def setUp(self):
feat_stride = 4
C = 16
H = 8
W = 12
self.proposal_creator_params = {
'n_train_post_nms': 10,
'n_test_post_nms': 5}
self.ratios = [0.25, 4]
self.anchor_scales = [2, 4]
self.link = RegionProposalNetwork(
in_channels=C, mid_channels=24,
ratios=self.ratios, anchor_scales=self.anchor_scales,
feat_stride=feat_stride,
proposal_creator_params=self.proposal_creator_params
)
self.x = np.random.uniform(size=(self.B, C, H, W)).astype(np.float32)
self.img_size = (H * feat_stride, W * feat_stride)
def setUp(self):
feat_stride = 4
self.B = 2
C = 16
H = 8
W = 12
self.proposal_creator_params = {
'n_train_post_nms': 10,
'n_test_post_nms': 5}
self.ratios = [0.25, 4]
self.anchor_scales = [2, 4]
self.link = RegionProposalNetwork(
in_channels=C, mid_channels=24,
ratios=self.ratios, anchor_scales=self.anchor_scales,
feat_stride=feat_stride,
proposal_creator_params=self.proposal_creator_params
)
self.x = np.random.uniform(size=(self.B, C, H, W)).astype(np.float32)
self.img_size = (H * feat_stride, W * feat_stride)
chainer.config.train = self.train
class TestRegionProposalNetwork(unittest.TestCase):
def setUp(self):
feat_stride = 4
self.B = 2
C = 16
H = 8
W = 12
self.proposal_creator_params = {
'n_train_post_nms': 10,
'n_test_post_nms': 5}
self.ratios = [0.25, 4]
self.anchor_scales = [2, 4]
self.link = RegionProposalNetwork(
in_channels=C, mid_channels=24,
ratios=self.ratios, anchor_scales=self.anchor_scales,
feat_stride=feat_stride,
proposal_creator_params=self.proposal_creator_params
)
self.x = np.random.uniform(size=(self.B, C, H, W)).astype(np.float32)
self.img_size = (H * feat_stride, W * feat_stride)
chainer.config.train = self.train
def _check_call(self, x, img_size):
_, _, H, W = x.shape
rpn_locs, rpn_scores, rois, roi_indices, anchor = self.link(
chainer.Variable(x), img_size)
self.assertIsInstance(rpn_locs, chainer.Variable)
self.assertIsInstance(rpn_locs.array, type(x))
self.assertIsInstance(rpn_scores, chainer.Variable)
self.assertIsInstance(rpn_scores.array, type(x))
A = len(self.ratios) * len(self.anchor_scales)
self.assertEqual(rpn_locs.shape, (self.B, H * W * A, 4))
self.assertEqual(rpn_scores.shape, (self.B, H * W * A, 2))
if chainer.config.train:
roi_size = self.proposal_creator_params[
'n_train_post_nms']
else:
roi_size = self.proposal_creator_params[
'n_test_post_nms']
self.assertIsInstance(rois, type(x))
self.assertIsInstance(roi_indices, type(x))
self.assertLessEqual(rois.shape[0], self.B * roi_size)
self.assertLessEqual(roi_indices.shape[0], self.B * roi_size)
# Depending randomly generated bounding boxes, this is not true.
if roi_indices.shape[0] == self.B * roi_size:
for i in range(self.B):
s = slice(i * roi_size, (i + 1) * roi_size)
np.testing.assert_equal(
cuda.to_cpu(roi_indices[s]),
i * np.ones((roi_size,), dtype=np.int32))
self.assertIsInstance(anchor, type(x))
self.assertEqual(anchor.shape, (A * H * W, 4))
def test_call_cpu(self):
self._check_call(self.x, self.img_size)
@attr.gpu
def test_call_gpu(self):
self.link.to_gpu()
self._check_call(
chainer.cuda.to_gpu(self.x), self.img_size)
class TestRegionProposalNetwork(unittest.TestCase):
def setUp(self):
feat_stride = 4
C = 16
H = 8
W = 12
self.proposal_creator_params = {
'n_train_post_nms': 10,
'n_test_post_nms': 5}
self.ratios = [0.25, 4]
self.anchor_scales = [2, 4]
self.link = RegionProposalNetwork(
in_channels=C, mid_channels=24,
ratios=self.ratios, anchor_scales=self.anchor_scales,
feat_stride=feat_stride,
proposal_creator_params=self.proposal_creator_params
)
self.x = np.random.uniform(size=(self.B, C, H, W)).astype(np.float32)
self.img_size = (H * feat_stride, W * feat_stride)
def _check_call(self, x, img_size, scales):
_, _, H, W = x.shape
with chainer.using_config('train', self.train):
rpn_locs, rpn_scores, rois, roi_indices, anchor = self.link(
chainer.Variable(x), img_size, scales)
self.assertIsInstance(rpn_locs, chainer.Variable)
self.assertIsInstance(rpn_locs.array, type(x))
self.assertIsInstance(rpn_scores, chainer.Variable)
self.assertIsInstance(rpn_scores.array, type(x))
A = len(self.ratios) * len(self.anchor_scales)
self.assertEqual(rpn_locs.shape, (self.B, H * W * A, 4))
self.assertEqual(rpn_scores.shape, (self.B, H * W * A, 2))
if self.train:
roi_size = self.proposal_creator_params[
'n_train_post_nms']
else:
roi_size = self.proposal_creator_params[
'n_test_post_nms']
self.assertIsInstance(rois, type(x))
self.assertIsInstance(roi_indices, type(x))
self.assertLessEqual(rois.shape[0], self.B * roi_size)
self.assertLessEqual(roi_indices.shape[0], self.B * roi_size)
# Depending randomly generated bounding boxes, this is not true.
if roi_indices.shape[0] == self.B * roi_size:
for i in range(self.B):
s = slice(i * roi_size, (i + 1) * roi_size)
np.testing.assert_equal(
cuda.to_cpu(roi_indices[s]),
i * np.ones((roi_size,), dtype=np.int32))
self.assertIsInstance(anchor, type(x))
self.assertEqual(anchor.shape, (A * H * W, 4))
def test_call_cpu(self):
self._check_call(self.x, self.img_size, self.scales)
@attr.gpu
def test_call_gpu(self):
self.link.to_gpu()
self._check_call(
chainer.backends.cuda.to_gpu(self.x), self.img_size, self.scales)