train.py

import keras.backend as K
import numpy as np
from keras.callbacks import (EarlyStopping, ModelCheckpoint, ReduceLROnPlateau,
                             TensorBoard)
from keras.layers import Input, Lambda
from keras.models import Model
from keras.optimizers import Adam

from nets.loss import yolo_loss
from nets.yolo4 import yolo_body
from utils.utils import (LossHistory, WarmUpCosineDecayScheduler,
                         get_random_data, get_random_data_with_Mosaic)


#---------------------------------------------------#
#   获得类和先验框
#---------------------------------------------------#
def get_classes(classes_path):
    '''loads the classes'''
    with open(classes_path) as f:
        class_names = f.readlines()
    class_names = [c.strip() for c in class_names]
    return class_names

def get_anchors(anchors_path):
    '''loads the anchors from a file'''
    with open(anchors_path) as f:
        anchors = f.readline()
    anchors = [float(x) for x in anchors.split(',')]
    return np.array(anchors).reshape(-1, 2)

#---------------------------------------------------#
#   训练数据生成器
#---------------------------------------------------#
def data_generator(annotation_lines, batch_size, input_shape, anchors, num_classes, mosaic=False, random=True):
    n = len(annotation_lines)
    i = 0
    flag = True
    while True:
        image_data = []
        box_data = []
        for b in range(batch_size):
            if i==0:
                np.random.shuffle(annotation_lines)
            if mosaic:
                if flag and (i+4) < n:
                    image, box = get_random_data_with_Mosaic(annotation_lines[i:i+4], input_shape)
                    i = (i+4) % n
                else:
                    image, box = get_random_data(annotation_lines[i], input_shape, random=random)
                    i = (i+1) % n
                flag = bool(1-flag)
            else:
                image, box = get_random_data(annotation_lines[i], input_shape, random=random)
                i = (i+1) % n
            image_data.append(image)
            box_data.append(box)
        image_data = np.array(image_data)
        box_data = np.array(box_data)
        y_true = preprocess_true_boxes(box_data, input_shape, anchors, num_classes)
        yield [image_data, *y_true], np.zeros(batch_size)

#---------------------------------------------------#
#   读入xml文件，并输出y_true
#---------------------------------------------------#
def preprocess_true_boxes(true_boxes, input_shape, anchors, num_classes):
    assert (true_boxes[..., 4]<num_classes).all(), 'class id must be less than num_classes'
    # 一共有三个特征层数
    num_layers = len(anchors)//3
    #-----------------------------------------------------------#
    #   13x13的特征层对应的anchor是[142, 110], [192, 243], [459, 401]
    #   26x26的特征层对应的anchor是[36, 75], [76, 55], [72, 146]
    #   52x52的特征层对应的anchor是[12, 16], [19, 36], [40, 28]
    #-----------------------------------------------------------#
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]]

    #-----------------------------------------------------------#
    #   获得框的坐标和图片的大小
    #-----------------------------------------------------------#
    true_boxes = np.array(true_boxes, dtype='float32')
    input_shape = np.array(input_shape, dtype='int32')
    #-----------------------------------------------------------#
    #   通过计算获得真实框的中心和宽高
    #   中心点(m,n,2) 宽高(m,n,2)
    #-----------------------------------------------------------#
    boxes_xy = (true_boxes[..., 0:2] + true_boxes[..., 2:4]) // 2
    boxes_wh = true_boxes[..., 2:4] - true_boxes[..., 0:2]
    #-----------------------------------------------------------#
    #   将真实框归一化到小数形式
    #-----------------------------------------------------------#
    true_boxes[..., 0:2] = boxes_xy/input_shape[::-1]
    true_boxes[..., 2:4] = boxes_wh/input_shape[::-1]

    # m为图片数量，grid_shapes为网格的shape
    m = true_boxes.shape[0]
    grid_shapes = [input_shape//{0:32, 1:16, 2:8}[l] for l in range(num_layers)]
    #-----------------------------------------------------------#
    #   y_true的格式为(m,13,13,3,85)(m,26,26,3,85)(m,52,52,3,85)
    #-----------------------------------------------------------#
    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)]

    #-----------------------------------------------------------#
    #   [9,2] -> [1,9,2]
    #-----------------------------------------------------------#
    anchors = np.expand_dims(anchors, 0)
    anchor_maxes = anchors / 2.
    anchor_mins = -anchor_maxes

    #-----------------------------------------------------------#
    #   长宽要大于0才有效
    #-----------------------------------------------------------#
    valid_mask = boxes_wh[..., 0]>0

    for b in range(m):
        # 对每一张图进行处理
        wh = boxes_wh[b, valid_mask[b]]
        if len(wh)==0: continue
        #-----------------------------------------------------------#
        #   [n,2] -> [n,1,2]
        #-----------------------------------------------------------#
        wh = np.expand_dims(wh, -2)
        box_maxes = wh / 2.
        box_mins = -box_maxes

        #-----------------------------------------------------------#
        #   计算所有真实框和先验框的交并比
        #   intersect_area  [n,9]
        #   box_area        [n,1]
        #   anchor_area     [1,9]
        #   iou             [n,9]
        #-----------------------------------------------------------#
        intersect_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)
        #-----------------------------------------------------------#
        #   维度是[n,] 感谢 消尽不死鸟 的提醒
        #-----------------------------------------------------------#
        best_anchor = np.argmax(iou, axis=-1)

        for t, n in enumerate(best_anchor):
            #-----------------------------------------------------------#
            #   找到每个真实框所属的特征层
            #-----------------------------------------------------------#
            for l in range(num_layers):
                if n in anchor_mask[l]:
                    #-----------------------------------------------------------#
                    #   floor用于向下取整，找到真实框所属的特征层对应的x、y轴坐标
                    #-----------------------------------------------------------#
                    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指的的当前这个特征点的第k个先验框
                    #-----------------------------------------------------------#
                    k = anchor_mask[l].index(n)
                    #-----------------------------------------------------------#
                    #   c指的是当前这个真实框的种类
                    #-----------------------------------------------------------#
                    c = true_boxes[b, t, 4].astype('int32')
                    #-----------------------------------------------------------#
                    #   y_true的shape为(m,13,13,3,85)(m,26,26,3,85)(m,52,52,3,85)
                    #   最后的85可以拆分成4+1+80，4代表的是框的中心与宽高、
                    #   1代表的是置信度、80代表的是种类
                    #-----------------------------------------------------------#
                    y_true[l][b, j, i, k, 0:4] = true_boxes[b, t, 0:4]
                    y_true[l][b, j, i, k, 4] = 1
                    y_true[l][b, j, i, k, 5+c] = 1

    return y_true

#----------------------------------------------------#
#   检测精度mAP和pr曲线计算参考视频
#   https://www.bilibili.com/video/BV1zE411u7Vw
#----------------------------------------------------#
if __name__ == "__main__":
    #----------------------------------------------------#
    #   获得图片路径和标签
    #----------------------------------------------------#
    annotation_path = '2007_train.txt'
    #------------------------------------------------------#
    #   训练后的模型保存的位置，保存在logs文件夹里面
    #------------------------------------------------------#
    log_dir = 'logs/'
    #----------------------------------------------------#
    #   classes和anchor的路径，非常重要
    #   训练前一定要修改classes_path，使其对应自己的数据集
    #----------------------------------------------------#
    classes_path = 'model_data/new_classes.txt'    
    anchors_path = 'model_data/yolo_anchors.txt'
    #------------------------------------------------------#
    #   权值文件请看README，百度网盘下载
    #   训练自己的数据集时提示维度不匹配正常
    #   预测的东西都不一样了自然维度不匹配
    #------------------------------------------------------#
    weights_path = 'model_data/yolo4_weight.h5'
    #------------------------------------------------------#
    #   训练用图片大小
    #   一般在416x416和608x608选择
    #------------------------------------------------------#
    input_shape = (416,416)
    #------------------------------------------------------#
    #   是否对损失进行归一化，用于改变loss的大小
    #   用于决定计算最终loss是除上batch_size还是除上正样本数量
    #------------------------------------------------------#
    normalize = False

    #----------------------------------------------------#
    #   获取classes和anchor
    #----------------------------------------------------#
    class_names = get_classes(classes_path)
    anchors     = get_anchors(anchors_path)
    #------------------------------------------------------#
    #   一共有多少类和多少先验框
    #------------------------------------------------------#
    num_classes = len(class_names)
    num_anchors = len(anchors)
    #------------------------------------------------------#
    #   Yolov4的tricks应用
    #   mosaic 马赛克数据增强 True or False 
    #   实际测试时mosaic数据增强并不稳定，所以默认为False
    #   Cosine_scheduler 余弦退火学习率 True or False
    #   label_smoothing 标签平滑 0.01以下一般 如0.01、0.005
    #------------------------------------------------------#
    mosaic = True #马赛克增强
    Cosine_scheduler = False #余弦学习率衰减
    label_smoothing = 0
    #0.001 if you need

    K.clear_session()
    #------------------------------------------------------#
    #   创建yolo模型
    #------------------------------------------------------#
    image_input = Input(shape=(None, None, 3))
    h, w = input_shape
    print('Create YOLOv4 model with {} anchors and {} classes.'.format(num_anchors, num_classes))
    model_body = yolo_body(image_input, num_anchors//3, num_classes)
    
    #------------------------------------------------------#
    #   载入预训练权重
    #------------------------------------------------------#
    print('Load weights {}.'.format(weights_path))
    model_body.load_weights(weights_path, by_name=True, skip_mismatch=True)
    
    #------------------------------------------------------#
    #   在这个地方设置损失，将网络的输出结果传入loss函数
    #   把整个模型的输出作为loss
    #------------------------------------------------------#
    y_true = [Input(shape=(h//{0:32, 1:16, 2:8}[l], w//{0:32, 1:16, 2:8}[l], \
        num_anchors//3, num_classes+5)) for l in range(3)]
    loss_input = [*model_body.output, *y_true]
    model_loss = Lambda(yolo_loss, output_shape=(1,), name='yolo_loss',
        arguments={'anchors': anchors, 'num_classes': num_classes, 'ignore_thresh': 0.5, 
            'label_smoothing': label_smoothing, 'normalize': normalize})(loss_input)

    model = Model([model_body.input, *y_true], model_loss)

    #-------------------------------------------------------------------------------#
    #   训练参数的设置
    #   logging表示tensorboard的保存地址
    #   checkpoint用于设置权值保存的细节，period用于修改多少epoch保存一次
    #   reduce_lr用于设置学习率下降的方式
    #   early_stopping用于设定早停，val_loss多次不下降自动结束训练，表示模型基本收敛
    #-------------------------------------------------------------------------------#
    logging = TensorBoard(log_dir=log_dir)
    checkpoint = ModelCheckpoint(log_dir + 'ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}.h5',
        monitor='val_loss', save_weights_only=True, save_best_only=False, period=1)
    early_stopping = EarlyStopping(monitor='val_loss', min_delta=0, patience=10, verbose=1)
    loss_history = LossHistory(log_dir)

    #----------------------------------------------------------------------#
    #   验证集的划分在train.py代码里面进行
    #   2007_test.txt和2007_val.txt里面没有内容是正常的。训练不会使用到。
    #   当前划分方式下，验证集和训练集的比例为1:9
    #----------------------------------------------------------------------#
    val_split = 0.1
    with open(annotation_path) as f:
        lines = f.readlines()
    np.random.seed(10101)
    np.random.shuffle(lines)
    np.random.seed(None)
    num_val = int(len(lines)*val_split)
    num_train = len(lines) - num_val
    
    freeze_layers = 249
    for i in range(freeze_layers): model_body.layers[i].trainable = False
    print('Freeze the first {} layers of total {} layers.'.format(freeze_layers, len(model_body.layers)))

    #------------------------------------------------------#
    #   主干特征提取网络特征通用，冻结训练可以加快训练速度
    #   也可以在训练初期防止权值被破坏。
    #   Init_Epoch为起始世代
    #   Freeze_Epoch为冻结训练的世代
    #   Epoch总训练世代
    #   提示OOM或者显存不足请调小Batch_size
    #------------------------------------------------------#
    if True:
        Init_epoch          = 0
        Freeze_epoch        = 50
        batch_size          = 8
        learning_rate_base  = 1e-3

        if Cosine_scheduler:
            # 预热期
            warmup_epoch    = int((Freeze_epoch-Init_epoch)*0.2)
            # 总共的步长
            total_steps     = int((Freeze_epoch-Init_epoch) * num_train / batch_size)
            # 预热步长
            warmup_steps    = int(warmup_epoch * num_train / batch_size)
            # 学习率
            reduce_lr       = WarmUpCosineDecayScheduler(learning_rate_base=learning_rate_base,
                                                        total_steps=total_steps,
                                                        warmup_learning_rate=1e-4,
                                                        warmup_steps=warmup_steps,
                                                        hold_base_rate_steps=num_train,
                                                        min_learn_rate=1e-6
                                                        )
            model.compile(optimizer=Adam(), loss={'yolo_loss': lambda y_true, y_pred: y_pred})
        else:
            reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=3, verbose=1)
            model.compile(optimizer=Adam(learning_rate_base), loss={'yolo_loss': lambda y_true, y_pred: y_pred})

        epoch_size          = num_train // batch_size
        epoch_size_val      = num_val // batch_size

        if epoch_size == 0 or epoch_size_val == 0:
            raise ValueError("数据集过小，无法进行训练，请扩充数据集。")

        print('Train on {} samples, val on {} samples, with batch size {}.'.format(num_train, num_val, batch_size))
        model.fit_generator(data_generator(lines[:num_train], batch_size, input_shape, anchors, num_classes, mosaic=mosaic, random=True),
                steps_per_epoch=epoch_size,
                validation_data=data_generator(lines[num_train:], batch_size, input_shape, anchors, num_classes, mosaic=False, random=False),
                validation_steps=epoch_size_val,
                epochs=Freeze_epoch,
                initial_epoch=Init_epoch,
                callbacks=[logging, checkpoint, reduce_lr, early_stopping, loss_history])
        model.save_weights(log_dir + 'trained_weights_stage_1.h5')

    for i in range(freeze_layers): model_body.layers[i].trainable = True

    if True:
        Freeze_epoch        = 50
        Epoch               = 100
        batch_size          = 2
        learning_rate_base  = 1e-4

        if Cosine_scheduler:
            # 预热期
            warmup_epoch = int((Epoch-Freeze_epoch)*0.2)
            # 总共的步长
            total_steps = int((Epoch-Freeze_epoch) * num_train / batch_size)
            # 预热步长
            warmup_steps = int(warmup_epoch * num_train / batch_size)
            # 学习率
            reduce_lr = WarmUpCosineDecayScheduler(learning_rate_base=learning_rate_base,
                                                        total_steps=total_steps,
                                                        warmup_learning_rate=1e-5,
                                                        warmup_steps=warmup_steps,
                                                        hold_base_rate_steps=num_train//2,
                                                        min_learn_rate=1e-6
                                                        )
            model.compile(optimizer=Adam(), loss={'yolo_loss': lambda y_true, y_pred: y_pred})
        else:
            reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=3, verbose=1)
            model.compile(optimizer=Adam(learning_rate_base), loss={'yolo_loss': lambda y_true, y_pred: y_pred})

        epoch_size          = num_train // batch_size
        epoch_size_val      = num_val // batch_size

        if epoch_size == 0 or epoch_size_val == 0:
            raise ValueError("数据集过小，无法进行训练，请扩充数据集。")

        print('Train on {} samples, val on {} samples, with batch size {}.'.format(num_train, num_val, batch_size))
        model.fit_generator(data_generator(lines[:num_train], batch_size, input_shape, anchors, num_classes, mosaic=mosaic, random=True),
                steps_per_epoch=epoch_size,
                validation_data=data_generator(lines[num_train:], batch_size, input_shape, anchors, num_classes, mosaic=False, random=False),
                validation_steps=epoch_size_val,
                epochs=Epoch,
                initial_epoch=Freeze_epoch,
                callbacks=[logging, checkpoint, reduce_lr, early_stopping, loss_history])
        model.save_weights(log_dir + 'last2_morepics.h5')