import keras.backend as K import numpy as np import tensorflow as tf from keras.backend.tensorflow_backend import set_session 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 (WarmUpCosineDecayScheduler, get_random_data, get_random_data_with_Mosaic, rand) #---------------------------------------------------# # 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+1) % 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] [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/voc_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 #------------------------------------------------------# normalize = True #----------------------------------------------------# # 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 # 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 Cosine_scheduler = False label_smoothing = 0 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) #----------------------------------------------------------------------# # 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}) 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=max(1, num_train//batch_size), validation_data=data_generator(lines[num_train:], batch_size, input_shape, anchors, num_classes, mosaic=False, random=False), validation_steps=max(1, num_val//batch_size), epochs=Freeze_epoch, initial_epoch=Init_epoch, callbacks=[logging, checkpoint, reduce_lr, early_stopping]) 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}) 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=max(1, num_train//batch_size), validation_data=data_generator(lines[num_train:], batch_size, input_shape, anchors, num_classes, mosaic=False, random=False), validation_steps=max(1, num_val//batch_size), epochs=Epoch, initial_epoch=Freeze_epoch, callbacks=[logging, checkpoint, reduce_lr, early_stopping]) model.save_weights(log_dir + 'last1.h5')