def test_yolov3_neck():
# num_scales, in_channels, out_channels must be same length
with pytest.raises(AssertionError):
YOLOV3Neck(num_scales=3, in_channels=[16, 8, 4], out_channels=[8, 4])
# len(feats) must equal to num_scales
with pytest.raises(AssertionError):
neck = YOLOV3Neck(num_scales=3,
in_channels=[16, 8, 4],
out_channels=[8, 4, 2])
feats = (torch.rand(1, 4, 16, 16), torch.rand(1, 8, 16, 16))
neck(feats)
# test normal channels
s = 32
in_channels = [16, 8, 4]
out_channels = [8, 4, 2]
feat_sizes = [s // 2**i for i in range(len(in_channels) - 1, -1, -1)]
feats = [
torch.rand(1, in_channels[i], feat_sizes[i], feat_sizes[i])
for i in range(len(in_channels) - 1, -1, -1)
]
neck = YOLOV3Neck(num_scales=3,
in_channels=in_channels,
out_channels=out_channels)
outs = neck(feats)
assert len(outs) == len(feats)
for i in range(len(outs)):
assert outs[i].shape == \
(1, out_channels[i], feat_sizes[i], feat_sizes[i])
# test more flexible setting
s = 32
in_channels = [32, 8, 16]
out_channels = [19, 21, 5]
feat_sizes = [s // 2**i for i in range(len(in_channels) - 1, -1, -1)]
feats = [
torch.rand(1, in_channels[i], feat_sizes[i], feat_sizes[i])
for i in range(len(in_channels) - 1, -1, -1)
]
neck = YOLOV3Neck(num_scales=3,
in_channels=in_channels,
out_channels=out_channels)
outs = neck(feats)
assert len(outs) == len(feats)
for i in range(len(outs)):
assert outs[i].shape == \
(1, out_channels[i], feat_sizes[i], feat_sizes[i])
def yolo_neck_config(test_step_name):
"""Config yolov3 Neck."""
in_channels = [16, 8, 4]
out_channels = [8, 4, 2]
# The data of yolov3_neck.pkl contains a list of
# torch.Tensor, where each torch.Tensor is generated by
# torch.rand and each tensor size is:
# (1, 4, 64, 64), (1, 8, 32, 32), (1, 16, 16, 16).
yolov3_neck_data = 'yolov3_neck.pkl'
feats = mmcv.load(osp.join(data_path, yolov3_neck_data))
if (yolo_test_step_names[test_step_name] == 0):
yolo_model = YOLOV3Neck(
in_channels=in_channels, out_channels=out_channels, num_scales=3)
return yolo_model, feats