Keras-RetinaNet训练自己的数据详细教程

水深无声 2022-03-12 08:28 331阅读 0赞

#### 代码地址：[Github][] ####

--------------------

**准备工作：**

1、代码开源框架使用的是 [fizyr/keras-retinanet][fizyr_keras-retinanet]  
2、Keras版本要2.2.4以上

下面进入正题。

--------------------

**第一部分：模型准备**

（1）模型下载地址：[fizyr/keras-retinanet][fizyr_keras-retinanet]  
（2）模型安装可以使用以下命令：

pip install numpy --user

pip install . --user

安装过程中，会检查依赖库，比如opencv-python，如果没有安装，会加载并安装。这里提一句，如果在安装时某个包下载安装不成功，自己记下来版本，比如opencv-python 3.4.5.20，可以直接先去利用pip或conda安装，但是一定要记得对应的版本。

（3）模型编译可以使用以下命令：

python setup.py build_ext --inplace

编译的时候可能会出现提示，没有某个版本C++的编译器，我提示的时没有2014版，把错误提示直接百度，就会出现解决方法，我是下载了一个3M的14版的编译工具。（当然，最好就是有相应版本的完整C++）

**第二部分：数据准备**

（1）在`keras-retinanet-master/keras_retinanet/`文件夹下面新建一个文件夹`CSV`用来存放自己制作的数据集。

**数据文件夹格式如下：**

（2）根据官网的样例，自己制作的Annotations数据集格式如下：

path/to/image.jpg,x1,y1,x2,y2,class_name

如果一张图片中没有包含任何要检测的物体，则格式如下：

path/to/image.jpg,,,,,

一个完整的例子：

/data/imgs/img_001.jpg,837,346,981,456,cow
    /data/imgs/img_002.jpg,215,312,279,391,cat
    /data/imgs/img_002.jpg,22,5,89,84,bird
    /data/imgs/img_003.jpg,,,,,

下面，我就贴出自己写的一个代码：

def restrict_image_info(label_path):
    
        with open(label_path, 'r') as load_f:
            load_dict = json.load(load_f)
            image_collect = load_dict['images']
            image_num = len(image_collect)
            anno_collect = load_dict['annotations']
            anno_num = len(anno_collect)
    
            img_path_list = []
            x1_list = []
            y1_list = []
            x2_list = []
            y2_list = []
            category_list = []
    
            img_path_val_list = []
            x1_val_list = []
            y1_val_list = []
            x2_val_list = []
            y2_val_list = []
            category_val_list = []
    
            mapper = {0: 'tieke', 1: 'heiding',
                      2: 'daoju', 3: 'dian', 4: 'jiandao'}
    
            train_rate = 0.9
            hight = image_num*train_rate
            train_img_id = np.random.randint(0, image_num, size=int(hight))
            print(len(train_img_id))
    
            for i in range(image_num):
                img = image_collect[i]
                img_name = img['file_name']
                img_id = img['id']
                img_height = img['height']
                img_width = img['width']
                if i in train_img_id:
                    for j in range(anno_num):
                        if anno_collect[j]['image_id'] == img_id:
                            bbox = anno_collect[j]['bbox']
                            img_path_list.append(restrict_rele_path+img_name)
                            x1_list.append(int(np.rint(bbox[0])))
                            y1_list.append(int(np.rint(bbox[1])))
                            x2_list.append(
                                int(np.rint(bbox[0] + bbox[2])))
                            y2_list.append(
                                int(np.rint((bbox[1]+bbox[3]))))
                            category_list.append(anno_collect[j]['category_id']-1)
    
                    anno = pd.DataFrame()
                    anno['img_path'] = img_path_list
                    anno['x1'] = x1_list
                    anno['y1'] = y1_list
                    anno['x2'] = x2_list
                    anno['y2'] = y2_list
                    anno['class'] = category_list
                    anno['class'] = anno['class'].map(mapper)
                else:
                    for j in range(anno_num):
                        if anno_collect[j]['image_id'] == img_id:
                            bbox = anno_collect[j]['bbox']
                            img_path_val_list.append(restrict_rele_path+img_name)
                            x1_val_list.append(int(np.rint(bbox[0])))
                            y1_val_list.append(int(np.rint(bbox[1])))
                            x2_val_list.append(
                                int(np.rint(bbox[0] + bbox[2])))
                            y2_val_list.append(
                                int(np.rint((bbox[1]+bbox[3]))))
                            category_val_list.append(
                                anno_collect[j]['category_id']-1)
    
                    anno_val = pd.DataFrame()
                    anno_val['img_path'] = img_path_val_list
                    anno_val['x1'] = x1_val_list
                    anno_val['y1'] = y1_val_list
                    anno_val['x2'] = x2_val_list
                    anno_val['y2'] = y2_val_list
                    anno_val['class'] = category_val_list
                    anno_val['class'] = anno_val['class'].map(mapper)
    
            anno.to_csv('CSV/train_annotations.csv', index=None, header=None)
            anno_val.to_csv('CSV/val_annotations.csv', index=None, header=None)

训练图片生成的数据格式如下：

data/jinnan2_round1_train_20190305/restricted/190119_184244_00166940.jpg,88,253,206,295,daoju
    data/jinnan2_round1_train_20190305/restricted/190119_184244_00166940.jpg,296,244,414,344,jiandao
    data/jinnan2_round1_train_20190305/restricted/190119_184244_00166940.jpg,231,239,299,341,jiandao
    data/jinnan2_round1_train_20190305/restricted/190119_184244_00166940.jpg,99,278,194,320,dian

验证图片生成的数据格式如下：

data/jinnan2_round1_train_20190305/restricted/190119_182957_00166754.jpg,314,237,326,265,dian
    data/jinnan2_round1_train_20190305/restricted/190127_100838_00177153.jpg,246,229,304,279,tieke
    data/jinnan2_round1_train_20190305/restricted/190119_184522_00166980.jpg,668,409,717,432,dian
    data/jinnan2_round1_train_20190305/restricted/190119_183142_00166782.jpg,565,326,708,432,jiandao
    data/jinnan2_round1_train_20190305/restricted/190127_143529_00178527.jpg,8,262,45,326,heiding

（3）根据官网的样例，自己制作的classes数据集格式如下：

class_name,id

一个完整的例子：

cow,0
    cat,1
    bird,2

最后生成的数据格式如下：

tieke,0
    heiding,1
    daoju,2
    dian,3
    jiandao,4

**注意：保存的csv文件是没有头部行的，不然后续代码会报错！**

（4）检查生成的数据是否合格

要进行这一步，必须先要完成第一步中模型的下载与编译！

检查数据可以使用以下命令：

python keras_retinanet/bin/debug.py csv keras_retinanet/CSV/train_annotations.csv keras_retinanet/CSV/classes.csv

其中第一个参数`csv`代表要检查的数据是自己制作的数据集，第二个参数是`train_annotations.csv`对应的路径，第三个参数是`classes.csv`对应的路径。

（5）图片存放位置

这个可以根据自己的需要定，但是最好放在上面新建的`CSV`文件夹下面，这个使用路径比较方便。在我自己这个代码中，我是在`CSV`文件夹下新建一个`data`文件夹下存放自己的图片，此时注意与`train_annotations.csv`文件中的图片路径要一致，比如我这时候就应该是这样：

data/jinnan2_round1_train_20190222/restricted/190119_185206_00167075.jpg,125,279,177,339,tieke
    data/jinnan2_round1_train_20190222/restricted/190119_185206_00167075.jpg,153,363,238,549,daoju

（6）关于模型的图片输入尺寸

在`https://github.com/fizyr/keras-retinanet/blob/master/keras_retinanet/bin/train.py`中的409、410行有设置输入的默认参数（800\*1333）：

parser.add_argument('--image-min-side',   help='Rescale the image so the smallest side is min_side.', type=int, default=800)
    parser.add_argument('--image-max-side',   help='Rescale the image if the largest side is larger than max_side.', type=int, default=1333)

**第三部分：模型训练**

模型训练可以使用以下命令：

python keras_retinanet/bin/train.py csv keras_retinanet/CSV/train_annotations.csv keras_retinanet/CSV/classes.csv --val-annotations keras_retinanet/CSV/val_annotations.csv

其中第一个参数`csv`代表要检查的数据是自己制作的数据集，第二个参数是`train_annotations.csv`对应的路径，第三个参数是`classes.csv`对应的路径，第四个参数`--val-annotations`是`val_annotations.csv`对应的路径。

多卡训练可用如下命令：

python keras_retinanet/bin/train.py --multi-gpu-force --multi-gpu 2 --batch-size 2 csv keras_retinanet/CSV/train_annotations.csv keras_retinanet/CSV/classes.csv --val-annotations keras_retinanet/CSV/val_annotations.csv

替换backbone可用如下命令（可选的包括vgg16，vgg19，resnet50，resnet101，densenet121，densenet169，densenet201）：

python keras_retinanet/bin/train.py --steps 1000 --backbone vgg16 --gpu 2 csv keras_retinanet/CSV/train_annotations.csv keras_retinanet/CSV/classes.csv --val-annotations keras_retinanet/CSV/val_annotations.csv

**第四部分：模型测试**

#!/usr/bin/env python
    # coding=UTF-8
    '''
    @Description: 
    @Author: HuangQinJian
    @LastEditors: HuangQinJian
    @Date: 2019-03-07 08:59:34
    @LastEditTime: 2019-03-07 11:13:20
    '''
    
    import os
    import time
    
    import keras
    # import miscellaneous modules
    import matplotlib.pyplot as plt
    import numpy as np
    import pandas as pd
    # set tf backend to allow memory to grow, instead of claiming everything
    import tensorflow as tf
    from skimage.io import imsave
    
    import cv2
    from keras_retinanet import models
    from keras_retinanet.utils.colors import label_color
    from keras_retinanet.utils.image import (preprocess_image, read_image_bgr,
                                             resize_image)
    from keras_retinanet.utils.visualization import draw_box, draw_caption
    
    if not os.path.exists('result'):
        os.mkdir('result')
    
    
    def get_session():
        config = tf.ConfigProto()
        config.gpu_options.allow_growth = True
        return tf.Session(config=config)
    
    
    os.environ["CUDA_VISIBLE_DEVICES"] = "1"
    
    
    def predict_save(model, test_img_fold, test_img_list):
        # load image
        img_name_list = []
        bboxes_list = []
        class_list = []
        score_list = []
        for i in range(len(test_img_list)):
            # for i in range(1):
            img_name = test_img_list[i]
            img_path = os.path.join(test_img_fold, img_name)
            image = read_image_bgr(img_path)
            # copy to draw on
            draw = image.copy()
            draw = cv2.cvtColor(draw, cv2.COLOR_BGR2RGB)
            # preprocess image for network
            image = preprocess_image(image)
            image, scale = resize_image(image)
            # process image
            start = time.time()
            # print(image.shape)
            # print(scale)
            boxes, scores, labels = model.predict_on_batch(
                np.expand_dims(image, axis=0))
            print("processing time: ", time.time() - start)
            # correct for image scale
            boxes /= scale
            i = 0
            for box, score, label in zip(boxes[0], scores[0], labels[0]):
                # scores are sorted so we can break
                if score < 0.5:
                    break
                color = label_color(label)
                b = box.astype(int)
                img_name_list.append(img_name)
                bboxes_list.append(b)
                class_list.append(labels[0][i])
                score_list.append(score)
                i += 1
                draw_box(draw, b, color=color)
                caption = "{} {:.3f}".format(labels_to_names[label], score)
                draw_caption(draw, b, caption)
            imsave('result/'+img_name, draw)
        submit = pd.DataFrame()
        submit['img_name'] = img_name_list
        submit['bbox'] = bboxes_list
        submit['class'] = class_list
        submit['score'] = score_list
        # submit.to_csv('submit.csv', index=None)
        submit.to_pickle('submit.pkl')
    
    
    if __name__ == "__main__":
        # set the modified tf session as backend in keras
        keras.backend.tensorflow_backend.set_session(get_session())
    
        # adjust this to point to your downloaded/trained model
        # models can be downloaded here: https://github.com/fizyr/keras-retinanet/releases
    
        model_path = os.path.join('snapshots', 'old.h5')
    
        # load retinanet model
        model = models.load_model(model_path, backbone_name='resnet50')
    
        # if the model is not converted to an inference model, use the line below
        # see: https://github.com/fizyr/keras-retinanet#converting-a-training-model-to-inference-model
        model = models.convert_model(model)
    
        # print(model.summary())
    
        # load label to names mapping for visualization purposes
        labels_to_names = {0: 'tieke', 1: 'heiding',
                           2: 'daoju', 3: 'dian', 4: 'jiandao'}
    
        test_img_fold = 'keras_retinanet/CSV/data/jinnan2_round1_test_a_20190306/'
        test_img_list = os.listdir(test_img_fold)
        print(len(test_img_list))
        predict_save(model, test_img_fold, test_img_list)

**可能会遇到的错误：**

（1）`ImportError: No module named 'keras_resnet'`

解决办法：`pip install keras-resnet --user`

（2）在第四部分模型预测的时候，必须运行：

# if the model is not converted to an inference model, use the line below
     # see: https://github.com/fizyr/keras-retinanet#converting-a-training-model-to-inference-model
     model = models.convert_model(model)

否则会报如下错误：

'boxes, scores, labels ' not enough values to unpack (expected 3, got 2)

#### 代码地址：[Github][] ####

--------------------

**参考资料：**  
1、[Retinanet训练自己的数据（2）：模型准备][Retinanet_2]

[Github]: https://github.com/HuangQinJian/2019_jinnan2_x-ray-detection/tree/master
[fizyr_keras-retinanet]: https://github.com/fizyr/keras-retinanet
[Retinanet_2]: https://blog.csdn.net/mingyang_wang/article/details/85955425