File size: 11,716 Bytes
d3dbf03 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 |
# Copyright (c) OpenMMLab. All rights reserved.
import copy
import unittest
import numpy as np
import pytest
import torch
from mmengine.structures import InstanceData
from mmengine.testing import assert_dict_has_keys
from numpy.testing import assert_array_equal
from mmaction.datasets.transforms import (FormatAudioShape, FormatGCNInput,
FormatShape, PackActionInputs,
Transpose)
from mmaction.registry import TRANSFORMS
from mmaction.structures import ActionDataSample
from mmaction.utils import register_all_modules
register_all_modules()
class TestPackActionInputs(unittest.TestCase):
def test_transform(self):
# none input
with self.assertRaises(ValueError):
results = PackActionInputs()(dict())
# keypoint input
results = dict(keypoint=np.random.randn(2, 300, 17, 3), label=1)
transform = PackActionInputs()
results = transform(results)
self.assertIn('inputs', results)
self.assertIn('data_samples', results)
self.assertIsInstance(results['inputs'], torch.Tensor)
self.assertEqual(results['inputs'].shape, (2, 300, 17, 3))
self.assertEqual(results['data_samples'].gt_label,
torch.LongTensor([1]))
# heatmap_imgs input
results = dict(heatmap_imgs=np.random.randn(2, 17, 56, 56), label=1)
transform = PackActionInputs()
results = transform(results)
self.assertIn('inputs', results)
self.assertIn('data_samples', results)
self.assertIsInstance(results['inputs'], torch.Tensor)
self.assertEqual(results['inputs'].shape, (2, 17, 56, 56))
self.assertEqual(results['data_samples'].gt_label,
torch.LongTensor([1]))
# audios input
results = dict(audios=np.random.randn(3, 1, 128, 80), label=[1])
transform = PackActionInputs()
results = transform(results)
self.assertIn('inputs', results)
self.assertIn('data_samples', results)
self.assertEqual(results['inputs'].shape, (3, 1, 128, 80))
self.assertIsInstance(results['inputs'], torch.Tensor)
# text input
results = dict(text=np.random.randn(77))
transform = PackActionInputs()
results = transform(results)
self.assertIn('inputs', results)
self.assertIn('data_samples', results)
self.assertEqual(results['inputs'].shape, (77, ))
self.assertIsInstance(results['inputs'], torch.Tensor)
# imgs input with label
data = dict(
imgs=np.random.randn(2, 256, 256, 3),
label=[1],
filename='test.txt',
original_shape=(256, 256, 3),
img_shape=(256, 256, 3),
flip_direction='vertical')
transform = PackActionInputs()
results = transform(copy.deepcopy(data))
self.assertIn('inputs', results)
self.assertIn('data_samples', results)
self.assertIsInstance(results['inputs'], torch.Tensor)
self.assertIsInstance(results['data_samples'], ActionDataSample)
self.assertEqual(results['data_samples'].img_shape, (256, 256, 3))
self.assertEqual(results['data_samples'].gt_label,
torch.LongTensor([1]))
# Test grayscale image
data['imgs'] = data['imgs'].mean(-1)
results = transform(copy.deepcopy(data))
self.assertIn('inputs', results)
self.assertIsInstance(results['inputs'], torch.Tensor)
self.assertEqual(results['inputs'].shape, (2, 256, 256))
# imgs input with gt_bboxes
data = dict(
imgs=np.random.randn(256, 256, 3),
gt_bboxes=np.array([[0, 0, 340, 224]]),
gt_labels=[1],
proposals=np.array([[0, 0, 340, 224]]),
filename='test.txt')
transform = PackActionInputs()
results = transform(copy.deepcopy(data))
self.assertIn('inputs', results)
self.assertIsInstance(results['inputs'], torch.Tensor)
self.assertIn('data_samples', results)
self.assertIsInstance(results['data_samples'], ActionDataSample)
self.assertIsInstance(results['data_samples'].gt_instances,
InstanceData)
self.assertIsInstance(results['data_samples'].proposals, InstanceData)
# imgs and text input
data = dict(
imgs=np.random.randn(2, 256, 256, 3), text=np.random.randn(77))
transform = PackActionInputs(collect_keys=('imgs', 'text'))
results = transform(copy.deepcopy(data))
self.assertIn('inputs', results)
self.assertIn('data_samples', results)
self.assertIsInstance(results['inputs'], dict)
self.assertEqual(results['inputs']['imgs'].shape, (2, 256, 256, 3))
self.assertEqual(results['inputs']['text'].shape, (77, ))
def test_repr(self):
cfg = dict(
type='PackActionInputs', meta_keys=['flip_direction', 'img_shape'])
transform = TRANSFORMS.build(cfg)
self.assertEqual(
repr(transform), 'PackActionInputs(collect_keys=None, '
"meta_keys=['flip_direction', 'img_shape'])")
class TestPackLocalizationInputs(unittest.TestCase):
def test_transform(self):
# raw_feature input
data = dict(
raw_feature=np.random.randn(400, 5),
gt_bbox=np.array([[0.1, 0.3], [0.375, 0.625]]),
filename='test.txt')
cfg = dict(type='PackLocalizationInputs', keys=('gt_bbox', ))
transform = TRANSFORMS.build(cfg)
results = transform(copy.deepcopy(data))
self.assertIn('inputs', results)
self.assertIsInstance(results['inputs'], torch.Tensor)
self.assertIn('data_samples', results)
self.assertIsInstance(results['data_samples'], ActionDataSample)
self.assertIsInstance(results['data_samples'].gt_instances,
InstanceData)
del data['raw_feature']
with self.assertRaises(ValueError):
transform(copy.deepcopy(data))
# bsp_feature input
data['bsp_feature'] = np.random.randn(100, 32)
results = transform(copy.deepcopy(data))
self.assertIn('inputs', results)
self.assertIsInstance(results['inputs'], torch.Tensor)
self.assertIn('data_samples', results)
self.assertIsInstance(results['data_samples'], ActionDataSample)
self.assertIsInstance(results['data_samples'].gt_instances,
InstanceData)
def test_repr(self):
cfg = dict(
type='PackLocalizationInputs',
meta_keys=['video_name', 'feature_frame'])
transform = TRANSFORMS.build(cfg)
self.assertEqual(
repr(transform),
"PackLocalizationInputs(meta_keys=['video_name', 'feature_frame'])"
)
def test_transpose():
results = dict(imgs=np.random.randn(256, 256, 3))
keys = ['imgs']
order = [2, 0, 1]
transpose = Transpose(keys, order)
results = transpose(results)
assert results['imgs'].shape == (3, 256, 256)
assert repr(transpose) == transpose.__class__.__name__ + \
f'(keys={keys}, order={order})'
def test_format_shape():
with pytest.raises(ValueError):
# invalid input format
FormatShape('NHWC')
# 'NCHW' input format (RGB Modality)
results = dict(
imgs=np.random.randn(3, 224, 224, 3), num_clips=1, clip_len=3)
format_shape = FormatShape('NCHW')
assert format_shape(results)['input_shape'] == (3, 3, 224, 224)
# `NCHW` input format (Flow Modality)
results = dict(
imgs=np.random.randn(3, 224, 224, 2),
num_clips=1,
clip_len=3,
modality='Flow')
format_shape = FormatShape('NCHW')
assert format_shape(results)['input_shape'] == (1, 6, 224, 224)
# `NCTHW` input format with num_clips=1, clip_len=3
results = dict(
imgs=np.random.randn(3, 224, 224, 3), num_clips=1, clip_len=3)
format_shape = FormatShape('NCTHW')
assert format_shape(results)['input_shape'] == (1, 3, 3, 224, 224)
# `NCTHW` input format with num_clips=2, clip_len=3
results = dict(
imgs=np.random.randn(18, 224, 224, 3), num_clips=2, clip_len=3)
assert format_shape(results)['input_shape'] == (6, 3, 3, 224, 224)
target_keys = ['imgs', 'input_shape']
assert assert_dict_has_keys(results, target_keys)
# `NCTHW` input format with imgs and heatmap_imgs
results = dict(
imgs=np.random.randn(6, 224, 224, 3),
heatmap_imgs=np.random.randn(12, 17, 56, 56),
num_clips=2,
clip_len=dict(RGB=3, Pose=6))
results = format_shape(results)
assert results['input_shape'] == (2, 3, 3, 224, 224)
assert results['heatmap_input_shape'] == (2, 17, 6, 56, 56)
assert repr(format_shape) == "FormatShape(input_format='NCTHW')"
# `NCTHW_Heatmap` input format
results = dict(
imgs=np.random.randn(12, 17, 56, 56), num_clips=2, clip_len=6)
format_shape = FormatShape('NCTHW_Heatmap')
assert format_shape(results)['input_shape'] == (2, 17, 6, 56, 56)
# `NPTCHW` input format
results = dict(
imgs=np.random.randn(72, 224, 224, 3),
num_clips=9,
clip_len=1,
num_proposals=8)
format_shape = FormatShape('NPTCHW')
assert format_shape(results)['input_shape'] == (8, 9, 3, 224, 224)
def test_format_audio_shape():
with pytest.raises(ValueError):
# invalid input format
FormatAudioShape('XXXX')
# `NCTF` input format
results = dict(audios=np.random.randn(3, 128, 8))
format_shape = FormatAudioShape('NCTF')
assert format_shape(results)['input_shape'] == (3, 1, 128, 8)
assert repr(format_shape) == format_shape.__class__.__name__ + \
"(input_format='NCTF')"
def test_format_gcn_input():
with pytest.raises(AssertionError):
FormatGCNInput(mode='invalid')
results = dict(
keypoint=np.random.randn(2, 10, 17, 2),
keypoint_score=np.random.randn(2, 10, 17))
format_shape = FormatGCNInput(num_person=2, mode='zero')
results = format_shape(results)
assert results['keypoint'].shape == (1, 2, 10, 17, 3)
assert repr(format_shape) == 'FormatGCNInput(num_person=2, mode=zero)'
results = dict(keypoint=np.random.randn(2, 40, 25, 3), num_clips=4)
format_shape = FormatGCNInput(num_person=2, mode='zero')
results = format_shape(results)
assert results['keypoint'].shape == (4, 2, 10, 25, 3)
results = dict(keypoint=np.random.randn(1, 10, 25, 3))
format_shape = FormatGCNInput(num_person=2, mode='zero')
results = format_shape(results)
assert results['keypoint'].shape == (1, 2, 10, 25, 3)
assert_array_equal(results['keypoint'][:, 1], np.zeros((1, 10, 25, 3)))
results = dict(keypoint=np.random.randn(1, 10, 25, 3))
format_shape = FormatGCNInput(num_person=2, mode='loop')
results = format_shape(results)
assert results['keypoint'].shape == (1, 2, 10, 25, 3)
assert_array_equal(results['keypoint'][:, 1], results['keypoint'][:, 0])
results = dict(keypoint=np.random.randn(3, 10, 25, 3))
format_shape = FormatGCNInput(num_person=2, mode='zero')
results = format_shape(results)
assert results['keypoint'].shape == (1, 2, 10, 25, 3)
|