【YOLO系列】YOLOv3代码详解(四)：模型脚本model.py

前言

        以下内容仅为个人在学习人工智能中所记录的笔记，先将目标识别算法yolo系列的整理出来分享给大家，供大家学习参考。

        本文仅对YOLOV3代码中关键部分进行了注释，未掌握基础代码的铁汁可以自己百度一下。

        若文中内容有误，希望大家批评指正。

---

资料下载

        YOLOV3论文下载地址：YOLOv3：An Incremental Improvement

回顾

        YOLO V1：【YOLO系列】YOLO V1论文思想详解

        YOLO V2：【YOLO系列】YOLO V2论文思想详解

        YOLO V3：【YOLO系列】 YOLOv3论文思想详解

项目地址

        YOLOV3 keras版本：下载地址

        YOLOV3 Tensorflow版本：下载地址

        YOLOV3 Pytorch版本：下载地址

Gitee仓库

        YOLOV3 各版本：yolov3各版本

---

YOLO V3代码详解

        YOLO V3代码详解（一）：【YOLO系列】YOLOv3代码详解(一)：主脚本yolo_video.py

        YOLO V3代码详解（二）：【YOLO系列】YOLOv3代码详解(二)：检测脚本yolo.py

        YOLO V3代码详解（三）：【YOLO系列】YOLOv3代码详解(三)：训练脚本train.py

---

        本文主要基于keras版本进行讲解

        话不多说，直接上代码

---

一、代码详解

1、定义卷积神经网络函数

@wraps(Conv2D)
def DarknetConv2D(*args, **kwargs):"""Wrapper to set Darknet parameters for Convolution2D."""# 定义一个darknet_conv_kwargs字典，传递“kernel_regularizer”、“padding”参数darknet_conv_kwargs = {'kernel_regularizer': l2(5e-4)}# 如果输入的kwargs中定义了strides为（2,2），则padding模式为valid，否则为same模式darknet_conv_kwargs['padding'] = 'valid' if kwargs.get('strides') == (2, 2) else 'same'# 将输入的kwargs值更新到darknet_conv_kwargs字典中darknet_conv_kwargs.update(kwargs)return Conv2D(*args, **darknet_conv_kwargs)def DarknetConv2D_BN_Leaky(*args, **kwargs):"""Darknet Convolution2D followed by BatchNormalization and LeakyReLU."""# 定义一个no_bias_kwargs字典no_bias_kwargs = {'use_bias': False}# 将传递里面的kwargs值更新到no_bias_kwargs字典中no_bias_kwargs.update(kwargs)# 返回一个组合函数，由DarknetConv2D、BN、LeakyRelu组成，LeakyRelu的alpha值为0.1，这意味着当输入值小于 0 时，输出为 0.1 倍的输入值；当输入值大于等于 0 时，输出为输入值本身。return compose(# 定义一个Conv2D层DarknetConv2D(*args, **no_bias_kwargs),BatchNormalization(),LeakyReLU(alpha=0.1))

2、定义残差结构块函数

def resblock_body(x, num_filters, num_blocks):"""A series of resblocks starting with a downsampling Convolution2D"""# Darknet uses left and top padding instead of 'same' mode# 进行零填充# 第一个元组(1, 0)指定了垂直方向（或高度方向）的填充。1表示在顶部填充1行零，0表示在底部不填充。# 第二个元组(1, 0)指定了水平方向（或宽度方向）的填充。1表示在左侧填充1列零，0表示在右侧不填充。x = ZeroPadding2D(((1, 0), (1, 0)))(x)# 创建一个DarknetConv2D_BN_Leaky卷积层，其中包括卷积层（filters=num_filters， kernel_size=(3, 3)，strides=(2, 2)，padding="same"）、归一化层BN、激活函数层LeakyRule# 这里strides=(2, 2)，代替了池化的作用x = DarknetConv2D_BN_Leaky(num_filters, (3, 3), strides=(2, 2))(x)# 残差结构for i in range(num_blocks):y = compose(DarknetConv2D_BN_Leaky(num_filters // 2, (1, 1)),DarknetConv2D_BN_Leaky(num_filters, (3, 3)))(x)x = Add()([x, y])return x

3、定义darknet_body()函数

def darknet_body(x):'''Darknent body having 52 Convolution2D layers'''# 创建一个DarknetConv2D_BN_Leaky卷积层，其中包括卷积层（filters=32， kernel_size=(3, 3)，strides=(1, 1)，padding="same"）、归一化层BN、激活函数层LeakyRulex = DarknetConv2D_BN_Leaky(32, (3, 3))(x)# 残差结构，（输入，filter数量, 残差block数量）x = resblock_body(x, 64, 1)x = resblock_body(x, 128, 2)x = resblock_body(x, 256, 8)x = resblock_body(x, 512, 8)x = resblock_body(x, 1024, 4)return x

4、定义最后输出层的神经网络函数

def make_last_layers(x, num_filters, out_filters):'''6 Conv2D_BN_Leaky layers followed by a Conv2D_linear layer'''x = compose(DarknetConv2D_BN_Leaky(num_filters, (1, 1)),DarknetConv2D_BN_Leaky(num_filters * 2, (3, 3)),DarknetConv2D_BN_Leaky(num_filters, (1, 1)),DarknetConv2D_BN_Leaky(num_filters * 2, (3, 3)),DarknetConv2D_BN_Leaky(num_filters, (1, 1)))(x)y = compose(DarknetConv2D_BN_Leaky(num_filters * 2, (3, 3)),DarknetConv2D(out_filters, (1, 1)))(x)return x, y

5、定义输出倒数三个特征图函数

def yolo_body(inputs, num_anchors, num_classes):"""Create YOLO_V3 model CNN body in Keras."""darknet = Model(inputs, darknet_body(inputs))# 输出三个特征图# 输出层的最后计算，包括6个Conv2D_BN_Leaky层和1个Conv2D_linear层x, y1 = make_last_layers(darknet.output, 512, num_anchors * (num_classes + 5))# 最后一层输出层进行Conv2D_BN_Leaky层与上采样操作后，与第152层的输出拼接x = compose(DarknetConv2D_BN_Leaky(256, (1, 1)),UpSampling2D(2))(x)x = Concatenate()([x, darknet.layers[152].output])x, y2 = make_last_layers(x, 256, num_anchors * (num_classes + 5))# 倒数第二层输出层进行Conv2D_BN_Leaky层与上采样操作后，与第92层的输出拼接x = compose(DarknetConv2D_BN_Leaky(128, (1, 1)),UpSampling2D(2))(x)x = Concatenate()([x, darknet.layers[92].output])x, y3 = make_last_layers(x, 128, num_anchors * (num_classes + 5))return Model(inputs, [y1, y2, y3])

6、定义tiny model的输出特征图函数

def tiny_yolo_body(inputs, num_anchors, num_classes):"""在keras架构上创建一个tiny YOLOV3模型，由8个CNN层+6个池化层+上采样层（CNN+upsampling）+2个输出（2个CNN+2个Conv）构成，总共20层"""# 生成一个卷积组合x1，输入为inputs，由5个DarknetConv2D_BN_Leaky与4个池化层构成x1 = compose(# 创建一个DarknetConv2D_BN_Leaky卷积层，其中包括卷积层（filters=16， kernel_size=(3, 3)，strides=(1, 1)，padding="same"）、归一化层BN、激活函数层LeakyRuleDarknetConv2D_BN_Leaky(16, (3, 3)),# 池化层，池化框尺寸为(2,2)，步长为(2,2)，表示特征图缩小4倍，即宽和高各缩小2倍；padding模式为same，'same'表示在输入特征图的边缘填充0，使得经过池化后，输出特征图的大小与输入特征图一致MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same'),DarknetConv2D_BN_Leaky(32, (3, 3)),MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same'),DarknetConv2D_BN_Leaky(64, (3, 3)),MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same'),DarknetConv2D_BN_Leaky(128, (3, 3)),MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same'),DarknetConv2D_BN_Leaky(256, (3, 3)))(inputs)# 生成一个卷积组合x2，输入为x1，由3个DarknetConv2D_BN_Leaky与2个池化层构成x2 = compose(MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same'),DarknetConv2D_BN_Leaky(512, (3, 3)),MaxPooling2D(pool_size=(2, 2), strides=(1, 1), padding='same'),DarknetConv2D_BN_Leaky(1024, (3, 3)),DarknetConv2D_BN_Leaky(256, (1, 1)))(x1)# 生成一个预测层，输入为x2，由1个DarknetConv2D_BN_Leaky与1个卷积层构成，输出一个N*N*Anchor个数*(类别数量+5)的tensory1 = compose(DarknetConv2D_BN_Leaky(512, (3, 3)),DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x2)# 生成一个卷积组合x2，输入为x2，由1个DarknetConv2D_BN_Leaky与1个上采样层构成x2 = compose(DarknetConv2D_BN_Leaky(128, (1, 1)),UpSampling2D(2))(x2)# 将经过上采样的x2与x1拼接在一起，再1个DarknetConv2D_BN_Leaky层与1个卷积层，输出一个N*N*Anchor个数*(类别数量+5)的tensory2 = compose(Concatenate(),DarknetConv2D_BN_Leaky(256, (3, 3)),DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))([x2, x1])return Model(inputs, [y1, y2])

7、计算bbox坐标、置信度与类别概率

def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):"""Convert final layer features to bounding box parameters.预测box的坐标，置信度与分类"""num_anchors = len(anchors)# 生成一个tensor，形状为(batch, height, width, num_anchors, box_params).anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])# 获取输出层的height, width的维度grid_shape = K.shape(feats)[1:3]# 绘制x、y坐标，y-height， x-width# K.arange(0, stop=grid_shape[0]) 表示生成一个0-(grid_shape[0]-1)的张量grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),[1, grid_shape[1], 1, 1])grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),[grid_shape[0], 1, 1, 1])grid = K.concatenate([grid_x, grid_y])grid = K.cast(grid, K.dtype(feats))feats = K.reshape(feats, [-1, grid_shape[0], grid_shape[1], num_anchors, 5 + num_classes])# 这一步对应论文中Bounding box Prediction.同时做了归一化box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))# 获取置信度值与分类值box_confidence = K.sigmoid(feats[..., 4:5])box_class_probs = K.sigmoid(feats[..., 5:])# 计算坐标损失if calc_loss == True:return grid, feats, box_xy, box_whreturn box_xy, box_wh, box_confidence, box_class_probs

8、修正bbox坐标

def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):"""Get corrected boxes修正box的坐标将得到的特征图与原图相比，求出偏移量，修正box的坐标"""box_yx = box_xy[..., ::-1]box_hw = box_wh[..., ::-1]input_shape = K.cast(input_shape, K.dtype(box_yx))image_shape = K.cast(image_shape, K.dtype(box_yx))# 新生成一个以（input_shape / image_shape）中最小比例的尺寸图片new_shape = K.round(image_shape * K.min(input_shape / image_shape))# 计算新生成的最小比例的图片与放大后的特征图的相对偏移量offset = (input_shape - new_shape) / 2. / input_shape# 计算放大后的特征图与新生成的最小比例的图片的比例scale = input_shape / new_shape# 修正box的坐标box_yx = (box_yx - offset) * scalebox_hw *= scalebox_mins = box_yx - (box_hw / 2.)box_maxes = box_yx + (box_hw / 2.)boxes = K.concatenate([box_mins[..., 0:1],  # y_minbox_mins[..., 1:2],  # x_minbox_maxes[..., 0:1],  # y_maxbox_maxes[..., 1:2]  # x_max])# Scale boxes back to original image shape.# 反归一化，求得box在输入图片的实际坐标值boxes *= K.concatenate([image_shape, image_shape])return boxes

9、预测box的坐标(x, y, w, h)，置信度与分类（用于预测）

def yolo_boxes_and_scores(feats, anchors, num_classes, input_shape, image_shape):"""Process Conv layer outputfeats: 输出层，shape=(m,N,N,3,5+80)anchors: 输出层对应的Anchornum_classes：类别的数量input_shape： 特征图放大32倍的尺寸image_shape：输入图片的大小"""# 预测box的坐标(x, y, w, h)，置信度与分类box_xy, box_wh, box_confidence, box_class_probs = yolo_head(feats,anchors, num_classes, input_shape)# 修正每个特征图中box的坐标boxes = yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape)boxes = K.reshape(boxes, [-1, 4])# 计算每个box的置信度box_scores = box_confidence * box_class_probsbox_scores = K.reshape(box_scores, [-1, num_classes])return boxes, box_scores

10、评估函数()

def yolo_eval(yolo_outputs,anchors,num_classes,image_shape,max_boxes=20,score_threshold=.6,iou_threshold=.5):"""评估函数Evaluate YOLO model on given input and return filtered boxes.yolo_outputs：输出层，shape=(m,N,N,3,5+80)anchors：Anchor Boxnum_classes：类别的数量image_shape：输入图像的尺寸max_boxes：box的最大数量score_threshold：预测分数的阈值iou_threshold：IOU的阈值"""# 将Anchor Box与输出层对应num_layers = len(yolo_outputs)anchor_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]] if num_layers == 3 else [[3, 4, 5], [1, 2, 3]]  # default setting# 将特征图尺寸放大32倍input_shape = K.shape(yolo_outputs[0])[1:3] * 32boxes = []box_scores = []for l in range(num_layers):# 计算输出的box与分值_boxes, _box_scores = yolo_boxes_and_scores(yolo_outputs[l],anchors[anchor_mask[l]], num_classes, input_shape, image_shape)boxes.append(_boxes)box_scores.append(_box_scores)boxes = K.concatenate(boxes, axis=0)box_scores = K.concatenate(box_scores, axis=0)# 筛选出分值大于阈值的mask = box_scores >= score_thresholdmax_boxes_tensor = K.constant(max_boxes, dtype='int32')boxes_ = []scores_ = []classes_ = []for c in range(num_classes):# TODO: use keras backend instead of tf.# 将box_scores >= score_threshold的box，box score取出来class_boxes = tf.boolean_mask(boxes, mask[:, c])class_box_scores = tf.boolean_mask(box_scores[:, c], mask[:, c])# 非极大值抑制,去除IOU>iou_threshold的框nms_index = tf.image.non_max_suppression(class_boxes, class_box_scores, max_boxes_tensor, iou_threshold=iou_threshold)# 将剩下的class_boxes、class_box_scores、class取出来class_boxes = K.gather(class_boxes, nms_index)class_box_scores = K.gather(class_box_scores, nms_index)classes = K.ones_like(class_box_scores, 'int32') * cboxes_.append(class_boxes)scores_.append(class_box_scores)classes_.append(classes)boxes_ = K.concatenate(boxes_, axis=0)scores_ = K.concatenate(scores_, axis=0)classes_ = K.concatenate(classes_, axis=0)return boxes_, scores_, classes_

11、对GT框进行预处理

def preprocess_true_boxes(true_boxes, input_shape, anchors, num_classes):'''Preprocess true boxes to training input formatParameters----------true_boxes: array, shape=(m, T, 5)Absolute x_min, y_min, x_max, y_max, class_id relative to input_shape.input_shape: array-like, hw, multiples of 32anchors: array, shape=(N, 2), whnum_classes: integerReturns-------y_true: list of array, shape like yolo_outputs, xywh are reletive value'''# 首先判断GT框中的class_id是否超过了类别的总数assert (true_boxes[..., 4] < num_classes).all(), 'class id must be less than num_classes'# 判断Anchor Box是否能分为3组,并指定每一组中Anchor Box的索引值# 这里对应原文中 作者选择了9种不同Anchor Box来对3种不同的尺度进行预测# 特征图较大的用较小的Anchor([0, 1, 2])去预测,特征图较小的用较大的Anchor([6, 7, 8])去预测num_layers = len(anchors) // 3  # default settinganchor_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]] if num_layers == 3 else [[3, 4, 5], [1, 2, 3]]true_boxes = np.array(true_boxes, dtype='float32')input_shape = np.array(input_shape, dtype='int32')# 计算GT框的中心点左边与宽、高,boxes_xy.shape=(m, T, 2)=boxes_wh.shapeboxes_xy = (true_boxes[..., 0:2] + true_boxes[..., 2:4]) // 2boxes_wh = true_boxes[..., 2:4] - true_boxes[..., 0:2]true_boxes[..., 0:2] = boxes_xy / input_shape[::-1]true_boxes[..., 2:4] = boxes_wh / input_shape[::-1]m = true_boxes.shape[0]# 生成倒数三层输出层的特征图大小（13,13），（26,26），（52,52）grid_shapes = [input_shape // {0: 32, 1: 16, 2: 8}[l] for l in range(num_layers)]# 创建倒数三层输出层的y_true零数组（m，13,13,3,5+80），（m，26,26,3,5+80），（m，52,52,3,5+80）y_true = [np.zeros((m, grid_shapes[l][0], grid_shapes[l][1], len(anchor_mask[l]), 5 + num_classes),dtype='float32') for l in range(num_layers)]# Expand dim to apply broadcasting.# 在anchor box数组中增加一维，shape=(1, N, 2)anchors = np.expand_dims(anchors, 0)anchor_maxes = anchors / 2.anchor_mins = -anchor_maxes# 要求所有维度的第一维元素要>0,返回的数组为（1,n）的bool值valid_mask = boxes_wh[..., 0] > 0for b in range(m):# Discard zero rows.# 判断wh是否为0，若为0则跳过该轮循环, wh.shape=(1, 2)wh = boxes_wh[b, valid_mask[b]]if len(wh) == 0: continue# Expand dim to apply broadcasting.# 在倒数第二维增加一维，wh.shape=(1, 1, 2)wh = np.expand_dims(wh, -2)box_maxes = wh / 2.box_mins = -box_maxesintersect_mins = np.maximum(box_mins, anchor_mins)intersect_maxes = np.minimum(box_maxes, anchor_maxes)intersect_wh = np.maximum(intersect_maxes - intersect_mins, 0.)intersect_area = intersect_wh[..., 0] * intersect_wh[..., 1]box_area = wh[..., 0] * wh[..., 1]anchor_area = anchors[..., 0] * anchors[..., 1]iou = intersect_area / (box_area + anchor_area - intersect_area)# Find best anchor for each true box# 获取与GT IOU最大的anchor box，记为best anchorbest_anchor = np.argmax(iou, axis=-1)# 将这个IOU最大的anchor box对应的GT的y_true记为1for t, n in enumerate(best_anchor):for l in range(num_layers):if n in anchor_mask[l]:i = np.floor(true_boxes[b, t, 0] * grid_shapes[l][1]).astype('int32')j = np.floor(true_boxes[b, t, 1] * grid_shapes[l][0]).astype('int32')k = anchor_mask[l].index(n)c = true_boxes[b, t, 4].astype('int32')y_true[l][b, j, i, k, 0:4] = true_boxes[b, t, 0:4]y_true[l][b, j, i, k, 4] = 1y_true[l][b, j, i, k, 5 + c] = 1return y_true

12、IOU计算

def box_iou(b1, b2):'''Return iou tensorParameters----------b1: tensor, shape=(i1,...,iN, 4), xywhb2: tensor, shape=(j, 4), xywhReturns-------iou: tensor, shape=(i1,...,iN, j)'''# Expand dim to apply broadcasting.b1 = K.expand_dims(b1, -2)b1_xy = b1[..., :2]b1_wh = b1[..., 2:4]b1_wh_half = b1_wh / 2.b1_mins = b1_xy - b1_wh_halfb1_maxes = b1_xy + b1_wh_half# Expand dim to apply broadcasting.b2 = K.expand_dims(b2, 0)b2_xy = b2[..., :2]b2_wh = b2[..., 2:4]b2_wh_half = b2_wh / 2.b2_mins = b2_xy - b2_wh_halfb2_maxes = b2_xy + b2_wh_halfintersect_mins = K.maximum(b1_mins, b2_mins)intersect_maxes = K.minimum(b1_maxes, b2_maxes)intersect_wh = K.maximum(intersect_maxes - intersect_mins, 0.)intersect_area = intersect_wh[..., 0] * intersect_wh[..., 1]b1_area = b1_wh[..., 0] * b1_wh[..., 1]b2_area = b2_wh[..., 0] * b2_wh[..., 1]iou = intersect_area / (b1_area + b2_area - intersect_area)return iou

13、损失计算

def yolo_loss(args, anchors, num_classes, ignore_thresh=.5, print_loss=False):"""Return yolo_loss tensorParameters----------yolo_outputs: list of tensor, the output of yolo_body or tiny_yolo_bodyy_true: list of array, the output of preprocess_true_boxesanchors: array, shape=(N, 2), whnum_classes: integerignore_thresh: float, the iou threshold whether to ignore object confidence lossReturns-------loss: tensor, shape=(1,)"""# 这个默认将Anchor Box分为3组num_layers = len(anchors) // 3  # default setting# 将前num_layers层(不含num_layers层)定义为输出层,yolo_outputs中输出的张量为(batch_size, height, width, channels)yolo_outputs = args[:num_layers]# 将后num_layers层定义为y_true层y_true = args[num_layers:]# 判断Anchor Box是否能分为3组,并指定每一组中Anchor Box的索引值# 这里对应原文中 作者选择了9种不同Anchor Box来对3种不同的尺度进行预测# 特征图较大的用较小的Anchor([0, 1, 2])去预测,特征图较小的用较大的Anchor([6, 7, 8])去预测anchor_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]] if num_layers == 3 else [[3, 4, 5], [1, 2, 3]]# K.cast()函数用于将一个值从一个类型转换为另一个类型# K.shape(yolo_outputs[0])[1:3]表示获取yolo_outputs[0]的第二维与第三维的形状，即（height, width）的形状# 然后再将（height, width）的形状放大32倍input_shape = K.cast(K.shape(yolo_outputs[0])[1:3] * 32, K.dtype(y_true[0]))grid_shapes = [K.cast(K.shape(yolo_outputs[l])[1:3], K.dtype(y_true[0])) for l in range(num_layers)]loss = 0# 获取batch sizem = K.shape(yolo_outputs[0])[0]mf = K.cast(m, K.dtype(yolo_outputs[0]))for l in range(num_layers):# [...]: 用于表示多个冒号，通常用于多维数组的索引，这里代表取第5维：是否为物体object_mask = y_true[l][..., 4:5]# 获取物体正确的分类true_class_probs = y_true[l][..., 5:]# 计算图像的每个像素点坐标grid，输出层raw_pred，shape=(m, N, N, 3, 5+80)，预测box的坐标(x, y, w, h)，grid, raw_pred, pred_xy, pred_wh = yolo_head(yolo_outputs[l],anchors[anchor_mask[l]], num_classes, input_shape, calc_loss=True)pred_box = K.concatenate([pred_xy, pred_wh])# Darknet raw box to calculate loss.# 计算坐标xy偏差raw_true_xy = y_true[l][..., :2] * grid_shapes[l][::-1] - grid# 计算wh的偏移量raw_true_wh = K.log(y_true[l][..., 2:4] / anchors[anchor_mask[l]] * input_shape[::-1])# 当object_mask是物体的时候，返回raw_true_wh，不是物体返回K.zeros_like(raw_true_wh)=0数组raw_true_wh = K.switch(object_mask, raw_true_wh, K.zeros_like(raw_true_wh))  # avoid log(0)=-infbox_loss_scale = 2 - y_true[l][..., 2:3] * y_true[l][..., 3:4]# Find ignore mask, iterate over each of batch.# 创建一个与y_true[0] 相同数据类型的动态数组，初始大小为 1。ignore_mask = tf.TensorArray(K.dtype(y_true[0]), size=1, dynamic_size=True)# 将object_mask数据类型转为boolobject_mask_bool = K.cast(object_mask, 'bool')# 计算某个bbox与ground truth的重合度是否超过某个阈值，超过则不计入损失计算def loop_body(b, ignore_mask):# 取出是物体的box坐标true_box = tf.boolean_mask(y_true[l][b, ..., 0:4], object_mask_bool[b, ..., 0])# 计算预测框与GT框的IOU值iou = box_iou(pred_box[b], true_box)# 输出best_iou < ignore_thresh判断的0,1值best_iou = K.max(iou, axis=-1)ignore_mask = ignore_mask.write(b, K.cast(best_iou < ignore_thresh, K.dtype(true_box)))return b + 1, ignore_mask_, ignore_mask = K.control_flow_ops.while_loop(lambda b, *args: b < m, loop_body, [0, ignore_mask])ignore_mask = ignore_mask.stack()ignore_mask = K.expand_dims(ignore_mask, -1)# K.binary_crossentropy is helpful to avoid exp overflow.# 坐标损失、置信度损失、分类损失计算xy_loss = object_mask * box_loss_scale * K.binary_crossentropy(raw_true_xy, raw_pred[..., 0:2], from_logits=True)wh_loss = object_mask * box_loss_scale * 0.5 * K.square(raw_true_wh - raw_pred[..., 2:4])confidence_loss = object_mask * K.binary_crossentropy(object_mask, raw_pred[..., 4:5], from_logits=True) + (1 - object_mask) * K.binary_crossentropy(object_mask, raw_pred[..., 4:5], from_logits=True) * ignore_maskclass_loss = object_mask * K.binary_crossentropy(true_class_probs, raw_pred[..., 5:], from_logits=True)xy_loss = K.sum(xy_loss) / mfwh_loss = K.sum(wh_loss) / mfconfidence_loss = K.sum(confidence_loss) / mfclass_loss = K.sum(class_loss) / mfloss += xy_loss + wh_loss + confidence_loss + class_lossif print_loss:loss = tf.Print(loss, [loss, xy_loss, wh_loss, confidence_loss, class_loss, K.sum(ignore_mask)], message='loss: ')return loss