1利用xlsxwriter自动生成excel报表，在unbuntu下公式不会自动运算，在windows下可以,主要功能代码：
workbook = xlsxwriter.Workbook('new_table.xlsx') sheetslit = workbook.add_worksheet('slit') sheetnormal = workbook.add_worksheet('normal') bold = workbook.add_format({'bold': 1}) headings = ['图片名称','漏检个数', '总数', '遮挡1','遮挡2'] sheetslit.write_row('A1', headings, bold)# sheetnormal.write_row('A1', headings, bold)# slit_count = 1 slit_miss_count = 0 slit_total_count = 0 normal_count = 1 normal_miss_count = 0 normal_total_count = 0 slit_set =set() normal_set = set() for p in path: #根据数据自动生成excel try: image = Image.open('/home/xd/Desktop/test_images/'+p) except: print('Open Error! Try again!') continue else: r_image,boxes = yolo.detect_image(image) for d in x: if d['name'] == p: miss = d['num'] - boxes if d['table_incline']==True: worksheet = sheetslit slit_count += 1 slit_miss_count += abs(miss) slit_total_count += d['num'] worksheet.write('A'+str(slit_count), d['name']) worksheet.write('B'+str(slit_count), miss) worksheet.write('C'+str(slit_count), d['num']) worksheet.write('D'+str(slit_count), d['seat_mask']) worksheet.write('E'+str(slit_count), d['stand_mask']) slit_set.add(miss) # else: worksheet = sheetnormal normal_count += 1 normal_miss_count += abs(miss) normal_total_count += d['num'] worksheet.write('A'+str(normal_count), d['name']) worksheet.write('B'+str(normal_count), miss) worksheet.write('C'+str(normal_count), d['num']) worksheet.write('D'+str(normal_count), d['seat_mask']) worksheet.write('E'+str(normal_count), d['stand_mask']) normal_set.add(miss) break sheetnormal.write('A'+str(normal_count+1),'总计') sheetnormal.write('B'+str(normal_count+1),str(normal_miss_count)) sheetnormal.write('C'+str(normal_count+1),str(normal_total_count)) count = 1 for ss in normal_set: sheetnormal.write('G'+str(count),ss) sheetnormal.write_formula('F'+str(count), '=COUNTIF(B2:B'+str(normal_count)+','+str(ss)+')') count += 1 chart_normal = workbook.add_chart({'type':'pie'}) global_miss_count = normal_miss_count + slit_miss_count chart_normal.set_title ({'name': '漏检个数'}) chart_normal.set_style(3) chart_normal.add_series({ 'name': '漏检个数', 'categories': '=normal!$G$1'+':$G$'+str(count-1), 'values': '=normal!$F$2:$F$5'+str(normal_count) }) print(sheetnormal.insert_chart('E52', chart_normal, {'x_offset': 25, 'y_offset': 10})) sheetslit.write('A'+str(slit_count+1),'总计') sheetslit.write('B'+str(slit_count+1),str(slit_miss_count)) sheetslit.write('C'+str(slit_count+1),str(slit_total_count)) print((slit_miss_count+normal_miss_count)) count = 1 for ss in slit_set: sheetslit.write_formula('F'+str(count), '=COUNTIF(B2:B'+str(slit_count)+','+str(ss)+')') sheetslit.write('G'+str(count),ss) count += 1 chart_slit = workbook.add_chart({'type':'pie'}) chart_slit.set_title ({'name': '漏检个数'}) chart_slit.set_style(3) chart_normal.add_series({ 'name': '漏检个数', 'categories': '=normal!$G$1'+':$G$'+str(count-1), 'values': '=normal!$F$2:$F$5'+str(slit_count) }) print(sheetslit.insert_chart('E52', chart_slit, {'x_offset': 25, 'y_offset': 10})) workbook.close()2使用opencv的翻转对图片进行数据增强，但对于yolo3的效果不怎么好，代码如下
def deal_flip_image(image,xmin,ymin,xmax,ymax,flip_type): #水平翻转 height = image.shape[0] width = image.shape[1] if flip_type==1: xmin = width - xmin xmax = width - xmax #垂直翻转 elif flip_type==0: ymin = height - ymin ymax = height - ymax #垂直水平翻转 else: xmin = width - xmin xmax = width - xmax ymin = height - ymin ymax = height - ymax return xmin,ymin,xmax,ymax # cv2.rectangle(image, (xmin, ymax), (xmax, ymin), (255, 0, 0), 2) # 获得根节点 def filp_augment(new_path= "D://train/transform/",new_image_path= "D://train/newimages/",old_path="D://train/Annotations/",old_image_path="D://train/Images/"): for p in os.listdir(old_path): for filp_type in [1,0,-1]: tree = ET.parse(old_path+p) root = tree.getroot() objects = root.findall("object") file_name = root.find("filename").text path = old_image_path + file_name orignal_image = cv2.imread(path) transfrom_image = cv2.flip(orignal_image, filp_type) for item in objects: bndbox = item.find('bndbox') xmin = int(bndbox.find("xmin").text) ymin = int(bndbox.find("ymin").text) xmax = int(bndbox.find("xmax").text) ymax = int(bndbox.find("ymax").text) # cv2.rectangle(orignal_image, (xmin, ymax), (xmax, ymin), (255, 0, 0), 2) xmin,ymin,xmax,ymax = deal_flip_image(transfrom_image,xmin,ymin,xmax,ymax,filp_type) bndbox.find("xmin").text = str(xmin) bndbox.find("ymin").text = str(ymin) bndbox.find("xmax").text = str(xmax) bndbox.find("ymax").text = str(ymax) prefix = {1:"horizontal_",0:"vertical_",-1:"h_v_"}[filp_type] new_file_name = prefix + file_name root.find("filename").text = new_file_name root.find("path").text = root.find("path").text.replace(file_name,new_file_name) cv2.imwrite(new_image_path + new_file_name,transfrom_image) tree.write(new_path+prefix + p,xml_declaration=True, encoding="utf-8") # print((xmin,ymin,xmax,ymax)) def verify_augment(annotations_path,image_path): image = cv2.imread(image_path) tree = ET.parse(annotations_path) root = tree.getroot() objects = root.findall("object") for item in objects: bndbox = item.find('bndbox') xmin = int(bndbox.find("xmin").text) ymin = int(bndbox.find("ymin").text) xmax = int(bndbox.find("xmax").text) ymax = int(bndbox.find("ymax").text) cv2.rectangle(image, (xmin, ymax), (xmax, ymin), (255, 0, 0), 2) cv2.imshow("test", image)3对github的基于keras的yolov3模型的使用，keras-yolo3。使用labelImg标注数据，使用train.py进行数据训练，配置cfg文件,调整yolov3模型的参数，之后使用python convert.py -w yolov3.cfg yolov3.weights model_data/yolo_weights.h5将权重和cfg转换成.h5文件作为模型载入权重。
目前认为几个比较重要的参数是input_shape,anchors，filters,filters数目=（class+5）*(anchors个数//尺度),yolov3三个尺度,anchors数目为9。anchors可以通过kmeans来获得。
4 对tensorboard有一定认识，使用 tensorboard --logdir=路径 命令,就可以在127.0.0.1:6006这个地址上查看计算图，网络结构，损失函数。对于keras，只需要在训练的callback里添加tensorboard对象就行了。
logging = TensorBoard(log_dir=log_dir) #训练模型 model.fit_generator(data_generator_wrapper(lines[:num_train], batch_size, input_shape, anchors, num_classes), steps_per_epoch=max(1, num_train//batch_size), validation_data=data_generator_wrapper(lines[num_train:], batch_size, input_shape, anchors, num_classes), validation_steps=max(1, num_val//batch_size), epochs=100, initial_epoch=0, callbacks=[logging, checkpoint])