yolo_v3.py

# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from __future__ import absolute_import
import math
import tqdm
import os.path as osp
import numpy as np
from multiprocessing.pool import ThreadPool
import paddle.fluid as fluid
from paddle.fluid.layers.learning_rate_scheduler import _decay_step_counter
from paddle.fluid.optimizer import ExponentialMovingAverage
import paddlex.utils.logging as logging
import paddlex
import copy
from paddlex.cv.transforms import arrange_transforms
from paddlex.cv.datasets import generate_minibatch
from .base import BaseAPI
from collections import OrderedDict
from .utils.detection_eval import eval_results, bbox2out


class YOLOv3(BaseAPI):
    """构建YOLOv3，并实现其训练、评估、预测和模型导出。

    Args:
        num_classes (int): 类别数。默认为80。
        backbone (str): YOLOv3的backbone网络，取值范围为['DarkNet53',
            'ResNet34', 'MobileNetV1', 'MobileNetV3_large']。默认为'MobileNetV1'。
        anchors (list|tuple): anchor框的宽度和高度，为None时表示使用默认值
                    [[10, 13], [16, 30], [33, 23], [30, 61], [62, 45],
                    [59, 119], [116, 90], [156, 198], [373, 326]]。
        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
                    [[6, 7, 8], [3, 4, 5], [0, 1, 2]]。
        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。
        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。
        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。
        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。
        nms_iou_threshold (float): 进行NMS时，用于剔除检测框IOU的阈值。默认为0.45。
        label_smooth (bool): 是否使用label smooth。默认值为False。
        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]。
    """

    def __init__(
            self,
            num_classes=80,
            backbone='MobileNetV1',
            with_dcn_v2=False,
            # YOLO Head
            anchors=None,
            anchor_masks=None,
            use_coord_conv=False,
            use_iou_aware=False,
            use_spp=False,
            use_drop_block=False,
            scale_x_y=1.0,
            # YOLOv3 Loss
            ignore_threshold=0.7,
            label_smooth=False,
            use_iou_loss=False,
            # NMS
            use_matrix_nms=False,
            nms_score_threshold=0.01,
            nms_topk=1000,
            nms_keep_topk=100,
            nms_iou_threshold=0.45,
            train_random_shapes=[
                320, 352, 384, 416, 448, 480, 512, 544, 576, 608
            ]):
        self.init_params = locals()
        super(YOLOv3, self).__init__('detector')
        backbones = [
            'DarkNet53', 'ResNet34', 'MobileNetV1', 'MobileNetV3_large',
            'ResNet50_vd'
        ]
        assert backbone in backbones, "backbone should be one of {}".format(
            backbones)
        self.backbone = backbone
        self.num_classes = num_classes
        self.anchors = anchors
        self.anchor_masks = anchor_masks
        if anchors is None:
            self.anchors = [[10, 13], [16, 30], [33, 23], [30, 61], [62, 45],
                            [59, 119], [116, 90], [156, 198], [373, 326]]
        if anchor_masks is None:
            self.anchor_masks = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
        self.ignore_threshold = ignore_threshold
        self.nms_score_threshold = nms_score_threshold
        self.nms_topk = nms_topk
        self.nms_keep_topk = nms_keep_topk
        self.nms_iou_threshold = nms_iou_threshold
        self.label_smooth = label_smooth
        self.sync_bn = True
        self.train_random_shapes = train_random_shapes
        self.fixed_input_shape = None
        self.use_fine_grained_loss = False
        if use_coord_conv or use_iou_aware or use_spp or use_drop_block or use_iou_loss:
            self.use_fine_grained_loss = True
        self.use_coord_conv = use_coord_conv
        self.use_iou_aware = use_iou_aware
        self.use_spp = use_spp
        self.use_drop_block = use_drop_block
        self.use_iou_loss = use_iou_loss
        self.scale_x_y = scale_x_y
        self.max_height = 608
        self.max_width = 608
        self.use_matrix_nms = use_matrix_nms
        self.use_ema = False
        self.with_dcn_v2 = with_dcn_v2

    def _get_backbone(self, backbone_name):
        if backbone_name == 'DarkNet53':
            backbone = paddlex.cv.nets.DarkNet(norm_type='sync_bn')
        elif backbone_name == 'ResNet34':
            backbone = paddlex.cv.nets.ResNet(
                norm_type='sync_bn',
                layers=34,
                freeze_norm=False,
                norm_decay=0.,
                feature_maps=[3, 4, 5],
                freeze_at=0)
        elif backbone_name == 'MobileNetV1':
            backbone = paddlex.cv.nets.MobileNetV1(norm_type='sync_bn')
        elif backbone_name.startswith('MobileNetV3'):
            model_name = backbone_name.split('_')[1]
            backbone = paddlex.cv.nets.MobileNetV3(
                norm_type='sync_bn', model_name=model_name)
        elif backbone_name == 'ResNet50_vd':
            backbone = paddlex.cv.nets.ResNet(
                norm_type='sync_bn',
                layers=50,
                freeze_norm=False,
                norm_decay=0.,
                feature_maps=[3, 4, 5],
                freeze_at=0,
                variant='d',
                dcn_v2_stages=[5] if self.with_dcn_v2 else [])
        return backbone

    def build_net(self, mode='train'):
        model = paddlex.cv.nets.detection.YOLOv3(
            backbone=self._get_backbone(self.backbone),
            num_classes=self.num_classes,
            mode=mode,
            anchors=self.anchors,
            anchor_masks=self.anchor_masks,
            ignore_threshold=self.ignore_threshold,
            label_smooth=self.label_smooth,
            nms_score_threshold=self.nms_score_threshold,
            nms_topk=self.nms_topk,
            nms_keep_topk=self.nms_keep_topk,
            nms_iou_threshold=self.nms_iou_threshold,
            fixed_input_shape=self.fixed_input_shape,
            coord_conv=self.use_coord_conv,
            iou_aware=self.use_iou_aware,
            scale_x_y=self.scale_x_y,
            spp=self.use_spp,
            drop_block=self.use_drop_block,
            use_matrix_nms=self.use_matrix_nms,
            use_fine_grained_loss=self.use_fine_grained_loss,
            use_iou_loss=self.use_iou_loss,
            batch_size=self.batch_size_per_gpu
            if hasattr(self, 'batch_size_per_gpu') else 8)
        if mode == 'train' and self.use_iou_loss or self.use_iou_aware:
            model.max_height = self.max_height
            model.max_width = self.max_width
        inputs = model.generate_inputs()
        model_out = model.build_net(inputs)
        outputs = OrderedDict([('bbox', model_out)])
        if mode == 'train':
            self.optimizer.minimize(model_out)
            outputs = OrderedDict([('loss', model_out)])
            if self.use_ema:
                global_steps = _decay_step_counter()
                self.ema = ExponentialMovingAverage(
                    self.ema_decay, thres_steps=global_steps)
                self.ema.update()
        return inputs, outputs

    def default_optimizer(self, learning_rate, warmup_steps, warmup_start_lr,
                          lr_decay_epochs, lr_decay_gamma,
                          num_steps_each_epoch):
        if warmup_steps > lr_decay_epochs[0] * num_steps_each_epoch:
            logging.error(
                "In function train(), parameters should satisfy: warmup_steps <= lr_decay_epochs[0]*num_samples_in_train_dataset",
                exit=False)
            logging.error(
                "See this doc for more information: https://github.com/PaddlePaddle/PaddleX/blob/develop/docs/appendix/parameters.md#notice",
                exit=False)
            logging.error(
                "warmup_steps should less than {} or lr_decay_epochs[0] greater than {}, please modify 'lr_decay_epochs' or 'warmup_steps' in train function".
                format(lr_decay_epochs[0] * num_steps_each_epoch, warmup_steps
                       // num_steps_each_epoch))
        boundaries = [b * num_steps_each_epoch for b in lr_decay_epochs]
        values = [(lr_decay_gamma**i) * learning_rate
                  for i in range(len(lr_decay_epochs) + 1)]
        lr_decay = fluid.layers.piecewise_decay(
            boundaries=boundaries, values=values)
        lr_warmup = fluid.layers.linear_lr_warmup(
            learning_rate=lr_decay,
            warmup_steps=warmup_steps,
            start_lr=warmup_start_lr,
            end_lr=learning_rate)
        optimizer = fluid.optimizer.Momentum(
            learning_rate=lr_warmup,
            momentum=0.9,
            regularization=fluid.regularizer.L2DecayRegularizer(5e-04))
        return optimizer

    def train(self,
              num_epochs,
              train_dataset,
              train_batch_size=8,
              eval_dataset=None,
              save_interval_epochs=20,
              log_interval_steps=2,
              save_dir='output',
              pretrain_weights='IMAGENET',
              optimizer=None,
              learning_rate=1.0 / 8000,
              warmup_steps=1000,
              warmup_start_lr=0.0,
              lr_decay_epochs=[213, 240],
              lr_decay_gamma=0.1,
              use_ema=False,
              ema_decay=0.9998,
              metric=None,
              use_vdl=False,
              sensitivities_file=None,
              eval_metric_loss=0.05,
              early_stop=False,
              early_stop_patience=5,
              resume_checkpoint=None):
        """训练。

        Args:
            num_epochs (int): 训练迭代轮数。
            train_dataset (paddlex.datasets): 训练数据读取器。
            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
                数量之商为验证数据batch大小。默认值为8。
            eval_dataset (paddlex.datasets): 验证数据读取器。
            save_interval_epochs (int): 模型保存间隔（单位：迭代轮数）。默认为20。
            log_interval_steps (int): 训练日志输出间隔（单位：迭代次数）。默认为10。
            save_dir (str): 模型保存路径。默认值为'output'。
            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'，
                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'，
                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'。
            optimizer (paddle.fluid.optimizer): 优化器。当该参数为None时，使用默认优化器：
                fluid.layers.piecewise_decay衰减策略，fluid.optimizer.Momentum优化方法。
            learning_rate (float): 默认优化器的学习率。默认为1.0/8000。
            warmup_steps (int):  默认优化器进行warmup过程的步数。默认为1000。
            warmup_start_lr (int): 默认优化器warmup的起始学习率。默认为0.0。
            lr_decay_epochs (list): 默认优化器的学习率衰减轮数。默认为[213, 240]。
            lr_decay_gamma (float): 默认优化器的学习率衰减率。默认为0.1。
            metric (bool): 训练过程中评估的方式，取值范围为['COCO', 'VOC']。默认值为None。
            use_vdl (bool): 是否使用VisualDL进行可视化。默认值为False。
            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'，
                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。
            eval_metric_loss (float): 可容忍的精度损失。默认为0.05。
            early_stop (bool): 是否使用提前终止训练策略。默认值为False。
            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
                连续下降或持平，则终止训练。默认值为5。
            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。

        Raises:
            ValueError: 评估类型不在指定列表中。
            ValueError: 模型从inference model进行加载。
        """
        if not self.trainable:
            raise ValueError("Model is not trainable from load_model method.")
        if metric is None:
            if isinstance(train_dataset, paddlex.datasets.CocoDetection):
                metric = 'COCO'
            elif isinstance(train_dataset, paddlex.datasets.VOCDetection) or \
                    isinstance(train_dataset, paddlex.datasets.EasyDataDet):
                metric = 'VOC'
            else:
                raise ValueError(
                    "train_dataset should be datasets.VOCDetection or datasets.COCODetection or datasets.EasyDataDet."
                )
        assert metric in ['COCO', 'VOC'], "Metric only support 'VOC' or 'COCO'"
        self.metric = metric

        self.labels = train_dataset.labels
        # 构建训练网络
        if optimizer is None:
            # 构建默认的优化策略
            num_steps_each_epoch = train_dataset.num_samples // train_batch_size
            optimizer = self.default_optimizer(
                learning_rate=learning_rate,
                warmup_steps=warmup_steps,
                warmup_start_lr=warmup_start_lr,
                lr_decay_epochs=lr_decay_epochs,
                lr_decay_gamma=lr_decay_gamma,
                num_steps_each_epoch=num_steps_each_epoch)
        self.optimizer = optimizer
        self.use_ema = use_ema
        self.ema_decay = ema_decay

        self.batch_size_per_gpu = int(train_batch_size /
                                      paddlex.env_info['num'])
        if self.use_fine_grained_loss:
            for transform in train_dataset.transforms.transforms:
                if isinstance(transform, paddlex.det.transforms.Resize):
                    self.max_height = transform.target_size
                    self.max_width = transform.target_size
                    break
        if train_dataset.transforms.batch_transforms is None:
            train_dataset.transforms.batch_transforms = list()
        define_random_shape = False
        for bt in train_dataset.transforms.batch_transforms:
            if isinstance(bt, paddlex.det.transforms.BatchRandomShape):
                define_random_shape = True
        if not define_random_shape:
            if isinstance(self.train_random_shapes,
                          (list, tuple)) and len(self.train_random_shapes) > 0:
                train_dataset.transforms.batch_transforms.append(
                    paddlex.det.transforms.BatchRandomShape(
                        random_shapes=self.train_random_shapes))
                if self.use_fine_grained_loss:
                    self.max_height = max(self.max_height,
                                          max(self.train_random_shapes))
                    self.max_width = max(self.max_width,
                                         max(self.train_random_shapes))
        if self.use_fine_grained_loss:
            define_generate_target = False
            for bt in train_dataset.transforms.batch_transforms:
                if isinstance(bt, paddlex.det.transforms.GenerateYoloTarget):
                    define_generate_target = True
            if not define_generate_target:
                train_dataset.transforms.batch_transforms.append(
                    paddlex.det.transforms.GenerateYoloTarget(
                        anchors=self.anchors,
                        anchor_masks=self.anchor_masks,
                        num_classes=self.num_classes,
                        downsample_ratios=[32, 16, 8]))
        # 构建训练、验证、预测网络
        self.build_program()
        # 初始化网络权重
        self.net_initialize(
            startup_prog=fluid.default_startup_program(),
            pretrain_weights=pretrain_weights,
            save_dir=save_dir,
            sensitivities_file=sensitivities_file,
            eval_metric_loss=eval_metric_loss,
            resume_checkpoint=resume_checkpoint)
        # 训练
        self.train_loop(
            num_epochs=num_epochs,
            train_dataset=train_dataset,
            train_batch_size=train_batch_size,
            eval_dataset=eval_dataset,
            save_interval_epochs=save_interval_epochs,
            log_interval_steps=log_interval_steps,
            save_dir=save_dir,
            use_vdl=use_vdl,
            early_stop=early_stop,
            early_stop_patience=early_stop_patience)

    def evaluate(self,
                 eval_dataset,
                 batch_size=1,
                 epoch_id=None,
                 metric=None,
                 return_details=False):
        """评估。

        Args:
            eval_dataset (paddlex.datasets): 验证数据读取器。
            batch_size (int): 验证数据批大小。默认为1。
            epoch_id (int): 当前评估模型所在的训练轮数。
            metric (bool): 训练过程中评估的方式，取值范围为['COCO', 'VOC']。默认为None，
                根据用户传入的Dataset自动选择，如为VOCDetection，则metric为'VOC';
                如为COCODetection，则metric为'COCO'。
            return_details (bool): 是否返回详细信息。

        Returns:
            tuple (metrics, eval_details) | dict (metrics): 当return_details为True时，返回(metrics, eval_details)，
                当return_details为False时，返回metrics。metrics为dict，包含关键字：'bbox_mmap'或者’bbox_map‘，
                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）。
                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、
                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。
        """
        arrange_transforms(
            model_type=self.model_type,
            class_name=self.__class__.__name__,
            transforms=eval_dataset.transforms,
            mode='eval')
        if metric is None:
            if hasattr(self, 'metric') and self.metric is not None:
                metric = self.metric
            else:
                if isinstance(eval_dataset, paddlex.datasets.CocoDetection):
                    metric = 'COCO'
                elif isinstance(eval_dataset, paddlex.datasets.VOCDetection):
                    metric = 'VOC'
                else:
                    raise Exception(
                        "eval_dataset should be datasets.VOCDetection or datasets.COCODetection."
                    )
        assert metric in ['COCO', 'VOC'], "Metric only support 'VOC' or 'COCO'"

        total_steps = math.ceil(eval_dataset.num_samples * 1.0 / batch_size)
        results = list()

        data_generator = eval_dataset.generator(
            batch_size=batch_size, drop_last=False)
        logging.info(
            "Start to evaluating(total_samples={}, total_steps={})...".format(
                eval_dataset.num_samples, total_steps))
        for step, data in tqdm.tqdm(
                enumerate(data_generator()), total=total_steps):
            images = np.array([d[0] for d in data])
            im_sizes = np.array([d[1] for d in data])
            feed_data = {'image': images, 'im_size': im_sizes}
            with fluid.scope_guard(self.scope):
                outputs = self.exe.run(
                    self.test_prog,
                    feed=[feed_data],
                    fetch_list=list(self.test_outputs.values()),
                    return_numpy=False)
            res = {
                'bbox': (np.array(outputs[0]),
                         outputs[0].recursive_sequence_lengths())
            }
            res_id = [np.array([d[2]]) for d in data]
            res['im_id'] = (res_id, [])
            if metric == 'VOC':
                res_gt_box = [d[3].reshape(-1, 4) for d in data]
                res_gt_label = [d[4].reshape(-1, 1) for d in data]
                res_is_difficult = [d[5].reshape(-1, 1) for d in data]
                res_id = [np.array([d[2]]) for d in data]
                res['gt_box'] = (res_gt_box, [])
                res['gt_label'] = (res_gt_label, [])
                res['is_difficult'] = (res_is_difficult, [])
            results.append(res)
            logging.debug("[EVAL] Epoch={}, Step={}/{}".format(epoch_id, step +
                                                               1, total_steps))
        box_ap_stats, eval_details = eval_results(
            results, metric, eval_dataset.coco_gt, with_background=False)
        evaluate_metrics = OrderedDict(
            zip(['bbox_mmap'
                 if metric == 'COCO' else 'bbox_map'], box_ap_stats))
        if return_details:
            return evaluate_metrics, eval_details
        return evaluate_metrics

    @staticmethod
    def _preprocess(images, transforms, model_type, class_name, thread_num=1):
        arrange_transforms(
            model_type=model_type,
            class_name=class_name,
            transforms=transforms,
            mode='test')
        pool = ThreadPool(thread_num)
        batch_data = pool.map(transforms, images)
        pool.close()
        pool.join()
        padding_batch = generate_minibatch(batch_data)
        im = np.array(
            [data[0] for data in padding_batch],
            dtype=padding_batch[0][0].dtype)
        im_size = np.array([data[1] for data in padding_batch], dtype=np.int32)

        return im, im_size

    @staticmethod
    def _postprocess(res, batch_size, num_classes, labels):
        clsid2catid = dict({i: i for i in range(num_classes)})
        xywh_results = bbox2out([res], clsid2catid)
        preds = [[] for i in range(batch_size)]
        for xywh_res in xywh_results:
            image_id = xywh_res['image_id']
            del xywh_res['image_id']
            xywh_res['category'] = labels[xywh_res['category_id']]
            preds[image_id].append(xywh_res)

        return preds

    def predict(self, img_file, transforms=None):
        """预测。

        Args:
            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。
            transforms (paddlex.det.transforms): 数据预处理操作。

        Returns:
            list: 预测结果列表，每个预测结果由预测框类别标签、
              预测框类别名称、预测框坐标(坐标格式为[xmin, ymin, w, h]）、
              预测框得分组成。
        """
        if transforms is None and not hasattr(self, 'test_transforms'):
            raise Exception("transforms need to be defined, now is None.")
        if isinstance(img_file, (str, np.ndarray)):
            images = [img_file]
        else:
            raise Exception("img_file must be str/np.ndarray")

        if transforms is None:
            transforms = self.test_transforms
        im, im_size = YOLOv3._preprocess(images, transforms, self.model_type,
                                         self.__class__.__name__)

        with fluid.scope_guard(self.scope):
            result = self.exe.run(self.test_prog,
                                  feed={'image': im,
                                        'im_size': im_size},
                                  fetch_list=list(self.test_outputs.values()),
                                  return_numpy=False,
                                  use_program_cache=True)

        res = {
            k: (np.array(v), v.recursive_sequence_lengths())
            for k, v in zip(list(self.test_outputs.keys()), result)
        }
        res['im_id'] = (np.array(
            [[i] for i in range(len(images))]).astype('int32'), [[]])
        preds = YOLOv3._postprocess(res,
                                    len(images), self.num_classes, self.labels)
        return preds[0]

    def batch_predict(self, img_file_list, transforms=None, thread_num=2):
        """预测。

        Args:
            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）
                且类型为float32且为BGR格式的数组。
            transforms (paddlex.det.transforms): 数据预处理操作。
            thread_num (int): 并发执行各图像预处理时的线程数。
        Returns:
            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、
              预测框类别名称、预测框坐标(坐标格式为[xmin, ymin, w, h]）、
              预测框得分组成。
        """
        if transforms is None and not hasattr(self, 'test_transforms'):
            raise Exception("transforms need to be defined, now is None.")

        if not isinstance(img_file_list, (list, tuple)):
            raise Exception("im_file must be list/tuple")

        if transforms is None:
            transforms = self.test_transforms
        im, im_size = YOLOv3._preprocess(img_file_list, transforms,
                                         self.model_type,
                                         self.__class__.__name__, thread_num)

        with fluid.scope_guard(self.scope):
            result = self.exe.run(self.test_prog,
                                  feed={'image': im,
                                        'im_size': im_size},
                                  fetch_list=list(self.test_outputs.values()),
                                  return_numpy=False,
                                  use_program_cache=True)

        res = {
            k: (np.array(v), v.recursive_sequence_lengths())
            for k, v in zip(list(self.test_outputs.keys()), result)
        }
        res['im_id'] = (np.array(
            [[i] for i in range(len(img_file_list))]).astype('int32'), [[]])
        preds = YOLOv3._postprocess(res,
                                    len(img_file_list), self.num_classes,
                                    self.labels)
        return preds