Spaces:
Sleeping
Sleeping
File size: 10,370 Bytes
b291f6a |
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 |
# Copyright (c) OpenMMLab. All rights reserved.
from unittest import TestCase
import pytest
import torch
from mmengine.config import Config
from mmengine.structures import InstanceData
from mmyolo.models.dense_heads import RTMDetRotatedHead
from mmyolo.utils import register_all_modules
register_all_modules()
class TestRTMDetRotatedHead(TestCase):
def setUp(self):
self.head_module = dict(
type='RTMDetRotatedSepBNHeadModule',
num_classes=4,
in_channels=1,
stacked_convs=1,
feat_channels=64,
featmap_strides=[4, 8, 16])
def test_init_weights(self):
head = RTMDetRotatedHead(head_module=self.head_module)
head.head_module.init_weights()
def test_predict_by_feat(self):
s = 256
img_metas = [{
'img_shape': (s, s, 3),
'ori_shape': (s, s, 3),
'scale_factor': (1.0, 1.0),
}]
test_cfg = dict(
multi_label=True,
decode_with_angle=True,
nms_pre=2000,
score_thr=0.01,
nms=dict(type='nms_rotated', iou_threshold=0.1),
max_per_img=300)
test_cfg = Config(test_cfg)
head = RTMDetRotatedHead(
head_module=self.head_module, test_cfg=test_cfg)
feat = [
torch.rand(1, 1, s // feat_size, s // feat_size)
for feat_size in [4, 8, 16]
]
cls_scores, bbox_preds, angle_preds = head.forward(feat)
head.predict_by_feat(
cls_scores,
bbox_preds,
angle_preds,
batch_img_metas=img_metas,
cfg=test_cfg,
rescale=True,
with_nms=True)
head.predict_by_feat(
cls_scores,
bbox_preds,
angle_preds,
batch_img_metas=img_metas,
cfg=test_cfg,
rescale=False,
with_nms=False)
def test_loss_by_feat(self):
if not torch.cuda.is_available():
pytest.skip('test requires GPU and torch+cuda')
s = 256
img_metas = [{
'img_shape': (s, s, 3),
'batch_input_shape': (s, s),
'scale_factor': 1,
}]
train_cfg = dict(
assigner=dict(
type='BatchDynamicSoftLabelAssigner',
num_classes=80,
topk=13,
iou_calculator=dict(type='mmrotate.RBboxOverlaps2D'),
batch_iou=False),
allowed_border=-1,
pos_weight=-1,
debug=False)
train_cfg = Config(train_cfg)
head = RTMDetRotatedHead(
head_module=self.head_module, train_cfg=train_cfg).cuda()
feat = [
torch.rand(1, 1, s // feat_size, s // feat_size).cuda()
for feat_size in [4, 8, 16]
]
cls_scores, bbox_preds, angle_preds = head.forward(feat)
# Test that empty ground truth encourages the network to predict
# background
gt_instances = InstanceData(
bboxes=torch.empty((0, 5)).cuda(),
labels=torch.LongTensor([]).cuda())
empty_gt_losses = head.loss_by_feat(cls_scores, bbox_preds,
angle_preds, [gt_instances],
img_metas)
# When there is no truth, the cls loss should be nonzero but there
# should be no box loss.
empty_cls_loss = empty_gt_losses['loss_cls'].sum()
empty_box_loss = empty_gt_losses['loss_bbox'].sum()
self.assertGreater(empty_cls_loss.item(), 0,
'classification loss should be non-zero')
self.assertEqual(
empty_box_loss.item(), 0,
'there should be no box loss when there are no true boxes')
# When truth is non-empty then both cls and box loss should be nonzero
# for random inputs
head = RTMDetRotatedHead(
head_module=self.head_module, train_cfg=train_cfg).cuda()
gt_instances = InstanceData(
bboxes=torch.Tensor([[130.6667, 86.8757, 100.6326, 70.8874,
0.2]]).cuda(),
labels=torch.LongTensor([1]).cuda())
one_gt_losses = head.loss_by_feat(cls_scores, bbox_preds, angle_preds,
[gt_instances], img_metas)
onegt_cls_loss = one_gt_losses['loss_cls'].sum()
onegt_box_loss = one_gt_losses['loss_bbox'].sum()
self.assertGreater(onegt_cls_loss.item(), 0,
'cls loss should be non-zero')
self.assertGreater(onegt_box_loss.item(), 0,
'box loss should be non-zero')
# test num_class = 1
self.head_module['num_classes'] = 1
head = RTMDetRotatedHead(
head_module=self.head_module, train_cfg=train_cfg).cuda()
gt_instances = InstanceData(
bboxes=torch.Tensor([[130.6667, 86.8757, 100.6326, 70.8874,
0.2]]).cuda(),
labels=torch.LongTensor([0]).cuda())
cls_scores, bbox_preds, angle_preds = head.forward(feat)
one_gt_losses = head.loss_by_feat(cls_scores, bbox_preds, angle_preds,
[gt_instances], img_metas)
onegt_cls_loss = one_gt_losses['loss_cls'].sum()
onegt_box_loss = one_gt_losses['loss_bbox'].sum()
self.assertGreater(onegt_cls_loss.item(), 0,
'cls loss should be non-zero')
self.assertGreater(onegt_box_loss.item(), 0,
'box loss should be non-zero')
def test_hbb_loss_by_feat(self):
s = 256
img_metas = [{
'img_shape': (s, s, 3),
'batch_input_shape': (s, s),
'scale_factor': 1,
}]
train_cfg = dict(
assigner=dict(
type='BatchDynamicSoftLabelAssigner',
num_classes=80,
topk=13,
iou_calculator=dict(type='mmrotate.RBboxOverlaps2D'),
batch_iou=False),
allowed_border=-1,
pos_weight=-1,
debug=False)
train_cfg = Config(train_cfg)
hbb_cfg = dict(
bbox_coder=dict(
type='DistanceAnglePointCoder', angle_version='le90'),
loss_bbox=dict(type='mmdet.GIoULoss', loss_weight=2.0),
angle_coder=dict(
type='mmrotate.CSLCoder',
angle_version='le90',
omega=1,
window='gaussian',
radius=1),
loss_angle=dict(
type='mmrotate.SmoothFocalLoss',
gamma=2.0,
alpha=0.25,
loss_weight=0.2),
use_hbbox_loss=True,
)
head = RTMDetRotatedHead(
head_module=self.head_module, **hbb_cfg, train_cfg=train_cfg)
feat = [
torch.rand(1, 1, s // feat_size, s // feat_size)
for feat_size in [4, 8, 16]
]
cls_scores, bbox_preds, angle_preds = head.forward(feat)
# Test that empty ground truth encourages the network to predict
# background
gt_instances = InstanceData(
bboxes=torch.empty((0, 5)), labels=torch.LongTensor([]))
empty_gt_losses = head.loss_by_feat(cls_scores, bbox_preds,
angle_preds, [gt_instances],
img_metas)
# When there is no truth, the cls loss should be nonzero but there
# should be no box loss.
empty_cls_loss = empty_gt_losses['loss_cls'].sum()
empty_box_loss = empty_gt_losses['loss_bbox'].sum()
empty_angle_loss = empty_gt_losses['loss_angle'].sum()
self.assertGreater(empty_cls_loss.item(), 0,
'classification loss should be non-zero')
self.assertEqual(
empty_box_loss.item(), 0,
'there should be no box loss when there are no true boxes')
self.assertEqual(
empty_angle_loss.item(), 0,
'there should be no angle loss when there are no true boxes')
# When truth is non-empty then both cls and box loss should be nonzero
# for random inputs
head = RTMDetRotatedHead(
head_module=self.head_module, **hbb_cfg, train_cfg=train_cfg)
gt_instances = InstanceData(
bboxes=torch.Tensor([[130.6667, 86.8757, 100.6326, 70.8874, 0.2]]),
labels=torch.LongTensor([1]))
one_gt_losses = head.loss_by_feat(cls_scores, bbox_preds, angle_preds,
[gt_instances], img_metas)
onegt_cls_loss = one_gt_losses['loss_cls'].sum()
onegt_box_loss = one_gt_losses['loss_bbox'].sum()
onegt_angle_loss = one_gt_losses['loss_angle'].sum()
self.assertGreater(onegt_cls_loss.item(), 0,
'cls loss should be non-zero')
self.assertGreater(onegt_box_loss.item(), 0,
'box loss should be non-zero')
self.assertGreater(onegt_angle_loss.item(), 0,
'angle loss should be non-zero')
# test num_class = 1
self.head_module['num_classes'] = 1
head = RTMDetRotatedHead(
head_module=self.head_module, **hbb_cfg, train_cfg=train_cfg)
gt_instances = InstanceData(
bboxes=torch.Tensor([[130.6667, 86.8757, 100.6326, 70.8874, 0.2]]),
labels=torch.LongTensor([0]))
cls_scores, bbox_preds, angle_preds = head.forward(feat)
one_gt_losses = head.loss_by_feat(cls_scores, bbox_preds, angle_preds,
[gt_instances], img_metas)
onegt_cls_loss = one_gt_losses['loss_cls'].sum()
onegt_box_loss = one_gt_losses['loss_bbox'].sum()
onegt_angle_loss = one_gt_losses['loss_angle'].sum()
self.assertGreater(onegt_cls_loss.item(), 0,
'cls loss should be non-zero')
self.assertGreater(onegt_box_loss.item(), 0,
'box loss should be non-zero')
self.assertGreater(onegt_angle_loss.item(), 0,
'angle loss should be non-zero')
|