说明：
本例程使用YOLOv3进行昆虫检测。例程分为数据处理、模型设计、损失函数、训练模型、模型预测和测试模型六个部分。本篇为第三部分，设计了物体边框、物体置信度和物体类别的损失函数。物体边框的x、y使用sigmoid_cross_entropy_with_logits损失函数，w、h使用绝对值L1损失函数。物体置信度和物体类别使用sigmoid_cross_entropy_with_logits损坏函数。
实验代码：
损失函数输出：
import paddle.fluid as fluid from paddle.fluid.dygraph.base import to_variable from source.data import single_thread_reader from source.model import YOLOv3 from source.loss import get_sum_loss with fluid.dygraph.guard(): # 读取数据 train_set = './dataset/train/' train_reader = single_thread_reader(train_set, 1, 'train') # 单线程读数据 image, gtbox, gtcls, image_size = next(train_reader()) # 读取一条数据 image = to_variable(image) # 转换数据格式 # 前向传播 num_classes = 7 # 类别数量 anchor_size = [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326] # 锚框大小 anchor_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]] # 锚框掩码 ignore_threshold = 0.7 # 样本阈值 downsample_ratio = 32 # 下采样率 model = YOLOv3(num_classes=num_classes, anchor_mask=anchor_mask) infer = model(image) # 计算损失 loss = get_sum_loss(infer, gtbox, gtcls, num_classes, anchor_size, anchor_mask, ignore_threshold, downsample_ratio) print(loss.numpy())
结果：
[8711.687]
loss.py文件
import numpy as np import paddle.fluid as fluid from paddle.fluid.dygraph.base import to_variable def sigmoid(x): """ 功能: 计算sigmoid函数 输入: x - 输入数值 输出: y - 输出数值 """ return 0.5 * (1.0 + np.tanh(0.5 * x)) # def sigmoid(x): # return 1.0 / (1.0 + np.exp(-x)) def get_box_iou_xywh(box1, box2): """ 功能: 计算边框交并比值 输入: box1 - 边界框1 box2 - 边界框2 输出: iou - 交并比值 """ # 计算交集面积 x1_min = box1[0] - box1[2]/2.0 y1_min = box1[1] - box1[3]/2.0 x1_max = box1[0] + box1[2]/2.0 y1_max = box1[1] + box1[3]/2.0 x2_min = box2[0] - box2[2]/2.0 y2_min = box2[1] - box2[3]/2.0 x2_max = box2[0] + box2[2]/2.0 y2_max = box2[1] + box2[3]/2.0 x_min = np.maximum(x1_min, x2_min) y_min = np.maximum(y1_min, y2_min) x_max = np.minimum(x1_max, x2_max) y_max = np.minimum(y1_max, y2_max) w = np.maximum(x_max - x_min, 0.0) h = np.maximum(y_max - y_min, 0.0) intersection = w * h # 交集面积 # 计算并集面积 s1 = box1[2] * box1[3] s2 = box2[2] * box2[3] union = s1 + s2 - intersection # 并集面积 # 计算交并比 iou = intersection / union return iou def get_ignore_label(infer, gtbox, num_classes, anchor_size, anchor_mask, ignore_threshold, downsample_ratio): """ 功能: 计算大于阈值的物体标签，设置为-1，不计算损失值 输入: infer - 特征图像 gtbox - 真实边框 num_classes - 类别数量 anchor_size - 锚框大小 anchor_mask - 锚框掩码 ignore_threshold - 忽略阈值 downsample_ratio - 下采样率 输出: lbobj - 物体标签 """ # 调整特征形状 batch_size = infer.shape[0] # 特征批数 num_rows = infer.shape[2] # 特征行数 num_cols = infer.shape[3] # 特征列数 num_anchor = len(anchor_mask) # 锚框数量 infer = infer.numpy() infer = infer.reshape([-1, num_anchor, 5 + num_classes, num_rows, num_cols]) # 转换特征形状 # 计算预测边框 pdloc = infer[:, :, 0:4, :, :] # 获取预测位置 pdbox = np.zeros(pdloc.shape) # 预测边框数组 image_h = num_rows * downsample_ratio # 预测图像高度 image_w = num_cols * downsample_ratio # 预测图像宽度 for m in range(batch_size): # 遍历图像 for i in range(num_rows): # 遍历行数 for j in range(num_cols): # 遍历列数 for k in range(num_anchor): # 遍历锚框 # 获取边框大小 anchor_w = anchor_size[2 * anchor_mask[k]] # 锚框宽度 anchor_h = anchor_size[2 * anchor_mask[k] + 1] # 锚框高度 # 设置预测边框 pdbox[m, k, 0, i, j] = j # 预测边框cx pdbox[m, k, 1, i, j] = i # 预测边框cy pdbox[m, k, 2, i, j] = anchor_w # 预测边框pw pdbox[m, k, 3, i, j] = anchor_h # 预测边框ph pdbox[:, :, 0, :, :] = (pdbox[:, :, 0, :, :] + sigmoid(pdloc[:, :, 0, :, :])) / num_cols # 预测边框x=cx + dx pdbox[:, :, 1, :, :] = (pdbox[:, :, 1, :, :] + sigmoid(pdloc[:, :, 1, :, :])) / num_rows # 预测边框y=cy + dy pdbox[:, :, 2, :, :] = (pdbox[:, :, 2, :, :] * np.exp(pdloc[:, :, 2, :, :])) / image_w # 预测边框w=pw * exp(tw) pdbox[:, :, 3, :, :] = (pdbox[:, :, 3, :, :] * np.exp(pdloc[:, :, 3, :, :])) / image_h # 预测边框h=ph * exp(th) pdbox = np.clip(pdbox, 0.0, 1.0) # 限制预测边框范围为[0,1] # 计算物体标签 lbobj = np.zeros([batch_size, num_anchor, num_rows, num_cols]) # 物体标签 for m in range(batch_size): # 遍历图像 for n in range(len(gtbox[m])): # 遍历真实边框 # 获取真实边框 gtbox_x = gtbox[m][n][0] # 真实边框gtx gtbox_y = gtbox[m][n][1] # 真实边框gty gtbox_w = gtbox[m][n][2] # 真实边框gtw gtbox_h = gtbox[m][n][3] # 真实边框gth # 是否存在物体 if gtbox_w < 1e-3 or gtbox_h < 1e-3: continue # 获取预测边框 pdbox_x = pdbox[m, :, 0, :, :] # 预测边框pdx pdbox_y = pdbox[m, :, 1, :, :] # 预测边框pdy pdbox_w = pdbox[m, :, 2, :, :] # 预测边框pdw pdbox_h = pdbox[m, :, 3, :, :] # 预测边框pdh # 计算交并比值 box1 = [pdbox_x, pdbox_y, pdbox_w, pdbox_h] # 设置预测边框 box2 = [gtbox_x, gtbox_y, gtbox_w, gtbox_h] # 设置真实边框 ious = get_box_iou_xywh(box1, box2) # 计算交并比值 # 计算物体标签 index = np.where(ious > ignore_threshold) # 大于阈值标签索引 lbobj[m][index] = -1 # 大于阈值物体标签 return lbobj def get_predict_label(infer, gtbox, gtcls, num_classes, anchor_size, anchor_mask, ignore_threshold, downsample_ratio): """ 功能: 计算预测标签 输入: infer - 特征图像 gtbox - 真实边框 gtcls - 真实类别 num_classes - 类别数量 anchor_size - 锚框大小 anchor_mask - 锚框掩码 ignore_threshold - 忽略阈值 downsample_ratio - 下采样率 输出: lbloc - 位置标签 lbobj - 物体标签 lbcls - 分类标签 wtloc - 位置权重 """ # 设置标签数据 batch_size = infer.shape[0] # 特征批数 num_rows = infer.shape[2] # 特征行数 num_cols = infer.shape[3] # 特征列数 num_anchor = len(anchor_mask) # 锚框数量 lbloc = np.zeros([batch_size, num_anchor, 4, num_rows, num_cols]) # 位置标签 lbcls = np.zeros([batch_size, num_anchor, num_classes, num_rows, num_cols]) # 类别标签 wtloc = np.ones([batch_size, num_anchor, num_rows, num_rows]) # 位置权重 # 大于阈值物体 # lbobj = np.zeros([batch_size, num_anchor, num_rows, num_cols]) # 物体标签 lbobj = get_ignore_label(infer, gtbox, num_classes, anchor_size, anchor_mask, ignore_threshold, downsample_ratio) # 计算预测标签 image_h = num_rows * downsample_ratio # 原图高度 image_w = num_cols * downsample_ratio # 原图宽度 for m in range(batch_size): # 遍历图像 for n in range(len(gtbox[m])): # 遍历真实边框 # 获取边框坐标 gtbox_x = gtbox[m][n][0] # 真实边框gtx gtbox_y = gtbox[m][n][1] # 真实边框gty gtbox_w = gtbox[m][n][2] # 真实边框gtw gtbox_h = gtbox[m][n][3] # 真实边框gth # 是否存在物体 if gtbox_w < 1e-3 or gtbox_h < 1e-3: continue # 计算交并比值 iou_list = [] # 交并比值列表 for k in range(num_anchor): # 遍历锚框 anchor_w = anchor_size[2 * anchor_mask[k]] # 锚框宽度 anchor_h = anchor_size[2 * anchor_mask[k] + 1] # 锚框高度 box1 = [0.0, 0.0, anchor_w/float(image_w), anchor_h/float(image_h)] # 设置锚框 box2 = [0.0, 0.0, float(gtbox_w), float(gtbox_h)] # 真实边框 iou = get_box_iou_xywh(box1, box2) # 计算交并比值 iou_list.append(iou) # 添加交并比值 # 获取锚框序号 iou_list = np.array(iou_list) # 转换数据类型 iou_sort = np.argsort(iou_list) # 交并比值排序 k = iou_sort[-1] # 最大锚框序号 # 设置标签坐标 i = int(gtbox_y * num_rows) # 特征图行坐标 j = int(gtbox_x * num_cols) # 特征图列坐标 # 设置位置标签 lbloc[m, k, 0, i, j] = gtbox_x * num_cols - j # 位置标签dx=sigmoid(tx)=gtx-cx lbloc[m, k, 1, i, j] = gtbox_y * num_rows - i # 位置标签dy=sigmoid(ty)=gty-cy lbloc[m, k, 2, i, j] = np.log(gtbox_w * image_w / anchor_size[2 * anchor_mask[k]]) # 位置标签tw=log(gtw/pw) lbloc[m, k, 3, i, j] = np.log(gtbox_h * image_h / anchor_size[2 * anchor_mask[k] + 1]) # 位置标签th=log(gth/ph) lbloc = lbloc.astype('float32') # 设置物体标签 lbobj[m, k, i, j] = 1 lbobj = lbobj.astype('float32') # 设置类别标签 c = gtcls[m][n] # 标签位置 lbcls[m, k, c, i, j] = 1.0 lbcls = lbcls.astype('float32') # 设置位置权重 wtloc[m, k, i, j] = 2.0 - gtbox_w * gtbox_h # 调节不同尺寸锚框对损失函数的贡献，作为加权系数和位置损失函数相乘 wtloc = wtloc.astype('float32') return lbloc, lbobj, lbcls, wtloc def get_loss(infer, gtbox, gtcls, num_classes, anchor_size, anchor_mask, ignore_threshold, downsample_ratio): """ 功能: 计算每张图像的损失总和 输入: infer - 特征图像 gtbox - 真实边框 gtcls - 真实类别 num_classes - 类别数量 anchor_size - 锚框大小 anchor_mask - 锚框掩码 ignore_threshold - 忽略阈值 downsample_ratio - 下采样率 输出: sum_loss - 损失总和 """ # 计算预测标签 lbloc, lbobj, lbcls, wtloc = get_predict_label(infer, gtbox, gtcls, num_classes, anchor_size, anchor_mask, ignore_threshold, downsample_ratio) # 转换标签格式 lbloc = to_variable(lbloc) lbobj = to_variable(lbobj) lbcls = to_variable(lbcls) wtloc = to_variable(wtloc) lbloc.stop_gradient=True # 停止梯度计算 lbobj.stop_gradient=True # 停止梯度计算 lbcls.stop_gradient=True # 停止梯度计算 wtloc.stop_gradient=True # 停止梯度计算 # 转换特征格式 infer = fluid.layers.reshape(infer, [-1, len(anchor_mask), 5 + num_classes, infer.shape[2], infer.shape[3]]) # 正样本值位置 ploss = lbobj > 0 # 正样本值位置 ploss = fluid.layers.cast(ploss, 'float32') # 转换数据格式 ploss.stop_gradient=True # 停止梯度计算 # 计算位置损失 pdloc_dx = infer[:, :, 0, :, :] # 预测位置dx=sigmoid(tx) pdloc_dy = infer[:, :, 1, :, :] # 预测位置dy=sigmoid(ty) pdloc_tw = infer[:, :, 2, :, :] # 预测位置tw pdloc_th = infer[:, :, 3, :, :] # 预测位置th lbloc_dx = lbloc[:, :, 0, :, :] # 标签位置dx=sigmoid(tx) lbloc_dy = lbloc[:, :, 1, :, :] # 标签位置dy=sigmoid(ty) lbloc_tw = lbloc[:, :, 2, :, :] # 标签位置tw lbloc_th = lbloc[:, :, 3, :, :] # 标签位置th loss_loc_dx = fluid.layers.sigmoid_cross_entropy_with_logits(pdloc_dx, lbloc_dx) # 计算位置损失dx loss_loc_dy = fluid.layers.sigmoid_cross_entropy_with_logits(pdloc_dy, lbloc_dy) # 计算位置损失dy loss_loc_tw = fluid.layers.abs(pdloc_tw - lbloc_tw) # 计算位置损失tw loss_loc_th = fluid.layers.abs(pdloc_th - lbloc_th) # 计算位置损失th loss_loc = loss_loc_dx + loss_loc_dy + loss_loc_tw + loss_loc_th # 计算总的位置损失 loss_loc = loss_loc * wtloc # 带权重的位置损失 loss_loc = loss_loc * ploss # 正样本的位置损失 # 计算物体损失 pdobj = infer[:, :, 4, :, :] # 物体预测数值 loss_obj = fluid.layers.sigmoid_cross_entropy_with_logits(pdobj, lbobj, ignore_index=-1) # 忽略标签为-1梯度 # 计算类别损失 pdcls = infer[:, :, 5:5+num_classes, :, :] # 类别预测数值 loss_cls = fluid.layers.sigmoid_cross_entropy_with_logits(pdcls, lbcls) # 计算类别损失 loss_cls = fluid.layers.reduce_sum(loss_cls, dim=2) # 对通道维损失求和 loss_cls = loss_cls * ploss # 正样本的类别损失 # 计算平均损失 sum_loss = loss_loc + loss_obj + loss_cls # 计算损失总和 sum_loss = fluid.layers.reduce_sum(sum_loss, dim=[1, 2, 3]) # 每张图像损失 return sum_loss def get_sum_loss(infer, gtbox, gtcls, num_classes, anchor_size, anchor_mask, ignore_threshold, downsample_ratio): """ 功能: 计算三个输出的损失总和 输入: infer - 特征列表 gtbox - 真实边框 gtcls - 真实类别 num_classes - 类别数量 anchor_size - 锚框大小 anchor_mask - 锚框掩码 ignore_threshold - 样本阈值 downsample_ratio - 下采样率 输出: sum_loss - 平均损失总和 """ # 计算平均损失 loss_list = [] # 平均损失列表 for i in range(len(infer)): # 计算平均损失 loss = get_loss(infer[i], gtbox, gtcls, num_classes, anchor_size, anchor_mask[i], ignore_threshold, downsample_ratio) loss_list.append(fluid.layers.reduce_mean(loss)) # 添加损失列表 # 减小下采样率 downsample_ratio //= 2 # 减小下采样率 # 计算损失总和 sum_loss = sum(loss_list) return sum_loss # def get_sum_loss(infer, gtbox, gtcls, num_classes, anchor_size, anchor_mask, ignore_threshold, downsample_ratio): # # 计算平均损失 # loss_list = [] # 平均损失列表 # gtbox = to_variable(gtbox) # gtcls = to_variable(gtcls) # for i in range(len(infer)): # # 计算平均损失 # loss = fluid.layers.yolov3_loss( # x=infer[i], # gt_box=gtbox, # gt_label=gtcls, # class_num=num_classes, # anchors=anchor_size, # anchor_mask=anchor_mask[i], # ignore_thresh=ignore_threshold, # downsample_ratio=downsample_ratio, # use_label_smooth=False) # loss_list.append(fluid.layers.reduce_mean(loss)) # 添加损失列表 # # 减小下采样率 # downsample_ratio //= 2 # 减小下采样率 # # 计算损失总和 # sum_loss = sum(loss_list) # return sum_loss
参考资料：
https://www.jianshu.com/p/47172eb86b39
https://www.cnblogs.com/houjun/p/10922352.html
https://blog.csdn.net/litt1e/article/details/88814417
https://blog.csdn.net/litt1e/article/details/88852745
https://blog.csdn.net/litt1e/article/details/88907542
https://aistudio.baidu.com/aistudio/projectdetail/742781
https://aistudio.baidu.com/aistudio/projectdetail/672017
https://aistudio.baidu.com/aistudio/projectdetail/868589
https://aistudio.baidu.com/aistudio/projectdetail/122277