def test_yolo_head_initialize_bias_class_probability(n_test: int = 5):
for n in range(n_test):
model = YOLOModel(os.path.join("tests", "test_configs",
"yolo_sample.yaml"),
verbose=n == 0)
class_probability = random.random()
n_object_per_image = (random.randint(4, 16), random.randint(320, 1024))
model.initialize_biases(class_probability=class_probability,
n_object_per_image=n_object_per_image)
class_prob = torch.log(
torch.tensor(class_probability /
(model.model[-1].n_classes - 0.99)))
for i in range(model.model[-1].n_layers):
class_bias = model.model[-1].conv[i].bias.view(3, -1)[:,
5:].mean(0)
assert torch.isclose(class_bias, class_prob, rtol=0.1).sum() == 10
for i in range(model.model[-1].n_layers):
obj_bias = model.model[-1].conv[i].bias.view(3, -1)[:, 4].mean()
obj_log = torch.log(
n_object_per_image[0] /
(n_object_per_image[1] / model.model[-1].stride[i])**2)
assert torch.isclose(obj_bias, obj_log, rtol=0.1)
def test_yolo_head_xyxy():
model = YOLOModel(
os.path.join("tests", "test_configs", "yolo_samplev2.yaml"),
verbose=True,
init_bias=True,
)
in_tensor = torch.rand((1, 3, 480, 380))
model.eval()
model.model[-1].out_xyxy = False
out_tensor_xywh = model(in_tensor)
model.model[-1].out_xyxy = True
out_tensor_xyxy = model(in_tensor)
xywh = out_tensor_xywh[0][0, :, :4]
xyxy = out_tensor_xyxy[0][0, :, :4]
x1y1_from_xywh = xywh[:, :2] - (xywh[:, 2:4] / 2)
x2y2_from_xywh = xywh[:, :2] + (xywh[:, 2:4] / 2)
wh_from_xyxy = xyxy[:, 2:] - xyxy[:, :2]
cxcy_from_xyxy = xyxy[:, :2] + (wh_from_xyxy / 2)
assert torch.isclose(torch.cat((x1y1_from_xywh, x2y2_from_xywh), -1),
xyxy).all()
assert torch.isclose(torch.cat((cxcy_from_xyxy, wh_from_xyxy), -1),
xywh).all()
def test_yolo_head():
model = YOLOModel(os.path.join("tests", "test_configs",
"yolo_sample.yaml"),
verbose=True)
model.initialize_biases()
in_tensor = torch.rand((1, 3, 480, 380))
out_tensor = model(in_tensor)
assert out_tensor[0].shape == (1, 3, 60, 48, 15)
assert out_tensor[1].shape == (1, 3, 30, 24, 15)
assert out_tensor[2].shape == (1, 3, 15, 12, 15)
model.eval()
out_tensor = model(in_tensor)
assert out_tensor[0].shape == (1, 11340, 15)
assert out_tensor[1][0].shape == (1, 3, 60, 48, 15)
assert out_tensor[1][1].shape == (1, 3, 30, 24, 15)
assert out_tensor[1][2].shape == (1, 3, 15, 12, 15)
profiler = model.profile(n_run=1)
n_param = profiler.get_parameter_numbers()
n_macs = profiler.get_macs()
assert (n_param == 7279367
), f"Number of parameter is not matched! Current: {n_param:,d}"
assert n_macs == 1350937664, f"Number of MACs is not matched! Current: {n_macs}"
def test_yolo_model_v2():
model = YOLOModel(
os.path.join("tests", "test_configs", "yolo_samplev2.yaml"),
verbose=True,
init_bias=True,
)
in_tensor = torch.rand((1, 3, 480, 380))
out_tensor = model(in_tensor)
assert out_tensor[0].shape == (1, 3, 60, 48, 15)
assert out_tensor[1].shape == (1, 3, 30, 24, 15)
assert out_tensor[2].shape == (1, 3, 15, 12, 15)
profiler = model.profile(n_run=1)
n_param = profiler.get_parameter_numbers()
n_macs = profiler.get_macs()
assert n_param == 7087815
assert n_macs == 1316596800
def test_yolo_model_mobilevit():
device = (torch.device("cuda:0")
if torch.cuda.is_available() else torch.device("cpu"))
model = YOLOModel(
os.path.join("tests", "test_configs", "yolo_mobilevit_sample.yaml"),
verbose=True,
init_bias=True,
).to(device)
in_tensor = torch.rand((1, 3, 512, 512)).to(device)
out_tensor = model(in_tensor)
assert out_tensor[0].shape == (1, 3, 64, 64, 15)
assert out_tensor[1].shape == (1, 3, 32, 32, 15)
assert out_tensor[2].shape == (1, 3, 16, 16, 15)
profiler = model.profile(n_run=100)
n_param = profiler.get_parameter_numbers()
n_macs = profiler.get_macs()
assert n_param == 4910455
assert n_macs == 1747833760
def test_yolo_head_initialize_bias(n_test: int = 5):
for n in range(n_test):
model = YOLOModel(os.path.join("tests", "test_configs",
"yolo_sample.yaml"),
verbose=n == 0)
class_frequency = torch.randint(100, 5000, (10, ))
n_object_per_image = (random.randint(4, 16), random.randint(320, 1024))
model.initialize_biases(class_frequency=class_frequency,
n_object_per_image=n_object_per_image)
freq_log = torch.log(class_frequency / class_frequency.sum())
for i in range(model.model[-1].n_layers):
class_bias = model.model[-1].conv[i].bias.view(3, -1)[:,
5:].mean(0)
assert torch.isclose(class_bias, freq_log, rtol=0.1).sum() == 10
for i in range(model.model[-1].n_layers):
obj_bias = model.model[-1].conv[i].bias.view(3, -1)[:, 4].mean()
obj_log = torch.log(
n_object_per_image[0] /
(n_object_per_image[1] / model.model[-1].stride[i])**2)
assert torch.isclose(obj_bias, obj_log, rtol=0.1)
def test_yolo_head_fused():
model = YOLOModel(os.path.join("tests", "test_configs",
"yolo_sample.yaml"),
verbose=True)
model.initialize_biases()
in_tensor = torch.rand((1, 3, 480, 380))
model.eval()
out_tensor = model(in_tensor)
model.fuse().eval()
out_tensor_fused = model(in_tensor)
assert torch.all(
torch.isclose(out_tensor[0], out_tensor_fused[0], rtol=1e-6))
for i in range(model.model[-1].n_layers):
assert torch.all(
torch.isclose(out_tensor[1][i], out_tensor_fused[1][i], rtol=1e-6))
def test_yolo_head_export():
model = YOLOModel(
os.path.join("tests", "test_configs", "yolo_samplev2.yaml"),
verbose=True,
init_bias=True,
)
in_tensor = torch.rand((1, 3, 480, 380))
model.eval()
out_tensor = model(in_tensor)
model.export(verbose=True)
out_tensor_export = model(in_tensor)
assert torch.all(
torch.isclose(out_tensor[0], out_tensor_export[0], rtol=1e-6))