inds: List[int]):
""" Ensure that datum ordering doesn't matter for generate_targets. """
# xyxy format
prop = np.array([
[-0.5, -0.5, 0.5,
0.5], # iou = 1 (truth 0) should be marked poitiive
[0.0, -0.5, 0.49,
0.5], # iou = 0.5 (truth 0) should be marked ignore
[0.0, -0.5, 0.39,
0.5], # iou = 0.39 (truth 0) should be marked negative
[10.0, 10.0, 11, 11],
]) # iou = 1 (truth 1) should be marked positive
# xyxy format
truth = np.array([[-0.5, -0.5, 0.5, 0.5], [10.0, 10.0, 11, 11]])
labels = np.array([label0, label1])
out_labels = np.array(
[label0, -1, 0,
label1]) # truth 0 / ignore / background / truth 1 from above
labels, reg = generate_targets(prop[shuffle_inds], truth, labels, 0.5,
0.4)
msg = "generate_targets is not invariant to datum-ordering"
assert_allclose(actual=labels,
desired=out_labels[shuffle_inds],
err_msg=msg)
def test_identical_proposed_and_truth(self, x: ndarray):
""" Ensure that generate_targets produces regression targets that are zero for identical proposal and truth. """
x = x.cumsum(axis=1) # ensure (l, t, r , b)
labels = np.array([0] * 5)
_, reg = generate_targets(x, x, labels, 0.5, 0.4)
msg = "generate_targets failed to produce the expected output when the proposed boxes are identical to ground truth"
assert_allclose(actual=reg,
desired=np.zeros_like(x),
atol=1e-5,
rtol=1e-5,
err_msg=msg)
def val_dataloader(self) -> DataLoader:
val_cls_targs, val_reg_targs = zip(
*(generate_targets(self.anchor_boxes, bxs, lbls, 0.2, 0.1)
for bxs, lbls in zip(self.val_boxes, self.val_labels)))
val_reg_targs = tr.tensor(val_reg_targs).float()
val_cls_targs = tr.tensor(val_cls_targs).long()
return DataLoader(
TensorDataset(self.val_images, val_cls_targs, val_reg_targs),
batch_size=16,
pin_memory=True,
num_workers=4,
shuffle=False,
drop_last=True,
)
def test_shapes(self, boxes: ndarray, truth: ndarray,
data: st.SearchStrategy):
""" Ensure the shape returned by generate_targets is correct, even in edge cases producing empty arrays. """
boxes = boxes.cumsum(
axis=1) # to ensure we don't hit 0-width or -height boxes
truth = truth.cumsum(
axis=1) # to ensure we don't hit 0-width or -height boxes
N = boxes.shape[0]
K = truth.shape[0]
labels = data.draw(hnp.arrays(dtype=int, shape=(K, )))
cls, reg = generate_targets(boxes, truth, labels, 0.5, 0.4)
msg = "generate_targets failed to produce classification targets of the correct shape"
assert cls.shape == (N, ), msg
msg = "generate_targets failed to produce regression targets of the correct shape"
assert reg.shape == (N, 4), msg
def test_known_regression_values(self, shuffle_inds: List[int]):
""" Ensure that generate_targets works for known values. Ensure that datum ordering does not matter. """
prop = np.array([
[-0.5, -0.5, 0.5, 0.5], # neither axis matches truth
[0, -0.5, np.exp(1), 0.5], # x matches truth
[-0.5, 0, 0.5, np.exp(1)], # y matches truth
])
truth = np.array([[0, 0, np.exp(1), np.exp(1)]])
labels = np.array([1.0])
out_reg = np.array([
[np.exp(1) / 2, np.exp(1) / 2, 1, 1],
[0, np.exp(1) / 2, 0, 1],
[np.exp(1) / 2, 0, 1, 0],
])
_, reg = generate_targets(prop[shuffle_inds], truth, labels, 0.5, 0.4)
msg = "generate_targets failed to produce known-correct regression values"
assert_allclose(actual=reg,
desired=out_reg[shuffle_inds],
atol=1e-5,
rtol=1e-5,
err_msg=msg)