系列文章
前言
本项为gluoncv官方源码解析
一、voc格式的mAP计算源码解析
先介绍如何使用:
根据网络的输出和label压入数组中保存,通过调用eval_metric.update()函数,来更新mAP值。
eval_metric.get()函数返回mAP的计算结果。
# 声明一个评价函数对象,传入之后的validate函数中
val_metric = voc_metric.VOC07MApMetric(iou_thresh=0.5, class_names=val_dataset.classes)
def validate(net, val_data, ctx, eval_metric):
"""Test on validation dataset."""
eval_metric.reset()
# set nms threshold and topk constraint
net.set_nms(nms_thresh=0.45, nms_topk=400)
mx.nd.waitall()
net.hybridize()
for batch in val_data:
data = gluon.utils.split_and_load(batch[0], ctx_list=ctx, batch_axis=0, even_split=False)
label = gluon.utils.split_and_load(batch[1], ctx_list=ctx, batch_axis=0, even_split=False)
det_bboxes = []
det_ids = []
det_scores = []
gt_bboxes = []
gt_ids = []
gt_difficults = []
for x, y in zip(data, label):
# get prediction results
ids, scores, bboxes = net(x)
det_ids.append(ids)
det_scores.append(scores)
# clip to image size
det_bboxes.append(bboxes.clip(0, batch[0].shape[2]))
# split ground truths
gt_ids.append(y.slice_axis(axis=-1, begin=4, end=5))
gt_bboxes.append(y.slice_axis(axis=-1, begin=0, end=4))
gt_difficults.append(y.slice_axis(axis=-1, begin=5, end=6) if y.shape[-1] > 5 else None)
# update metric
eval_metric.update(det_bboxes, det_ids, det_scores, gt_bboxes, gt_ids, gt_difficults)
return eval_metric.get()
VOC07MApMetric类继承自EvalMetric类,只是重写了计算mAP的方法,具体每个类别的AP计算方法,在EvalMetric类中。
"""Pascal VOC Detection evaluation."""
from __future__ import division
from collections import defaultdict
import numpy as np
import mxnet as mx
from yolov3.bbox import bbox_iou
class VOCMApMetric(mx.metric.EvalMetric):
"""
Calculate mean AP for object detection task
Parameters:
---------
iou_thresh : float
IOU overlap threshold for TP
class_names : list of str
optional, if provided, will print out AP for each class
"""
def __init__(self, iou_thresh=0.5, class_names=None):
super(VOCMApMetric, self).__init__('VOCMeanAP')
if class_names is None:
self.num = None
else:
assert isinstance(class_names, (list, tuple))
for name in class_names:
assert isinstance(name, str), "must provide names as str"
num = len(class_names)
self.name = list(class_names) + ['mAP']
self.num = num + 1
self.reset()
self.iou_thresh = iou_thresh
self.class_names = class_names
def reset(self):
"""Clear the internal statistics to initial state."""
if getattr(self, 'num', None) is None:
self.num_inst = 0
self.sum_metric = 0.0 # 存储map
else:
self.num_inst = [0] * self.num
self.sum_metric = [0.0] * self.num # 存储每一类的map值,最后一个存储总的map值
self._n_pos = defaultdict(int) # 默认字典,当字典中没有关键字时返回0
self._score = defaultdict(list) # 当字典中没有关键字时返回[]
self._match = defaultdict(list) # 匹配列表
def get(self):
"""Get the current evaluation result.
Returns
-------
name : str
Name of the metric.
value : float
Value of the evaluation.
"""
self._update() # update metric at this time
if self.num is None:
if self.num_inst == 0:
return (self.name, float('nan'))
else:
return (self.name, self.sum_metric / self.num_inst)
else:
names = ['%s'%(self.name[i]) for i in range(self.num)] # 输出各类的AP值,和最后的mAP值
values = [x / y if y != 0 else float('nan') \
for x, y in zip(self.sum_metric, self.num_inst)]
return (names, values)
# pylint: disable=arguments-differ, too-many-nested-blocks
def update(self, pred_bboxes, pred_labels, pred_scores,
gt_bboxes, gt_labels, gt_difficults=None):
"""Update internal buffer with latest prediction and gt pairs.
Parameters
----------
pred_bboxes : mxnet.NDArray or numpy.ndarray
Prediction bounding boxes with shape `B, N, 4`.
Where B is the size of mini-batch, N is the number of bboxes.
pred_labels : mxnet.NDArray or numpy.ndarray
Prediction bounding boxes labels with shape `B, N`.
pred_scores : mxnet.NDArray or numpy.ndarray
Prediction bounding boxes scores with shape `B, N`.
gt_bboxes : mxnet.NDArray or numpy.ndarray
Ground-truth bounding boxes with shape `B, M, 4`.
Where B is the size of mini-batch, M is the number of ground-truths.
gt_labels : mxnet.NDArray or numpy.ndarray
Ground-truth bounding boxes labels with shape `B, M`.
gt_difficults : mxnet.NDArray or numpy.ndarray, optional, default is None
Ground-truth bounding boxes difficulty labels with shape `B, M`.
"""
def as_numpy(a): # 转化为numpy格式
"""Convert a (list of) mx.NDArray into numpy.ndarray"""
if isinstance(a, (list, tuple)):
out = [x.asnumpy() if isinstance(x, mx.nd.NDArray) else x for x in a]
try:
out = np.concatenate(out, axis=0)
except ValueError:
out = np.array(out)
return out
elif isinstance(a, mx.nd.NDArray):
a = a.asnumpy()
return a
if gt_difficults is None:
gt_difficults = [None for _ in as_numpy(gt_labels)]
if isinstance(gt_labels, list):
gt_diff_shape = gt_difficults[0].shape[0] if hasattr(gt_difficults[0], 'shape') else 0
if len(gt_difficults) * gt_diff_shape != \
len(gt_labels) * gt_labels[0].shape[0]: # 检查gt_difficults和gt_labels形状是否一样
gt_difficults = [None] * len(gt_labels) * gt_labels[0].shape[0]
for pred_bbox, pred_label, pred_score, gt_bbox, gt_label, gt_difficult in zip(
*[as_numpy(x) for x in [pred_bboxes, pred_labels, pred_scores,
gt_bboxes, gt_labels, gt_difficults]]):
# strip padding -1 for pred and gt
valid_pred = np.where(pred_label.flat >= 0)[0] # label>=0为有效的标签,返回有效标签的下标
pred_bbox = pred_bbox[valid_pred, :] # 筛选有效的bbox
pred_label = pred_label.flat[valid_pred].astype(int) # 筛选有效的label
pred_score = pred_score.flat[valid_pred] # 筛选有效的score
valid_gt = np.where(gt_label.flat >= 0)[0]
gt_bbox = gt_bbox[valid_gt, :]
gt_label = gt_label.flat[valid_gt].astype(int)
if gt_difficult is None:
gt_difficult = np.zeros(gt_bbox.shape[0])
else:
gt_difficult = gt_difficult.flat[valid_gt]
for l in np.unique(np.concatenate((pred_label, gt_label)).astype(int)): # unique去除重复数字并排序
pred_mask_l = pred_label == l # 选出类别l
pred_bbox_l = pred_bbox[pred_mask_l] # 筛选pred_label类的bbox
pred_score_l = pred_score[pred_mask_l] # 筛选pred_label类的score
# sort by score
order = pred_score_l.argsort()[::-1] # 按分数降序排列
pred_bbox_l = pred_bbox_l[order]
pred_score_l = pred_score_l[order]
gt_mask_l = gt_label == l
gt_bbox_l = gt_bbox[gt_mask_l]
gt_difficult_l = gt_difficult[gt_mask_l]
self._n_pos[l] += np.logical_not(gt_difficult_l).sum() # 计算类别l有多少个不是difficult的标签
self._score[l].extend(pred_score_l) # 添加l类别的分数
if len(pred_bbox_l) == 0:
continue
if len(gt_bbox_l) == 0:
self._match[l].extend((0,) * pred_bbox_l.shape[0])
continue
# VOC evaluation follows integer typed bounding boxes.
pred_bbox_l = pred_bbox_l.copy()
pred_bbox_l[:, 2:] += 1
gt_bbox_l = gt_bbox_l.copy()
gt_bbox_l[:, 2:] += 1
iou = bbox_iou(pred_bbox_l, gt_bbox_l) # 计算iou
gt_index = iou.argmax(axis=1) # 最大iou的下标
# set -1 if there is no matching ground truth
gt_index[iou.max(axis=1) < self.iou_thresh] = -1 # 没有符合的bbox,则全部设置为-1
del iou
selec = np.zeros(gt_bbox_l.shape[0], dtype=bool)
for gt_idx in gt_index:
if gt_idx >= 0:
if gt_difficult_l[gt_idx]: # difficult=1,不参与ap计算
self._match[l].append(-1) # -1忽略
else:
if not selec[gt_idx]: # 可能有多个bbox与同一个gt_bbox的iou要大于阈值,标注第一个符合的即可
self._match[l].append(1) # 1为正确分类
else:
self._match[l].append(0) # 0为错误分类
selec[gt_idx] = True
else:
self._match[l].append(0)
def _update(self):
""" update num_inst and sum_metric """
aps = []
recall, precs = self._recall_prec()
for l, rec, prec in zip(range(len(precs)), recall, precs):
ap = self._average_precision(rec, prec)
aps.append(ap)
if self.num is not None and l < (self.num - 1):
self.sum_metric[l] = ap
self.num_inst[l] = 1
if self.num is None:
self.num_inst = 1
self.sum_metric = np.nanmean(aps)
else:
self.num_inst[-1] = 1
self.sum_metric[-1] = np.nanmean(aps)
def _recall_prec(self):
""" get recall and precision from internal records """
n_fg_class = max(self._n_pos.keys()) + 1 # 类别数
prec = [None] * n_fg_class # 准确率
rec = [None] * n_fg_class # 召回率
for l in self._n_pos.keys():
score_l = np.array(self._score[l]) # l类别的分数
match_l = np.array(self._match[l], dtype=np.int32)
order = score_l.argsort()[::-1] # 降序排列的下标索引
match_l = match_l[order] # 按分数降序排序
tp = np.cumsum(match_l == 1)
fp = np.cumsum(match_l == 0)
# If an element of fp + tp is 0,
# the corresponding element of prec[l] is nan.
with np.errstate(divide='ignore', invalid='ignore'):
prec[l] = tp / (fp + tp)
# If n_pos[l] is 0, rec[l] is None.
if self._n_pos[l] > 0:
rec[l] = tp / self._n_pos[l]
return rec, prec
def _average_precision(self, rec, prec):
"""
calculate average precision
Params:
----------
rec : numpy.array
cumulated recall
prec : numpy.array
cumulated precision
Returns:
----------
ap as float
"""
if rec is None or prec is None:
return np.nan
# append sentinel values at both ends
mrec = np.concatenate(([0.], rec, [1.]))
mpre = np.concatenate(([0.], np.nan_to_num(prec), [0.]))
# compute precision integration ladder
for i in range(mpre.size - 1, 0, -1):
mpre[i - 1] = np.maximum(mpre[i - 1], mpre[i])
# look for recall value changes
i = np.where(mrec[1:] != mrec[:-1])[0]
# sum (\delta recall) * prec
ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1])
return ap
class VOC07MApMetric(VOCMApMetric):
""" Mean average precision metric for PASCAL V0C 07 dataset
Parameters:
---------
iou_thresh : float
IOU overlap threshold for TP
class_names : list of str
optional, if provided, will print out AP for each class
"""
def __init__(self, *args, **kwargs):
super(VOC07MApMetric, self).__init__(*args, **kwargs)
def _average_precision(self, rec, prec):
"""
calculate average precision, override the default one,
special 11-point metric
Params:
----------
rec : numpy.array
cumulated recall
prec : numpy.array
cumulated precision
Returns:
----------
ap as float
"""
if rec is None or prec is None:
return np.nan
ap = 0.
for t in np.arange(0., 1.1, 0.1):
if np.sum(rec >= t) == 0: # 筛选大于t的召回率,没有则p=0
p = 0
else:
p = np.max(np.nan_to_num(prec)[rec >= t]) # 筛选大于t的召回率,选出此时最大的准确率
ap += p / 11.
return ap