def test_reorg():
inputs_shape = [1, 12, 12, 8]
inputs_np = np.random.uniform(-10., 10., size=inputs_shape).astype(np.float32)
inputs = tf.convert_to_tensor(inputs_np, dtype=tf.float32)
model = YoloV2(
anchors=[(1.0, 1.0), (1.5, 1.5)],
image_size=[128, 256],
batch_size=2,
num_max_boxes=5,
loss_warmup_steps=0,
)
outputs = model._reorg("reorg", inputs, stride=2, data_format="NHWC", use_space_to_depth=False)
tf.InteractiveSession()
outputs_np = outputs.eval()
assert outputs_np.shape == (inputs_shape[0], inputs_shape[1]/2, inputs_shape[2]/2, inputs_shape[3]*2*2,)
outputs2 = model._reorg("reorg", inputs, stride=2, data_format="NHWC", use_space_to_depth=True)
assert np.all(outputs_np == outputs2.eval())
def test_convert_boxes_space_inverse():
"""Test from_real_to_yolo is inverse function of from_yolo_to_real."""
model = YoloV2(
anchors=[(2.0, 4.0), (3.5, 4.5)],
image_size=[64, 96],
batch_size=1,
)
tf.InteractiveSession()
# shape is [batch_size, image_size[0]/32, image_size[1]/32, boxes_per_cell, 4(center_x, center_y, w, h)]
boxes = np.array([
[
[
[
[10, 20, 5, 8, ],
],
[
[5, 10, 40, 20, ],
],
[
[90, 32, 2, 10],
],
],
[
[
[1, 2, 30, 38, ],
],
[
[20, 40, 4, 4, ],
],
[
[90, 32, 2, 10],
],
],
],
])
boxes_tensor = tf.convert_to_tensor(boxes, dtype=tf.float32)
yolo_boxes = model.convert_boxes_space_from_real_to_yolo(boxes_tensor)
# real -> yolo -> real
reversed_boxes = model.convert_boxes_space_from_yolo_to_real(
model.convert_boxes_space_from_real_to_yolo(boxes_tensor)
)
assert np.allclose(reversed_boxes.eval(), boxes)
# yolo -> real -> yolo
yolo_boxes = model.convert_boxes_space_from_real_to_yolo(boxes_tensor)
reversed_boxes = model.convert_boxes_space_from_real_to_yolo(
model.convert_boxes_space_from_yolo_to_real(yolo_boxes)
)
assert np.allclose(reversed_boxes.eval(), yolo_boxes.eval())
def test_yolov2_post_process():
tf.InteractiveSession()
image_size = [96, 64]
batch_size = 2
classes = Pascalvoc2007.classes
anchors = [(0.1, 0.2), (1.2, 1.1)]
data_format = "NHWC"
score_threshold = 0.25
nms_iou_threshold = 0.5
model = YoloV2(
image_size=image_size,
batch_size=batch_size,
classes=classes,
anchors=anchors,
data_format=data_format,
score_threshold=score_threshold,
nms_iou_threshold=nms_iou_threshold,
)
post_process = Sequence([
FormatYoloV2(
image_size=image_size,
classes=classes,
anchors=anchors,
data_format=data_format,
),
ExcludeLowScoreBox(threshold=score_threshold),
NMS(
iou_threshold=nms_iou_threshold,
classes=classes,
),
])
shape = (batch_size, len(anchors) * (len(classes) + 5),
image_size[0] // 32, image_size[1] // 32)
np_output = np.random.uniform(-2., 2., size=shape).astype(np.float32)
output = tf.constant(np_output)
ys = model.post_process(output)
expected_ys = post_process(outputs=np_output)["outputs"]
for y, expected_y in zip(ys, expected_ys):
assert np.allclose(y.eval(), expected_y), (y.eval(), expected_y)
def test_offset_boxes():
"""Test from_real_to_yolo is inverse function of from_yolo_to_real."""
# shape is [batch_size, cell_size, cell_size, boxes_per_cell]
expected_x = np.array([
[
[
[0, 0],
[1, 1],
[2, 2],
[3, 3],
[4, 4],
],
[
[0, 0],
[1, 1],
[2, 2],
[3, 3],
[4, 4],
],
[
[0, 0],
[1, 1],
[2, 2],
[3, 3],
[4, 4],
],
[
[0, 0],
[1, 1],
[2, 2],
[3, 3],
[4, 4],
],
],
[
[
[0, 0],
[1, 1],
[2, 2],
[3, 3],
[4, 4],
],
[
[0, 0],
[1, 1],
[2, 2],
[3, 3],
[4, 4],
],
[
[0, 0],
[1, 1],
[2, 2],
[3, 3],
[4, 4],
],
[
[0, 0],
[1, 1],
[2, 2],
[3, 3],
[4, 4],
],
],
[
[
[0, 0],
[1, 1],
[2, 2],
[3, 3],
[4, 4],
],
[
[0, 0],
[1, 1],
[2, 2],
[3, 3],
[4, 4],
],
[
[0, 0],
[1, 1],
[2, 2],
[3, 3],
[4, 4],
],
[
[0, 0],
[1, 1],
[2, 2],
[3, 3],
[4, 4],
],
],
])
# shape is [batch_size, cell_size, cell_size, boxes_per_cell]
expected_y = np.array([
[[
[0, 0],
[0, 0],
[0, 0],
[0, 0],
[0, 0],
], [
[1, 1],
[1, 1],
[1, 1],
[1, 1],
[1, 1],
], [
[2, 2],
[2, 2],
[2, 2],
[2, 2],
[2, 2],
], [
[3, 3],
[3, 3],
[3, 3],
[3, 3],
[3, 3],
]],
[[
[0, 0],
[0, 0],
[0, 0],
[0, 0],
[0, 0],
], [
[1, 1],
[1, 1],
[1, 1],
[1, 1],
[1, 1],
], [
[2, 2],
[2, 2],
[2, 2],
[2, 2],
[2, 2],
], [
[3, 3],
[3, 3],
[3, 3],
[3, 3],
[3, 3],
]],
[[
[0, 0],
[0, 0],
[0, 0],
[0, 0],
[0, 0],
], [
[1, 1],
[1, 1],
[1, 1],
[1, 1],
[1, 1],
], [
[2, 2],
[2, 2],
[2, 2],
[2, 2],
[2, 2],
], [
[3, 3],
[3, 3],
[3, 3],
[3, 3],
[3, 3],
]],
])
anchors = [(4.0, 2.0), (3.5, 2.5)]
# shape is [batch_size, cell_size, cell_size, boxes_per_cell]
expected_w = np.broadcast_to(np.array([4.0, 3.5]), (3, 4, 5, 2))
# shape is [batch_size, cell_size, cell_size, boxes_per_cell]
expected_h = np.broadcast_to(np.array([2.0, 2.5]), (3, 4, 5, 2))
model = YoloV2(
anchors=anchors,
image_size=[128, 160],
batch_size=3,
is_dynamic_image_size=True,
)
tf.InteractiveSession()
offset_x, offset_y, offset_w, offset_h = model.offset_boxes()
# import ipdb; ipdb.set_trace()
assert np.all(offset_x.eval() == expected_x)
assert np.all(offset_x.eval()[:, :, 0, :] == 0)
assert np.all(offset_x.eval()[:, :, 1, :] == 1)
assert np.all(offset_x.eval()[:, :, 2, :] == 2)
assert np.all(offset_y.eval() == expected_y)
assert np.all(offset_y.eval()[:, 0, :, :] == 0)
assert np.all(offset_y.eval()[:, 1, :, :] == 1)
assert np.all(offset_y.eval()[:, 2, :, :] == 2)
assert np.all(offset_w.eval() == expected_w)
assert np.all(offset_w.eval()[:, :, :, 0] == anchors[0][0])
assert np.all(offset_w.eval()[:, :, :, 1] == anchors[1][0])
assert np.all(offset_h.eval() == expected_h)
assert np.all(offset_h.eval()[:, :, :, 0] == anchors[0][1])
assert np.all(offset_h.eval()[:, :, :, 1] == anchors[1][1])
def test_calculate_truth_and_masks():
model = YoloV2(
anchors=[(1.0, 1.0), (1.5, 1.5)],
image_size=[128, 256],
batch_size=2,
num_max_boxes=5,
loss_warmup_steps=0,
)
gt_boxes_list = [
[
[33, 43, 64, 50, 3],
[113, 62, 36, 26, 1],
[173, 53, 30, 32, 2],
[59, 93, 28, 18, 0],
[145, 79, 66, 67, 0],
],
[
[130, 67, 32, 22, 1],
[0, 0, 0, 0, -1],
[0, 0, 0, 0, -1],
[0, 0, 0, 0, -1],
[0, 0, 0, 0, -1],
],
]
expected_cell_gt_boxes = np.array([
[
[
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
],
[
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [33., 43., 64., 50., 3.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[113., 62., 36., 26., 1.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[173., 53., 30., 32., 2.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
],
[
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[59., 93., 28., 18., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [145., 79., 66., 67., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
],
[
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
],
],
[
[
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
],
[
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
],
[
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[130., 67., 32., 22., 1.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
],
[
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]],
],
],
])
expected_object_masks = [[
[
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
],
[
[[0.], [0.]],
[[0.], [1.]],
[[0.], [0.]],
[[1.], [0.]],
[[0.], [0.]],
[[1.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
],
[
[[0.], [0.]],
[[1.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [1.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
],
[
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
],
],
[
[
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
],
[
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
],
[
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[1.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
],
[
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
[[0.], [0.]],
],
]]
gt_boxes_list = tf.convert_to_tensor(gt_boxes_list, dtype=tf.float32)
# TODO(wakisaka): prepare numpy predict_boxes.
predict_boxes = tf.convert_to_tensor(expected_cell_gt_boxes,
dtype=tf.float32)
cell_gt_boxes, truth_confidence, object_masks, coordinate_masks =\
model.loss_function._calculate_truth_and_masks(gt_boxes_list, predict_boxes, global_step=0)
tf.InteractiveSession()
cell_gt_boxes_val = cell_gt_boxes.eval()
object_masks_val = object_masks.eval()
coordinate_masks_val = coordinate_masks.eval()
assert np.all(cell_gt_boxes_val == expected_cell_gt_boxes)
assert np.all(object_masks_val == expected_object_masks)
assert np.all(coordinate_masks_val == expected_object_masks)