Created by: GitLD

基于Fruit数据集 训练配置：基于PaddleDetection/configs/faster_rcnn_se154_vd_fpn_s1x.yml改写，主要修改数据集为fruit-detection，和评估指标VOC,11Point 训练过程给出的MAP： 实际评估的MAP： 但是同样的问题在YoloV3配置训练上没有问题： 训练MAP 评估MAP 评估代码如下： calculate_map.py

import os
import json
import numpy as np
import xml.etree.ElementTree as ET
from map_utils import DetectionMAP

import argparse

def parse_args():
    parser = argparse.ArgumentParser("Evaluation Parameters")
    parser.add_argument(
        '--anno_dir',
        type=str,
        default='/home/aistudio/work/insects/val/annotations/xmls',
        help='the directory of annotations')
    parser.add_argument(
        '--pred_result',
        type=str,
        default='./pred_results.json',
        help='the file path of prediction results')
    args = parser.parse_args()
    return args
args = parse_args()
# 存放测试集标签的目录
annotations_dir = args.anno_dir
# 存放预测结果的文件
pred_result_file = args.pred_result
# pred_result_file 中保存着测试结果，是包含所有图片预测结果的list，其构成如下
# [[img_name, [[label, score, x1, x2, y1, y2],..., [label, score, x1, x2, y1, y2]]], 
overlap_thresh = 0.5
map_type = '11point'
fruit_name = ['apple', 'orange', 'banana']

if __name__ == '__main__':
    cname2cid = {}
    for i, item in enumerate(fruit_name):
        cname2cid[item] = i
    filename = pred_result_file
    results = json.load(open(filename))
    num_classes = len(fruit_name)
    detection_map = DetectionMAP(class_num=num_classes,
                        overlap_thresh=overlap_thresh,
                        map_type=map_type,
                        is_bbox_normalized=False,
                        evaluate_difficult=False)
    for result in results:
        image_name = str(result[0])
        bboxes = np.array(result[1]).astype('float32')
        anno_file = os.path.join(annotations_dir, image_name + '.xml')
        tree = ET.parse(anno_file)
        objs = tree.findall('object')
        im_w = float(tree.find('size').find('width').text)
        im_h = float(tree.find('size').find('height').text)
        gt_bbox = np.zeros((len(objs), 4), dtype=np.float32)
        gt_class = np.zeros((len(objs), 1), dtype=np.int32)
        difficult = np.zeros((len(objs), 1), dtype=np.int32)
        for i, obj in enumerate(objs):
            cname = obj.find('name').text
            gt_class[i][0] = cname2cid[cname]
            _difficult = int(obj.find('difficult').text)
            x1 = float(obj.find('bndbox').find('xmin').text)
            y1 = float(obj.find('bndbox').find('ymin').text)
            x2 = float(obj.find('bndbox').find('xmax').text)
            y2 = float(obj.find('bndbox').find('ymax').text)
            x1 = max(0, x1)
            y1 = max(0, y1)
            x2 = min(im_w - 1, x2)
            y2 = min(im_h - 1, y2)
            gt_bbox[i] = [x1, y1, x2, y2]
            difficult[i][0] = _difficult
        detection_map.update(bboxes, gt_bbox, gt_class, difficult)

    print("Accumulating evaluatation results...")
    detection_map.accumulate()
    map_stat = 100. * detection_map.get_map()
    print("mAP({:.2f}, {}) = {:.2f}".format(overlap_thresh,
                            map_type, map_stat))

map_utils.py

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals

import sys
import numpy as np
import logging

logger = logging.getLogger(__name__)

__all__ = ['bbox_area', 'jaccard_overlap', 'DetectionMAP']


def bbox_area(bbox, is_bbox_normalized):
    """
    Calculate area of a bounding box
    """
    norm = 1. - float(is_bbox_normalized)
    width = bbox[2] - bbox[0] + norm
    height = bbox[3] - bbox[1] + norm
    return width * height


def jaccard_overlap(pred, gt, is_bbox_normalized=False):
    """
    Calculate jaccard overlap ratio between two bounding box
    """
    if pred[0] >= gt[2] or pred[2] <= gt[0] or \
        pred[1] >= gt[3] or pred[3] <= gt[1]:
        return 0.
    inter_xmin = max(pred[0], gt[0])
    inter_ymin = max(pred[1], gt[1])
    inter_xmax = min(pred[2], gt[2])
    inter_ymax = min(pred[3], gt[3])
    inter_size = bbox_area([inter_xmin, inter_ymin, inter_xmax, inter_ymax],
                           is_bbox_normalized)
    pred_size = bbox_area(pred, is_bbox_normalized)
    gt_size = bbox_area(gt, is_bbox_normalized)
    overlap = float(inter_size) / (pred_size + gt_size - inter_size)
    return overlap


class DetectionMAP(object):
    """
    Calculate detection mean average precision.
    Currently support two types: 11point and integral

    Args:
        class_num (int): the class number.
        overlap_thresh (float): The threshold of overlap
            ratio between prediction bounding box and 
            ground truth bounding box for deciding 
            true/false positive. Default 0.5.
        map_type (str): calculation method of mean average
            precision, currently support '11point' and
            'integral'. Default '11point'.
        is_bbox_normalized (bool): whther bounding boxes
            is normalized to range[0, 1]. Default False.
        evaluate_difficult (bool): whether to evaluate
            difficult bounding boxes. Default False.
    """

    def __init__(self,
                 class_num,
                 overlap_thresh=0.5,
                 map_type='11point',
                 is_bbox_normalized=False,
                 evaluate_difficult=False):
        self.class_num = class_num
        self.overlap_thresh = overlap_thresh
        assert map_type in ['11point', 'integral'], \
                "map_type currently only support '11point' "\
                "and 'integral'"
        self.map_type = map_type
        self.is_bbox_normalized = is_bbox_normalized
        self.evaluate_difficult = evaluate_difficult
        self.reset()

    def update(self, bbox, gt_box, gt_label, difficult=None):
        """
        Update metric statics from given prediction and ground
        truth infomations.
        """
        if difficult is None:
            difficult = np.zeros_like(gt_label)

        # record class gt count
        for gtl, diff in zip(gt_label, difficult):
            if self.evaluate_difficult or int(diff) == 0:
                self.class_gt_counts[int(np.array(gtl))] += 1

        # record class score positive
        visited = [False] * len(gt_label)
        for b in bbox:
            label, score, xmin, ymin, xmax, ymax = b.tolist()
            pred = [xmin, ymin, xmax, ymax]
            max_idx = -1
            max_overlap = -1.0
            for i, gl in enumerate(gt_label):
                if int(gl) == int(label):
                    overlap = jaccard_overlap(pred, gt_box[i],
                                              self.is_bbox_normalized)
                    #print('---------', int(gl), int(label), overlap)
                    if overlap > max_overlap:
                        max_overlap = overlap
                        max_idx = i

            if max_overlap > self.overlap_thresh:
                if self.evaluate_difficult or \
                        int(np.array(difficult[max_idx])) == 0:
                    if not visited[max_idx]:
                        self.class_score_poss[int(label)].append([score, 1.0])
                        visited[max_idx] = True
                    else:
                        self.class_score_poss[int(label)].append([score, 0.0])
            else:
                self.class_score_poss[int(label)].append([score, 0.0])

    def reset(self):
        """
        Reset metric statics
        """
        self.class_score_poss = [[] for _ in range(self.class_num)]
        self.class_gt_counts = [0] * self.class_num
        self.mAP = None

    def accumulate(self):
        """
        Accumulate metric results and calculate mAP
        """
        mAP = 0.
        valid_cnt = 0
        for score_pos, count in zip(self.class_score_poss,
                                    self.class_gt_counts):
            if count == 0 or len(score_pos) == 0:
                continue

            accum_tp_list, accum_fp_list = \
                    self._get_tp_fp_accum(score_pos)
            precision = []
            recall = []
            for ac_tp, ac_fp in zip(accum_tp_list, accum_fp_list):
                precision.append(float(ac_tp) / (ac_tp + ac_fp))
                recall.append(float(ac_tp) / count)

            if self.map_type == '11point':
                max_precisions = [0.] * 11
                start_idx = len(precision) - 1
                for j in range(10, -1, -1):
                    for i in range(start_idx, -1, -1):
                        if recall[i] < float(j) / 10.:
                            start_idx = i
                            if j > 0:
                                max_precisions[j - 1] = max_precisions[j]
                                break
                        else:
                            if max_precisions[j] < precision[i]:
                                max_precisions[j] = precision[i]
                mAP += sum(max_precisions) / 11.
                valid_cnt += 1
            elif self.map_type == 'integral':
                import math
                ap = 0.
                prev_recall = 0.
                for i in range(len(precision)):
                    recall_gap = math.fabs(recall[i] - prev_recall)
                    if recall_gap > 1e-6:
                        ap += precision[i] * recall_gap
                        prev_recall = recall[i]
                mAP += ap
                valid_cnt += 1
            else:
                logger.error("Unspported mAP type {}".format(self.map_type))
                sys.exit(1)

        self.mAP = mAP / float(valid_cnt) if valid_cnt > 0 else mAP

    def get_map(self):
        """
        Get mAP result
        """
        if self.mAP is None:
            logger.error("mAP is not calculated.")
        return self.mAP

    def _get_tp_fp_accum(self, score_pos_list):
        """
        Calculate accumulating true/false positive results from
        [score, pos] records
        """
        sorted_list = sorted(score_pos_list, key=lambda s: s[0], reverse=True)
        accum_tp = 0
        accum_fp = 0
        accum_tp_list = []
        accum_fp_list = []
        for (score, pos) in sorted_list:
            accum_tp += int(pos)
            accum_tp_list.append(accum_tp)
            accum_fp += 1 - int(pos)
            accum_fp_list.append(accum_fp)
        return accum_tp_list, accum_fp_list

推断文件

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import os
import glob
import json

import numpy as np
from PIL import Image


def set_paddle_flags(**kwargs):
    for key, value in kwargs.items():
        if os.environ.get(key, None) is None:
            os.environ[key] = str(value)


# NOTE(paddle-dev): All of these flags should be set before
# `import paddle`. Otherwise, it would not take any effect.
set_paddle_flags(
    FLAGS_eager_delete_tensor_gb=0,  # enable GC to save memory
)

from paddle import fluid

from ppdet.utils.cli import print_total_cfg
from ppdet.core.workspace import load_config, merge_config, create
from ppdet.modeling.model_input import create_feed
from ppdet.data.data_feed import create_reader

from ppdet.utils.eval_utils import parse_fetches
from ppdet.utils.cli import ArgsParser
from ppdet.utils.check import check_gpu
from ppdet.utils.visualizer import visualize_results
import ppdet.utils.checkpoint as checkpoint

import logging
FORMAT = '%(asctime)s-%(levelname)s: %(message)s'
logging.basicConfig(level=logging.INFO, format=FORMAT)
logger = logging.getLogger(__name__)


def get_save_image_name(output_dir, image_path):
    """
    Get save image name from source image path.
    """
    if not os.path.exists(output_dir):
        os.makedirs(output_dir)
    image_name = os.path.split(image_path)[-1]
    name, ext = os.path.splitext(image_name)
    return os.path.join(output_dir, "{}".format(name)) + ext


def get_test_images(infer_dir, infer_img):
    """
    Get image path list in TEST mode
    """
    assert infer_img is not None or infer_dir is not None, \
        "--infer_img or --infer_dir should be set"
    assert infer_img is None or os.path.isfile(infer_img), \
            "{} is not a file".format(infer_img)
    assert infer_dir is None or os.path.isdir(infer_dir), \
            "{} is not a directory".format(infer_dir)
    images = []

    # infer_img has a higher priority
    if infer_img and os.path.isfile(infer_img):
        images.append(infer_img)
        return images

    infer_dir = os.path.abspath(infer_dir)
    assert os.path.isdir(infer_dir), \
        "infer_dir {} is not a directory".format(infer_dir)
    exts = ['jpg', 'jpeg', 'png', 'bmp']
    exts += [ext.upper() for ext in exts]
    for ext in exts:
        images.extend(glob.glob('{}/*.{}'.format(infer_dir, ext)))

    assert len(images) > 0, "no image found in {}".format(infer_dir)
    logger.info("Found {} inference images in total.".format(len(images)))

    return images


def main():
    cfg = load_config(FLAGS.config)

    if 'architecture' in cfg:
        main_arch = cfg.architecture
    else:
        raise ValueError("'architecture' not specified in config file.")

    merge_config(FLAGS.opt)

    # check if set use_gpu=True in paddlepaddle cpu version
    check_gpu(cfg.use_gpu)
    print_total_cfg(cfg)

    if 'test_feed' not in cfg:
        test_feed = create(main_arch + 'TestFeed')
    else:
        test_feed = create(cfg.test_feed)

    test_images = get_test_images(FLAGS.infer_dir, FLAGS.infer_img)
    test_feed.dataset.add_images(test_images)

    place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace()
    exe = fluid.Executor(place)

    model = create(main_arch)

    startup_prog = fluid.Program()
    infer_prog = fluid.Program()
    with fluid.program_guard(infer_prog, startup_prog):
        with fluid.unique_name.guard():
            _, feed_vars = create_feed(test_feed, use_pyreader=False)
            test_fetches = model.test(feed_vars)
    infer_prog = infer_prog.clone(True)

    reader = create_reader(test_feed)
    feeder = fluid.DataFeeder(place=place, feed_list=feed_vars.values())

    exe.run(startup_prog)
    if cfg.weights:
        checkpoint.load_params(exe, infer_prog, cfg.weights)

    # parse infer fetches
    assert cfg.metric in ['COCO', 'VOC', 'WIDERFACE'], \
            "unknown metric type {}".format(cfg.metric)
    extra_keys = []
    if cfg['metric'] == 'COCO':
        extra_keys = ['im_info', 'im_id', 'im_shape']
    if cfg['metric'] == 'VOC' or cfg['metric'] == 'WIDERFACE':
        extra_keys = ['im_id', 'im_shape']
    keys, values, _ = parse_fetches(test_fetches, infer_prog, extra_keys)

    # parse dataset category
    if cfg.metric == 'COCO':
        from ppdet.utils.coco_eval import bbox2out, mask2out, get_category_info
    if cfg.metric == "VOC":
        from ppdet.utils.voc_eval import bbox2out, get_category_info
    if cfg.metric == "WIDERFACE":
        from ppdet.utils.widerface_eval_utils import bbox2out, get_category_info

    anno_file = getattr(test_feed.dataset, 'annotation', None)
    with_background = getattr(test_feed, 'with_background', True)
    use_default_label = getattr(test_feed, 'use_default_label', False)
    clsid2catid, catid2name = get_category_info(anno_file, with_background,
                                                use_default_label)

    # whether output bbox is normalized in model output layer
    is_bbox_normalized = False
    if hasattr(model, 'is_bbox_normalized') and \
            callable(model.is_bbox_normalized):
        is_bbox_normalized = model.is_bbox_normalized()

    # use tb-paddle to log image
    if FLAGS.use_tb:
        from tb_paddle import SummaryWriter
        tb_writer = SummaryWriter(FLAGS.tb_log_dir)
        tb_image_step = 0
        tb_image_frame = 0  # each frame can display ten pictures at most. 

    imid2path = reader.imid2path
    total_results = []
    for iter_id, data in enumerate(reader()):
        outs = exe.run(infer_prog,
                       feed=feeder.feed(data),
                       fetch_list=values,
                       return_numpy=False)
        res = {
            k: (np.array(v), v.recursive_sequence_lengths())
            for k, v in zip(keys, outs)
        }
        logger.info('Infer iter {}'.format(iter_id))

        bbox_results = None
        mask_results = None
        if 'bbox' in res:
            bbox_results = bbox2out([res], clsid2catid, is_bbox_normalized)
        if 'mask' in res:
            mask_results = mask2out([res], clsid2catid,
                                    model.mask_head.resolution)
        print(f"Num of BBox : {len(bbox_results)}")

        # Create json
        im_ids = res['im_id'][0]
        result = []
        for im_id in im_ids:
            image_path = imid2path[int(im_id)]
            img_name = image_path.split('/')[-1].split('.')[0]
        result.append(img_name)
        bbox_list = []
        for dt in np.array(bbox_results):
            catid, bbox, score = dt['category_id'], dt['bbox'], dt['score']
            xmin, ymin, w, h = bbox
            xmax = xmin + w
            ymax = ymin + h
            bbox_list.append([catid, score, xmin, ymin, xmax, ymax])
        result.append(bbox_list)
        total_results.append(result)
        
    json.dump(total_results, open(os.path.join(FLAGS.output_dir,'pred_results.json'), 'w'))
    print("Output finish!")

if __name__ == '__main__':
    parser = ArgsParser()
    parser.add_argument(
        "--infer_dir",
        type=str,
        default=None,
        help="Directory for images to perform inference on.")
    parser.add_argument(
        "--infer_img",
        type=str,
        default=None,
        help="Image path, has higher priority over --infer_dir")
    parser.add_argument(
        "--output_dir",
        type=str,
        default="output",
        help="Directory for storing the output visualization files.")
    parser.add_argument(
        "--draw_threshold",
        type=float,
        default=0.5,
        help="Threshold to reserve the result for visualization.")
    parser.add_argument(
        "--use_tb",
        type=bool,
        default=False,
        help="whether to record the data to Tensorboard.")
    parser.add_argument(
        '--tb_log_dir',
        type=str,
        default="tb_log_dir/image",
        help='Tensorboard logging directory for image.')
    FLAGS = parser.parse_args()
    main()