yolo_v3.py 27 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575
  1. # copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
  2. #
  3. # Licensed under the Apache License, Version 2.0 (the "License");
  4. # you may not use this file except in compliance with the License.
  5. # You may obtain a copy of the License at
  6. #
  7. # http://www.apache.org/licenses/LICENSE-2.0
  8. #
  9. # Unless required by applicable law or agreed to in writing, software
  10. # distributed under the License is distributed on an "AS IS" BASIS,
  11. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  12. # See the License for the specific language governing permissions and
  13. # limitations under the License.
  14. from __future__ import absolute_import
  15. import math
  16. import tqdm
  17. import os.path as osp
  18. import numpy as np
  19. from multiprocessing.pool import ThreadPool
  20. import paddle.fluid as fluid
  21. from paddle.fluid.layers.learning_rate_scheduler import _decay_step_counter
  22. from paddle.fluid.optimizer import ExponentialMovingAverage
  23. import paddlex.utils.logging as logging
  24. import paddlex
  25. import copy
  26. from paddlex.cv.transforms import arrange_transforms
  27. from paddlex.cv.datasets import generate_minibatch
  28. from .base import BaseAPI
  29. from collections import OrderedDict
  30. from .utils.detection_eval import eval_results, bbox2out
  31. class YOLOv3(BaseAPI):
  32. """构建YOLOv3,并实现其训练、评估、预测和模型导出。
  33. Args:
  34. num_classes (int): 类别数。默认为80。
  35. backbone (str): YOLOv3的backbone网络,取值范围为['DarkNet53',
  36. 'ResNet34', 'MobileNetV1', 'MobileNetV3_large']。默认为'MobileNetV1'。
  37. anchors (list|tuple): anchor框的宽度和高度,为None时表示使用默认值
  38. [[10, 13], [16, 30], [33, 23], [30, 61], [62, 45],
  39. [59, 119], [116, 90], [156, 198], [373, 326]]。
  40. anchor_masks (list|tuple): 在计算YOLOv3损失时,使用anchor的mask索引,为None时表示使用默认值
  41. [[6, 7, 8], [3, 4, 5], [0, 1, 2]]。
  42. ignore_threshold (float): 在计算YOLOv3损失时,IoU大于`ignore_threshold`的预测框的置信度被忽略。默认为0.7。
  43. nms_score_threshold (float): 检测框的置信度得分阈值,置信度得分低于阈值的框应该被忽略。默认为0.01。
  44. nms_topk (int): 进行NMS时,根据置信度保留的最大检测框数。默认为1000。
  45. nms_keep_topk (int): 进行NMS后,每个图像要保留的总检测框数。默认为100。
  46. nms_iou_threshold (float): 进行NMS时,用于剔除检测框IOU的阈值。默认为0.45。
  47. label_smooth (bool): 是否使用label smooth。默认值为False。
  48. train_random_shapes (list|tuple): 训练时从列表中随机选择图像大小。默认值为[320, 352, 384, 416, 448, 480, 512, 544, 576, 608]。
  49. """
  50. def __init__(
  51. self,
  52. num_classes=80,
  53. backbone='MobileNetV1',
  54. with_dcn_v2=False,
  55. # YOLO Head
  56. anchors=None,
  57. anchor_masks=None,
  58. use_coord_conv=False,
  59. use_iou_aware=False,
  60. use_spp=False,
  61. use_drop_block=False,
  62. scale_x_y=1.0,
  63. # YOLOv3 Loss
  64. ignore_threshold=0.7,
  65. label_smooth=False,
  66. use_iou_loss=False,
  67. # NMS
  68. use_matrix_nms=False,
  69. nms_score_threshold=0.01,
  70. nms_topk=1000,
  71. nms_keep_topk=100,
  72. nms_iou_threshold=0.45,
  73. train_random_shapes=[
  74. 320, 352, 384, 416, 448, 480, 512, 544, 576, 608
  75. ]):
  76. self.init_params = locals()
  77. super(YOLOv3, self).__init__('detector')
  78. backbones = [
  79. 'DarkNet53', 'ResNet34', 'MobileNetV1', 'MobileNetV3_large',
  80. 'ResNet50_vd'
  81. ]
  82. assert backbone in backbones, "backbone should be one of {}".format(
  83. backbones)
  84. self.backbone = backbone
  85. self.num_classes = num_classes
  86. self.anchors = anchors
  87. self.anchor_masks = anchor_masks
  88. if anchors is None:
  89. self.anchors = [[10, 13], [16, 30], [33, 23], [30, 61], [62, 45],
  90. [59, 119], [116, 90], [156, 198], [373, 326]]
  91. if anchor_masks is None:
  92. self.anchor_masks = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
  93. self.ignore_threshold = ignore_threshold
  94. self.nms_score_threshold = nms_score_threshold
  95. self.nms_topk = nms_topk
  96. self.nms_keep_topk = nms_keep_topk
  97. self.nms_iou_threshold = nms_iou_threshold
  98. self.label_smooth = label_smooth
  99. self.sync_bn = True
  100. self.train_random_shapes = train_random_shapes
  101. self.fixed_input_shape = None
  102. self.use_fine_grained_loss = False
  103. if use_coord_conv or use_iou_aware or use_spp or use_drop_block or use_iou_loss:
  104. self.use_fine_grained_loss = True
  105. self.use_coord_conv = use_coord_conv
  106. self.use_iou_aware = use_iou_aware
  107. self.use_spp = use_spp
  108. self.use_drop_block = use_drop_block
  109. self.use_iou_loss = use_iou_loss
  110. self.scale_x_y = scale_x_y
  111. self.max_height = 608
  112. self.max_width = 608
  113. self.use_matrix_nms = use_matrix_nms
  114. self.use_ema = False
  115. self.with_dcn_v2 = with_dcn_v2
  116. def _get_backbone(self, backbone_name):
  117. if backbone_name == 'DarkNet53':
  118. backbone = paddlex.cv.nets.DarkNet(norm_type='sync_bn')
  119. elif backbone_name == 'ResNet34':
  120. backbone = paddlex.cv.nets.ResNet(
  121. norm_type='sync_bn',
  122. layers=34,
  123. freeze_norm=False,
  124. norm_decay=0.,
  125. feature_maps=[3, 4, 5],
  126. freeze_at=0)
  127. elif backbone_name == 'MobileNetV1':
  128. backbone = paddlex.cv.nets.MobileNetV1(norm_type='sync_bn')
  129. elif backbone_name.startswith('MobileNetV3'):
  130. model_name = backbone_name.split('_')[1]
  131. backbone = paddlex.cv.nets.MobileNetV3(
  132. norm_type='sync_bn', model_name=model_name)
  133. elif backbone_name == 'ResNet50_vd':
  134. backbone = paddlex.cv.nets.ResNet(
  135. norm_type='sync_bn',
  136. layers=50,
  137. freeze_norm=False,
  138. norm_decay=0.,
  139. feature_maps=[3, 4, 5],
  140. freeze_at=0,
  141. variant='d',
  142. dcn_v2_stages=[5] if self.with_dcn_v2 else [])
  143. return backbone
  144. def build_net(self, mode='train'):
  145. model = paddlex.cv.nets.detection.YOLOv3(
  146. backbone=self._get_backbone(self.backbone),
  147. num_classes=self.num_classes,
  148. mode=mode,
  149. anchors=self.anchors,
  150. anchor_masks=self.anchor_masks,
  151. ignore_threshold=self.ignore_threshold,
  152. label_smooth=self.label_smooth,
  153. nms_score_threshold=self.nms_score_threshold,
  154. nms_topk=self.nms_topk,
  155. nms_keep_topk=self.nms_keep_topk,
  156. nms_iou_threshold=self.nms_iou_threshold,
  157. fixed_input_shape=self.fixed_input_shape,
  158. coord_conv=self.use_coord_conv,
  159. iou_aware=self.use_iou_aware,
  160. scale_x_y=self.scale_x_y,
  161. spp=self.use_spp,
  162. drop_block=self.use_drop_block,
  163. use_matrix_nms=self.use_matrix_nms,
  164. use_fine_grained_loss=self.use_fine_grained_loss,
  165. use_iou_loss=self.use_iou_loss,
  166. batch_size=self.batch_size_per_gpu
  167. if hasattr(self, 'batch_size_per_gpu') else 8)
  168. if mode == 'train' and self.use_iou_loss or self.use_iou_aware:
  169. model.max_height = self.max_height
  170. model.max_width = self.max_width
  171. inputs = model.generate_inputs()
  172. model_out = model.build_net(inputs)
  173. outputs = OrderedDict([('bbox', model_out)])
  174. if mode == 'train':
  175. self.optimizer.minimize(model_out)
  176. outputs = OrderedDict([('loss', model_out)])
  177. if self.use_ema:
  178. global_steps = _decay_step_counter()
  179. self.ema = ExponentialMovingAverage(
  180. self.ema_decay, thres_steps=global_steps)
  181. self.ema.update()
  182. return inputs, outputs
  183. def default_optimizer(self, learning_rate, warmup_steps, warmup_start_lr,
  184. lr_decay_epochs, lr_decay_gamma,
  185. num_steps_each_epoch):
  186. if warmup_steps > lr_decay_epochs[0] * num_steps_each_epoch:
  187. logging.error(
  188. "In function train(), parameters should satisfy: warmup_steps <= lr_decay_epochs[0]*num_samples_in_train_dataset",
  189. exit=False)
  190. logging.error(
  191. "See this doc for more information: https://github.com/PaddlePaddle/PaddleX/blob/develop/docs/appendix/parameters.md#notice",
  192. exit=False)
  193. logging.error(
  194. "warmup_steps should less than {} or lr_decay_epochs[0] greater than {}, please modify 'lr_decay_epochs' or 'warmup_steps' in train function".
  195. format(lr_decay_epochs[0] * num_steps_each_epoch, warmup_steps
  196. // num_steps_each_epoch))
  197. boundaries = [b * num_steps_each_epoch for b in lr_decay_epochs]
  198. values = [(lr_decay_gamma**i) * learning_rate
  199. for i in range(len(lr_decay_epochs) + 1)]
  200. lr_decay = fluid.layers.piecewise_decay(
  201. boundaries=boundaries, values=values)
  202. lr_warmup = fluid.layers.linear_lr_warmup(
  203. learning_rate=lr_decay,
  204. warmup_steps=warmup_steps,
  205. start_lr=warmup_start_lr,
  206. end_lr=learning_rate)
  207. optimizer = fluid.optimizer.Momentum(
  208. learning_rate=lr_warmup,
  209. momentum=0.9,
  210. regularization=fluid.regularizer.L2DecayRegularizer(5e-04))
  211. return optimizer
  212. def train(self,
  213. num_epochs,
  214. train_dataset,
  215. train_batch_size=8,
  216. eval_dataset=None,
  217. save_interval_epochs=20,
  218. log_interval_steps=2,
  219. save_dir='output',
  220. pretrain_weights='IMAGENET',
  221. optimizer=None,
  222. learning_rate=1.0 / 8000,
  223. warmup_steps=1000,
  224. warmup_start_lr=0.0,
  225. lr_decay_epochs=[213, 240],
  226. lr_decay_gamma=0.1,
  227. use_ema=False,
  228. ema_decay=0.9998,
  229. metric=None,
  230. use_vdl=False,
  231. sensitivities_file=None,
  232. eval_metric_loss=0.05,
  233. early_stop=False,
  234. early_stop_patience=5,
  235. resume_checkpoint=None):
  236. """训练。
  237. Args:
  238. num_epochs (int): 训练迭代轮数。
  239. train_dataset (paddlex.datasets): 训练数据读取器。
  240. train_batch_size (int): 训练数据batch大小。目前检测仅支持单卡评估,训练数据batch大小与显卡
  241. 数量之商为验证数据batch大小。默认值为8。
  242. eval_dataset (paddlex.datasets): 验证数据读取器。
  243. save_interval_epochs (int): 模型保存间隔(单位:迭代轮数)。默认为20。
  244. log_interval_steps (int): 训练日志输出间隔(单位:迭代次数)。默认为10。
  245. save_dir (str): 模型保存路径。默认值为'output'。
  246. pretrain_weights (str): 若指定为路径时,则加载路径下预训练模型;若为字符串'IMAGENET',
  247. 则自动下载在ImageNet图片数据上预训练的模型权重;若为字符串'COCO',
  248. 则自动下载在COCO数据集上预训练的模型权重;若为None,则不使用预训练模型。默认为'IMAGENET'。
  249. optimizer (paddle.fluid.optimizer): 优化器。当该参数为None时,使用默认优化器:
  250. fluid.layers.piecewise_decay衰减策略,fluid.optimizer.Momentum优化方法。
  251. learning_rate (float): 默认优化器的学习率。默认为1.0/8000。
  252. warmup_steps (int): 默认优化器进行warmup过程的步数。默认为1000。
  253. warmup_start_lr (int): 默认优化器warmup的起始学习率。默认为0.0。
  254. lr_decay_epochs (list): 默认优化器的学习率衰减轮数。默认为[213, 240]。
  255. lr_decay_gamma (float): 默认优化器的学习率衰减率。默认为0.1。
  256. metric (bool): 训练过程中评估的方式,取值范围为['COCO', 'VOC']。默认值为None。
  257. use_vdl (bool): 是否使用VisualDL进行可视化。默认值为False。
  258. sensitivities_file (str): 若指定为路径时,则加载路径下敏感度信息进行裁剪;若为字符串'DEFAULT',
  259. 则自动下载在ImageNet图片数据上获得的敏感度信息进行裁剪;若为None,则不进行裁剪。默认为None。
  260. eval_metric_loss (float): 可容忍的精度损失。默认为0.05。
  261. early_stop (bool): 是否使用提前终止训练策略。默认值为False。
  262. early_stop_patience (int): 当使用提前终止训练策略时,如果验证集精度在`early_stop_patience`个epoch内
  263. 连续下降或持平,则终止训练。默认值为5。
  264. resume_checkpoint (str): 恢复训练时指定上次训练保存的模型路径。若为None,则不会恢复训练。默认值为None。
  265. Raises:
  266. ValueError: 评估类型不在指定列表中。
  267. ValueError: 模型从inference model进行加载。
  268. """
  269. if not self.trainable:
  270. raise ValueError("Model is not trainable from load_model method.")
  271. if metric is None:
  272. if isinstance(train_dataset, paddlex.datasets.CocoDetection):
  273. metric = 'COCO'
  274. elif isinstance(train_dataset, paddlex.datasets.VOCDetection) or \
  275. isinstance(train_dataset, paddlex.datasets.EasyDataDet):
  276. metric = 'VOC'
  277. else:
  278. raise ValueError(
  279. "train_dataset should be datasets.VOCDetection or datasets.COCODetection or datasets.EasyDataDet."
  280. )
  281. assert metric in ['COCO', 'VOC'], "Metric only support 'VOC' or 'COCO'"
  282. self.metric = metric
  283. self.labels = train_dataset.labels
  284. # 构建训练网络
  285. if optimizer is None:
  286. # 构建默认的优化策略
  287. num_steps_each_epoch = train_dataset.num_samples // train_batch_size
  288. optimizer = self.default_optimizer(
  289. learning_rate=learning_rate,
  290. warmup_steps=warmup_steps,
  291. warmup_start_lr=warmup_start_lr,
  292. lr_decay_epochs=lr_decay_epochs,
  293. lr_decay_gamma=lr_decay_gamma,
  294. num_steps_each_epoch=num_steps_each_epoch)
  295. self.optimizer = optimizer
  296. self.use_ema = use_ema
  297. self.ema_decay = ema_decay
  298. self.batch_size_per_gpu = int(train_batch_size /
  299. paddlex.env_info['num'])
  300. if self.use_fine_grained_loss:
  301. for transform in train_dataset.transforms.transforms:
  302. if isinstance(transform, paddlex.det.transforms.Resize):
  303. self.max_height = transform.target_size
  304. self.max_width = transform.target_size
  305. break
  306. if train_dataset.transforms.batch_transforms is None:
  307. train_dataset.transforms.batch_transforms = list()
  308. define_random_shape = False
  309. for bt in train_dataset.transforms.batch_transforms:
  310. if isinstance(bt, paddlex.det.transforms.BatchRandomShape):
  311. define_random_shape = True
  312. if not define_random_shape:
  313. if isinstance(self.train_random_shapes,
  314. (list, tuple)) and len(self.train_random_shapes) > 0:
  315. train_dataset.transforms.batch_transforms.append(
  316. paddlex.det.transforms.BatchRandomShape(
  317. random_shapes=self.train_random_shapes))
  318. if self.use_fine_grained_loss:
  319. self.max_height = max(self.max_height,
  320. max(self.train_random_shapes))
  321. self.max_width = max(self.max_width,
  322. max(self.train_random_shapes))
  323. if self.use_fine_grained_loss:
  324. define_generate_target = False
  325. for bt in train_dataset.transforms.batch_transforms:
  326. if isinstance(bt, paddlex.det.transforms.GenerateYoloTarget):
  327. define_generate_target = True
  328. if not define_generate_target:
  329. train_dataset.transforms.batch_transforms.append(
  330. paddlex.det.transforms.GenerateYoloTarget(
  331. anchors=self.anchors,
  332. anchor_masks=self.anchor_masks,
  333. num_classes=self.num_classes,
  334. downsample_ratios=[32, 16, 8]))
  335. # 构建训练、验证、预测网络
  336. self.build_program()
  337. # 初始化网络权重
  338. self.net_initialize(
  339. startup_prog=fluid.default_startup_program(),
  340. pretrain_weights=pretrain_weights,
  341. save_dir=save_dir,
  342. sensitivities_file=sensitivities_file,
  343. eval_metric_loss=eval_metric_loss,
  344. resume_checkpoint=resume_checkpoint)
  345. # 训练
  346. self.train_loop(
  347. num_epochs=num_epochs,
  348. train_dataset=train_dataset,
  349. train_batch_size=train_batch_size,
  350. eval_dataset=eval_dataset,
  351. save_interval_epochs=save_interval_epochs,
  352. log_interval_steps=log_interval_steps,
  353. save_dir=save_dir,
  354. use_vdl=use_vdl,
  355. early_stop=early_stop,
  356. early_stop_patience=early_stop_patience)
  357. def evaluate(self,
  358. eval_dataset,
  359. batch_size=1,
  360. epoch_id=None,
  361. metric=None,
  362. return_details=False):
  363. """评估。
  364. Args:
  365. eval_dataset (paddlex.datasets): 验证数据读取器。
  366. batch_size (int): 验证数据批大小。默认为1。
  367. epoch_id (int): 当前评估模型所在的训练轮数。
  368. metric (bool): 训练过程中评估的方式,取值范围为['COCO', 'VOC']。默认为None,
  369. 根据用户传入的Dataset自动选择,如为VOCDetection,则metric为'VOC';
  370. 如为COCODetection,则metric为'COCO'。
  371. return_details (bool): 是否返回详细信息。
  372. Returns:
  373. tuple (metrics, eval_details) | dict (metrics): 当return_details为True时,返回(metrics, eval_details),
  374. 当return_details为False时,返回metrics。metrics为dict,包含关键字:'bbox_mmap'或者’bbox_map‘,
  375. 分别表示平均准确率平均值在各个IoU阈值下的结果取平均值的结果(mmAP)、平均准确率平均值(mAP)。
  376. eval_details为dict,包含关键字:'bbox',对应元素预测结果列表,每个预测结果由图像id、
  377. 预测框类别id、预测框坐标、预测框得分;’gt‘:真实标注框相关信息。
  378. """
  379. arrange_transforms(
  380. model_type=self.model_type,
  381. class_name=self.__class__.__name__,
  382. transforms=eval_dataset.transforms,
  383. mode='eval')
  384. if metric is None:
  385. if hasattr(self, 'metric') and self.metric is not None:
  386. metric = self.metric
  387. else:
  388. if isinstance(eval_dataset, paddlex.datasets.CocoDetection):
  389. metric = 'COCO'
  390. elif isinstance(eval_dataset, paddlex.datasets.VOCDetection):
  391. metric = 'VOC'
  392. else:
  393. raise Exception(
  394. "eval_dataset should be datasets.VOCDetection or datasets.COCODetection."
  395. )
  396. assert metric in ['COCO', 'VOC'], "Metric only support 'VOC' or 'COCO'"
  397. total_steps = math.ceil(eval_dataset.num_samples * 1.0 / batch_size)
  398. results = list()
  399. data_generator = eval_dataset.generator(
  400. batch_size=batch_size, drop_last=False)
  401. logging.info(
  402. "Start to evaluating(total_samples={}, total_steps={})...".format(
  403. eval_dataset.num_samples, total_steps))
  404. for step, data in tqdm.tqdm(
  405. enumerate(data_generator()), total=total_steps):
  406. images = np.array([d[0] for d in data])
  407. im_sizes = np.array([d[1] for d in data])
  408. feed_data = {'image': images, 'im_size': im_sizes}
  409. with fluid.scope_guard(self.scope):
  410. outputs = self.exe.run(
  411. self.test_prog,
  412. feed=[feed_data],
  413. fetch_list=list(self.test_outputs.values()),
  414. return_numpy=False)
  415. res = {
  416. 'bbox': (np.array(outputs[0]),
  417. outputs[0].recursive_sequence_lengths())
  418. }
  419. res_id = [np.array([d[2]]) for d in data]
  420. res['im_id'] = (res_id, [])
  421. if metric == 'VOC':
  422. res_gt_box = [d[3].reshape(-1, 4) for d in data]
  423. res_gt_label = [d[4].reshape(-1, 1) for d in data]
  424. res_is_difficult = [d[5].reshape(-1, 1) for d in data]
  425. res_id = [np.array([d[2]]) for d in data]
  426. res['gt_box'] = (res_gt_box, [])
  427. res['gt_label'] = (res_gt_label, [])
  428. res['is_difficult'] = (res_is_difficult, [])
  429. results.append(res)
  430. logging.debug("[EVAL] Epoch={}, Step={}/{}".format(epoch_id, step +
  431. 1, total_steps))
  432. box_ap_stats, eval_details = eval_results(
  433. results, metric, eval_dataset.coco_gt, with_background=False)
  434. evaluate_metrics = OrderedDict(
  435. zip(['bbox_mmap'
  436. if metric == 'COCO' else 'bbox_map'], box_ap_stats))
  437. if return_details:
  438. return evaluate_metrics, eval_details
  439. return evaluate_metrics
  440. @staticmethod
  441. def _preprocess(images, transforms, model_type, class_name, thread_num=1):
  442. arrange_transforms(
  443. model_type=model_type,
  444. class_name=class_name,
  445. transforms=transforms,
  446. mode='test')
  447. pool = ThreadPool(thread_num)
  448. batch_data = pool.map(transforms, images)
  449. pool.close()
  450. pool.join()
  451. padding_batch = generate_minibatch(batch_data)
  452. im = np.array(
  453. [data[0] for data in padding_batch],
  454. dtype=padding_batch[0][0].dtype)
  455. im_size = np.array([data[1] for data in padding_batch], dtype=np.int32)
  456. return im, im_size
  457. @staticmethod
  458. def _postprocess(res, batch_size, num_classes, labels):
  459. clsid2catid = dict({i: i for i in range(num_classes)})
  460. xywh_results = bbox2out([res], clsid2catid)
  461. preds = [[] for i in range(batch_size)]
  462. for xywh_res in xywh_results:
  463. image_id = xywh_res['image_id']
  464. del xywh_res['image_id']
  465. xywh_res['category'] = labels[xywh_res['category_id']]
  466. preds[image_id].append(xywh_res)
  467. return preds
  468. def predict(self, img_file, transforms=None):
  469. """预测。
  470. Args:
  471. img_file (str|np.ndarray): 预测图像路径,或者是解码后的排列格式为(H, W, C)且类型为float32且为BGR格式的数组。
  472. transforms (paddlex.det.transforms): 数据预处理操作。
  473. Returns:
  474. list: 预测结果列表,每个预测结果由预测框类别标签、
  475. 预测框类别名称、预测框坐标(坐标格式为[xmin, ymin, w, h])、
  476. 预测框得分组成。
  477. """
  478. if transforms is None and not hasattr(self, 'test_transforms'):
  479. raise Exception("transforms need to be defined, now is None.")
  480. if isinstance(img_file, (str, np.ndarray)):
  481. images = [img_file]
  482. else:
  483. raise Exception("img_file must be str/np.ndarray")
  484. if transforms is None:
  485. transforms = self.test_transforms
  486. im, im_size = YOLOv3._preprocess(images, transforms, self.model_type,
  487. self.__class__.__name__)
  488. with fluid.scope_guard(self.scope):
  489. result = self.exe.run(self.test_prog,
  490. feed={'image': im,
  491. 'im_size': im_size},
  492. fetch_list=list(self.test_outputs.values()),
  493. return_numpy=False,
  494. use_program_cache=True)
  495. res = {
  496. k: (np.array(v), v.recursive_sequence_lengths())
  497. for k, v in zip(list(self.test_outputs.keys()), result)
  498. }
  499. res['im_id'] = (np.array(
  500. [[i] for i in range(len(images))]).astype('int32'), [[]])
  501. preds = YOLOv3._postprocess(res,
  502. len(images), self.num_classes, self.labels)
  503. return preds[0]
  504. def batch_predict(self, img_file_list, transforms=None, thread_num=2):
  505. """预测。
  506. Args:
  507. img_file_list (list|tuple): 对列表(或元组)中的图像同时进行预测,列表中的元素可以是图像路径,也可以是解码后的排列格式为(H,W,C)
  508. 且类型为float32且为BGR格式的数组。
  509. transforms (paddlex.det.transforms): 数据预处理操作。
  510. thread_num (int): 并发执行各图像预处理时的线程数。
  511. Returns:
  512. list: 每个元素都为列表,表示各图像的预测结果。在各图像的预测结果列表中,每个预测结果由预测框类别标签、
  513. 预测框类别名称、预测框坐标(坐标格式为[xmin, ymin, w, h])、
  514. 预测框得分组成。
  515. """
  516. if transforms is None and not hasattr(self, 'test_transforms'):
  517. raise Exception("transforms need to be defined, now is None.")
  518. if not isinstance(img_file_list, (list, tuple)):
  519. raise Exception("im_file must be list/tuple")
  520. if transforms is None:
  521. transforms = self.test_transforms
  522. im, im_size = YOLOv3._preprocess(img_file_list, transforms,
  523. self.model_type,
  524. self.__class__.__name__, thread_num)
  525. with fluid.scope_guard(self.scope):
  526. result = self.exe.run(self.test_prog,
  527. feed={'image': im,
  528. 'im_size': im_size},
  529. fetch_list=list(self.test_outputs.values()),
  530. return_numpy=False,
  531. use_program_cache=True)
  532. res = {
  533. k: (np.array(v), v.recursive_sequence_lengths())
  534. for k, v in zip(list(self.test_outputs.keys()), result)
  535. }
  536. res['im_id'] = (np.array(
  537. [[i] for i in range(len(img_file_list))]).astype('int32'), [[]])
  538. preds = YOLOv3._postprocess(res,
  539. len(img_file_list), self.num_classes,
  540. self.labels)
  541. return preds