1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
|
- import os.path
- import tempfile
- import unittest
- import numpy as np
- import onnx
- import onnxruntime as ort
- import torch.jit
- from super_gradients.training.datasets.data_formats.bbox_formats import NormalizedXYWHCoordinateFormat, CXCYWHCoordinateFormat, YXYXCoordinateFormat
- from super_gradients.training.datasets.data_formats.output_adapters.detection_adapter import DetectionOutputAdapter
- from super_gradients.training.datasets.data_formats import (
- ConcatenatedTensorFormat,
- BoundingBoxesTensorSliceItem,
- TensorSliceItem,
- )
- NORMALIZED_XYWH_SCORES_LABELS = ConcatenatedTensorFormat(
- layout=(
- BoundingBoxesTensorSliceItem(name="bboxes", format=NormalizedXYWHCoordinateFormat()),
- TensorSliceItem(length=1, name="scores"),
- TensorSliceItem(length=1, name="labels"),
- )
- )
- CXCYWH_SCORES_LABELS = ConcatenatedTensorFormat(
- layout=(
- BoundingBoxesTensorSliceItem(name="bboxes", format=CXCYWHCoordinateFormat()),
- TensorSliceItem(length=1, name="scores"),
- TensorSliceItem(length=1, name="labels"),
- )
- )
- CXCYWH_LABELS_SCORES_DISTANCE_ATTR = ConcatenatedTensorFormat(
- layout=(
- BoundingBoxesTensorSliceItem(name="bboxes", format=CXCYWHCoordinateFormat()),
- TensorSliceItem(length=1, name="labels"),
- TensorSliceItem(length=1, name="scores"),
- TensorSliceItem(length=1, name="distance"),
- TensorSliceItem(length=4, name="attributes"),
- )
- )
- ATTR_YXYX = ConcatenatedTensorFormat(
- layout=(
- TensorSliceItem(length=4, name="attributes"),
- BoundingBoxesTensorSliceItem(name="bboxes", format=YXYXCoordinateFormat()),
- )
- )
- class TestDetectionOutputAdapter(unittest.TestCase):
- @torch.no_grad()
- def test_select_only_some_outputs(self):
- adapter = DetectionOutputAdapter(CXCYWH_LABELS_SCORES_DISTANCE_ATTR, ATTR_YXYX, image_shape=(640, 640)).eval()
- example_inputs = (
- torch.randn((300, CXCYWH_LABELS_SCORES_DISTANCE_ATTR.num_channels)),
- torch.randn((4, 300, CXCYWH_LABELS_SCORES_DISTANCE_ATTR.num_channels)),
- )
- for expected_input in example_inputs:
- intermediate = adapter(expected_input)
- self.assertEqual(ATTR_YXYX.num_channels, intermediate.size(-1))
- @torch.no_grad()
- def test_output_adapter_convert_vice_versa(self):
- adapter = DetectionOutputAdapter(NORMALIZED_XYWH_SCORES_LABELS, CXCYWH_SCORES_LABELS, image_shape=(640, 640)).eval()
- adapter_back = DetectionOutputAdapter(CXCYWH_SCORES_LABELS, NORMALIZED_XYWH_SCORES_LABELS, image_shape=(640, 640)).eval()
- example_inputs = (
- torch.randn((300, NORMALIZED_XYWH_SCORES_LABELS.num_channels)),
- torch.randn((4, 300, NORMALIZED_XYWH_SCORES_LABELS.num_channels)),
- )
- for expected_input in example_inputs:
- intermediate = adapter(expected_input)
- output_actual = adapter_back(intermediate)
- self.assertTrue(torch.allclose(expected_input, output_actual, atol=1e-4))
- @torch.no_grad()
- def test_output_adapter_can_be_traced(self):
- adapter = DetectionOutputAdapter(NORMALIZED_XYWH_SCORES_LABELS, CXCYWH_SCORES_LABELS, image_shape=(640, 640)).eval()
- example_inputs = (
- torch.randn((300, NORMALIZED_XYWH_SCORES_LABELS.num_channels)),
- torch.randn((4, 300, NORMALIZED_XYWH_SCORES_LABELS.num_channels)),
- )
- for inp in example_inputs:
- traced_adapter = torch.jit.trace(adapter, example_inputs=inp, strict=True)
- output_expected = adapter(inp)
- output_actual = traced_adapter(inp)
- self.assertTrue(output_expected.eq(output_actual).all())
- @torch.no_grad()
- def test_output_adapter_can_be_scripted(self):
- adapter = DetectionOutputAdapter(NORMALIZED_XYWH_SCORES_LABELS, CXCYWH_SCORES_LABELS, image_shape=(640, 640)).eval()
- example_inputs = (
- torch.randn((300, NORMALIZED_XYWH_SCORES_LABELS.num_channels)),
- torch.randn((4, 300, NORMALIZED_XYWH_SCORES_LABELS.num_channels)),
- )
- for inp in example_inputs:
- scripted_adapter = torch.jit.script(adapter, example_inputs=[inp])
- output_expected = adapter(inp)
- output_actual = scripted_adapter(inp)
- self.assertTrue(output_expected.eq(output_actual).all())
- @torch.no_grad()
- def test_output_adapter_can_be_onnx_exported(self):
- adapter = DetectionOutputAdapter(NORMALIZED_XYWH_SCORES_LABELS, CXCYWH_SCORES_LABELS, image_shape=(640, 640)).eval()
- example_inputs = (
- torch.randn((300, NORMALIZED_XYWH_SCORES_LABELS.num_channels)),
- torch.randn((4, 300, NORMALIZED_XYWH_SCORES_LABELS.num_channels)),
- )
- for inp in example_inputs:
- expected_output = adapter(inp).numpy()
- with tempfile.TemporaryDirectory() as tmpdirname:
- adapter_fname = os.path.join(tmpdirname, "adapter.onnx")
- torch.onnx.export(adapter, inp, f=adapter_fname, input_names=["predictions"], output_names=["output_predictions"], opset_version=11)
- onnx_model = onnx.load(adapter_fname)
- onnx.checker.check_model(onnx_model)
- ort_sess = ort.InferenceSession(adapter_fname)
- actual_output = ort_sess.run(None, {"predictions": inp.numpy()})[0]
- np.testing.assert_allclose(actual_output, expected_output)
- if __name__ == "__main__":
- unittest.main()
|