def test_write_npy(self):
from det3.ops import read_npy, write_npy
data = np.random.randn(1000).astype(np.float32)
path = "./unit-test/result/test_write_npy.npy"
write_npy(data, path)
est = read_npy(path)
gt = data
self.assertTrue(np.array_equal(gt, est))
def test_get_pts_idx(self):
from det3.ops import crop_pts_3drot, read_npy
from det3.dataloader.carladata import CarlaObj, CarlaObj, CarlaLabel, CarlaCalib
pts = read_npy("./unit-test/data/test_CarlaAugmentor_000250.npy")
calib = CarlaCalib(
"./unit-test/data/test_CarlaAugmentor_000250_calib.txt"
).read_calib_file()
label = CarlaLabel(
"./unit-test/data/test_CarlaAugmentor_000250_label_imu.txt"
).read_label_file()
pts = calib.lidar2imu(pts, key='Tr_imu_to_velo_top')
for i in range(len(label)):
obj = UdiObj(arr=np.array(label.bboxes3d[i,
[3, 4, 5, 2, 1, 0, 6]]),
cls="Car",
score=0.99)
idx = obj.get_pts_idx(pts)
gt = label.data[i].get_pts_idx(pts)
gt = np.where(gt)[0].flatten()
self.assertTrue(set(idx.tolist()) == set(gt.tolist()))
def read_data(self):
'''
-> res: dict
res[calib]: LyftCalib / None
res[label]: LyftLabel / None
res[pc]: dict {lidar: np.ndarray (#pts, >=3) (in Flidar)} / None
'''
calib = LyftCalib(self._calib_path).read_calib_file() if self._output_dict["calib"] else None
label = LyftLabel(self._label_path).read_label_file() if self._output_dict["label"] else None
if self._output_dict["lidar"]:
pc = dict()
for k, v in self._lidar_paths.items():
pc[k] = read_npy(v).reshape(-1, 4)
else:
pc = None
res = {k: None for k in self._output_dict.keys()}
res["calib"] = calib
res["label"] = label
res["lidar"] = pc
return res
def test_corner_transformation(self):
from det3.ops import crop_pts_3drot, read_npy
from det3.dataloader.carladata import CarlaObj, CarlaObj, CarlaLabel, CarlaCalib
pts = read_npy("./unit-test/data/test_CarlaAugmentor_000250.npy")
calib = CarlaCalib(
"./unit-test/data/test_CarlaAugmentor_000250_calib.txt"
).read_calib_file()
label = CarlaLabel(
"./unit-test/data/test_CarlaAugmentor_000250_label_imu.txt"
).read_label_file()
pts = calib.lidar2imu(pts, key='Tr_imu_to_velo_top')
for i in range(len(label)):
obj = UdiObj(arr=np.array(label.bboxes3d[i,
[3, 4, 5, 2, 1, 0, 6]]),
cls="Car",
score=0.99)
cns = obj.get_bbox3d_corners()
gt = label.data[i].get_bbox3dcorners()
self.assertTrue(np.allclose(cns, gt, atol=1e-2))
obj_cp = UdiObj()
obj_cp.from_corners(cns, obj.cls, obj.score)
self.assertTrue(obj_cp.equal(obj))
def test_crop_pts_3drot(self):
print("test_crop_pts_3drot")
from det3.ops import crop_pts_3drot, read_npy
from det3.dataloader.carladata import CarlaObj, CarlaObj, CarlaLabel, CarlaCalib
pts = read_npy("./unit-test/data/test_CarlaAugmentor_000250.npy")
calib = CarlaCalib(
"./unit-test/data/test_CarlaAugmentor_000250_calib.txt"
).read_calib_file()
label = CarlaLabel(
"./unit-test/data/test_CarlaAugmentor_000250_label_imu.txt"
).read_label_file()
pts = calib.lidar2imu(pts, key='Tr_imu_to_velo_top')
for test_dtype in [np.float32, np.float64]:
# obj.h, obj.w, obj.l, obj.x, obj.y, obj.z, obj.ry
boxes = label.bboxes3d[:, [3, 4, 5, 2, 1, 0, 6]].astype(test_dtype)
pts = pts.astype(test_dtype)
gt = [obj.get_pts_idx(pts) for obj in label.data]
gt = [np.where(itm)[0].flatten() for itm in gt]
# np
est = crop_pts_3drot(boxes, pts)
self.assertTrue(len(gt) == len(est))
for gt_, est_ in zip(gt, est):
set_gt = set(gt_.tolist())
set_est = set(est_.tolist())
self.assertTrue(set_gt == set_est)
times = 1000
t1 = time.time()
for i in range(times):
est = crop_pts_3drot(boxes, pts)
t = (time.time() - t1) / times * 1000
print(f"np {times}: {t:.3f} ms {est[0].dtype}")
# torch
pts_ts = torch.from_numpy(pts)
boxes_ts = torch.from_numpy(boxes)
est_ts = crop_pts_3drot(boxes_ts, pts_ts)
self.assertTrue(len(gt) == len(est_ts))
for gt_, est_ts_ in zip(gt, est_ts):
set_gt = set(gt_.tolist())
set_est_ts = set(est_ts_.numpy().tolist())
self.assertTrue(set_gt == set_est_ts)
t1 = time.time()
for i in range(times):
est_ts = crop_pts_3drot(boxes_ts, pts_ts)
t = (time.time() - t1) / times * 1000
print(f"torchcpu {times}: {t:.3f} ms {est_ts[0].dtype}")
# torchgpu
pts_tsgpu = torch.from_numpy(pts).cuda()
boxes_tsgpu = torch.from_numpy(boxes).cuda()
est_tsgpu = crop_pts_3drot(boxes_tsgpu, pts_tsgpu)
self.assertTrue(len(gt) == len(est_tsgpu))
for gt_, est_tsgpu_ in zip(gt, est_tsgpu):
set_gt = set(gt_.tolist())
set_est_tsgpu = set(est_tsgpu_.cpu().numpy().tolist())
self.assertTrue(set_gt == set_est_tsgpu)
torch.cuda.synchronize()
t1 = time.time()
for i in range(times):
est_tsgpu = crop_pts_3drot(boxes_tsgpu, pts_tsgpu)
torch.cuda.synchronize()
t = (time.time() - t1) / times * 1000
print(f"torchgpu {times}: {t:.3f} ms {est_tsgpu[0].dtype}")
def test_read_npy(self):
from det3.ops import read_npy, write_npy
gt = np.random.randn(1000).astype(np.float32)
np.save("./unit-test/result/test_read_npy.npy", gt)
est = read_npy("./unit-test/result/test_read_npy.npy")
self.assertTrue(np.array_equal(gt, est))