當前位置：首頁 > 编程资源 > 编程问答 >内容正文

编程问答

飞桨2.0 PaddleDetection：瓶装酒瑕疵检测迁移学习教程

發布時間：2023/12/31 编程问答 33 豆豆

生活随笔收集整理的這篇文章主要介紹了飞桨2.0 PaddleDetection：瓶装酒瑕疵检测迁移学习教程小編覺得挺不錯的,現在分享給大家,幫大家做個參考.

序言

瓶裝酒的生產過程中，受到原材料質量（酒瓶）以及加工工藝（灌裝）等因素的影響，產品中可能存在各類瑕疵影響產品質量。一條產線一般有三到五個質檢環節分別檢測不同類型的瑕疵。由于瑕疵種類多樣、有的瑕疵體積小不易察覺，瓶裝酒廠家往往需要投入大量人力成本用于產品質檢。高效、可靠的自動化質檢能夠降低大量人工成本，創造經濟效益。

PaddleDetection提供了這類工業質檢場景的通用解決方案，本文是在PaddlePaddle 2.0.0rc框架版本下PaddleDetection的使用教程。

PaddleDetection模型庫與依賴安裝

PaddleDetection的文檔，本文主要參考以下模塊

使用教程

安裝說明
快速開始
訓練、評估流程
數據預處理及自定義數據集
配置模塊設計和介紹
詳細的配置信息和參數說明示例
IPython Notebook demo
遷移學習教程

模型庫

模型庫

!git clone https://gitee.com/paddlepaddle/PaddleDetection.git # 官方的國內鏡像 Cloning into 'PaddleDetection'... remote: Enumerating objects: 8784, done.[K remote: Counting objects: 100% (8784/8784), done.[K remote: Compressing objects: 100% (4820/4820), done.[K remote: Total 8784 (delta 6342), reused 5425 (delta 3850), pack-reused 0[K Receiving objects: 100% (8784/8784), 51.41 MiB | 4.19 MiB/s, done. Resolving deltas: 100% (6342/6342), done. Checking connectivity... done. # 中文字體鏡像 !git clone https://gitee.com/mirrors/noto-cjk.git # 安裝cocoAPI # !git clone https://github.com/cocodataset/cocoapi.git # 如果要加速的話可以找一個國內鏡像 !git clone https://gitee.com/firefox1200/cocoapi.git !cd cocoapi/PythonAPI&&make install !pip install pycocotools -i https://mirror.baidu.com/pypi/simple !pip install -r PaddleDetection/requirements.txt -i https://mirror.baidu.com/pypi/simple !pip install scikit-image -i https://mirror.baidu.com/pypi/simple !pip install kmeans -i https://mirror.baidu.com/pypi/simple

數據集信息

數智重慶.全球產業賦能創新大賽【賽場一】

數據準備

訓練集和測試集有多份，目前還可以直接從服務器上獲取
如果數據集的url失效，也可以從項目掛載的數據集中解壓，請讀者自行調整解壓位置

!wget http://tianchi-competition.oss-cn-hangzhou.aliyuncs.com/231763/round1/chongqing1_round1_train1_20191223.zip !wget http://tianchi-competition.oss-cn-hangzhou.aliyuncs.com/231763/round1/chongqing1_round1_testA_20191223.zip !wget http://tianchi-competition.oss-cn-hangzhou.aliyuncs.com/231763/round2/chongqing1_round2_train_20200213.zip !wget http://tianchi-competition.oss-cn-hangzhou.aliyuncs.com/231763/round1/chongqing1_round1_testB_20200210.zip --password=GXrhrvhEp92O*s2A # 解壓數據集 !unzip chongqing1_round1_train1_20191223.zip -d ./data !unzip chongqing1_round1_testA_20191223.zip -d ./data !unzip chongqing1_round2_train_20200213.zip -d ./data !unzip -P GXrhrvhEp92O*s2A chongqing1_round1_testB_20200210.zip -d ./data # !unzip chongqing1_round2_image_info.zip -d ./data !mkdir PaddleDetection/dataset/coco/annotations !mkdir PaddleDetection/dataset/coco/train2017 !mkdir PaddleDetection/dataset/coco/val2017 !mkdir PaddleDetection/dataset/coco/test2017 !cp data/chongqing1_round1_testA_20191223/images/*.jpg -d PaddleDetection/dataset/coco/test2017

EDA

此處提供了數據集簡單的EDA分析，供讀者參考
完整內容請查看：數據的簡單分析和可視

import os import json import pandas as pd import seaborn as sns import pandas as pd import matplotlib.pyplot as plt from matplotlib.font_manager import FontProperties import os, sys, zipfile import urllib.request import shutil import random import numpy as np from tqdm import tqdm import cv2 import kmeansmyfont = FontProperties(fname=r"/home/aistudio/noto-cjk/NotoSerifSC-Light.otf", size=12) plt.rcParams['figure.figsize'] = (12, 12) plt.rcParams['font.family']= myfont.get_family() plt.rcParams['font.sans-serif'] = myfont.get_name() plt.rcParams['axes.unicode_minus'] = False def generate_anno_eda(dataset_path):with open(os.path.join(dataset_path, 'annotations.json')) as f:anno = json.load(f)print('標簽類別:', anno['categories'])print('類別數量：', len(anno['categories']))print('訓練集圖片數量：', len(anno['images']))print('訓練集標簽數量：', len(anno['annotations']))total=[]for img in anno['images']:hw = (img['height'],img['width'])total.append(hw)unique = set(total)for k in unique:print('長寬為(%d,%d)的圖片數量為：'%k,total.count(k))ids=[]images_id=[]for i in anno['annotations']:ids.append(i['id'])images_id.append(i['image_id'])print('訓練集圖片數量:', len(anno['images']))print('unique id 數量：', len(set(ids)))print('unique image_id 數量', len(set(images_id)))# 創建類別標簽字典category_dic=dict([(i['id'],i['name']) for i in anno['categories']])counts_label=dict([(i['name'],0) for i in anno['categories']])for i in anno['annotations']:counts_label[category_dic[i['category_id']]] += 1label_list = counts_label.keys() # 各部分標簽print('標簽列表:', label_list)size = counts_label.values() # 各部分大小color = ['#FFB6C1', '#D8BFD8', '#9400D3', '#483D8B', '#4169E1', '#00FFFF','#B1FFF0','#ADFF2F','#EEE8AA','#FFA500','#FF6347'] # 各部分顏色# explode = [0.05, 0, 0] # 各部分突出值patches, l_text, p_text = plt.pie(size, labels=label_list, colors=color, labeldistance=1.1, autopct="%1.1f%%", shadow=False, startangle=90, pctdistance=0.6, textprops={'fontproperties':myfont})plt.axis("equal") # 設置橫軸和縱軸大小相等，這樣餅才是圓的plt.legend(prop=myfont)plt.show() generate_anno_eda('data/chongqing1_round1_train1_20191223') 標簽類別: [{'supercategory': '瓶蓋破損', 'id': 1, 'name': '瓶蓋破損'}, {'supercategory': '噴碼正常', 'id': 9, 'name': '噴碼正常'}, {'supercategory': '瓶蓋斷點', 'id': 5, 'name': '瓶蓋斷點'}, {'supercategory': '瓶蓋壞邊', 'id': 3, 'name': '瓶蓋壞邊'}, {'supercategory': '瓶蓋打旋', 'id': 4, 'name': '瓶蓋打旋'}, {'supercategory': '背景', 'id': 0, 'name': '背景'}, {'supercategory': '瓶蓋變形', 'id': 2, 'name': '瓶蓋變形'}, {'supercategory': '標貼氣泡', 'id': 8, 'name': '標貼氣泡'}, {'supercategory': '標貼歪斜', 'id': 6, 'name': '標貼歪斜'}, {'supercategory': '噴碼異常', 'id': 10, 'name': '噴碼異常'}, {'supercategory': '標貼起皺', 'id': 7, 'name': '標貼起皺'}] 類別數量： 11 訓練集圖片數量： 4516 訓練集標簽數量： 6945 長寬為(492,658)的圖片數量為： 4105 長寬為(3000,4096)的圖片數量為： 411 訓練集圖片數量: 4516 unique id 數量： 2011 unique image_id 數量 4516 標簽列表: dict_keys(['瓶蓋破損', '噴碼正常', '瓶蓋斷點', '瓶蓋壞邊', '瓶蓋打旋', '背景', '瓶蓋變形', '標貼氣泡', '標貼歪斜', '噴碼異常', '標貼起皺'])

[外鏈圖片轉存失敗,源站可能有防盜鏈機制,建議將圖片保存下來直接上傳(img-zBf63mJs-1610175324779)(output_19_1.png)]

generate_anno_eda('data/chongqing1_round2_train_20200213') 標簽類別: [{'supercategory': '瓶蓋破損', 'id': 1, 'name': '瓶蓋破損'}, {'supercategory': '噴碼正常', 'id': 9, 'name': '噴碼正常'}, {'supercategory': '瓶蓋斷點', 'id': 5, 'name': '瓶蓋斷點'}, {'supercategory': '瓶蓋壞邊', 'id': 3, 'name': '瓶蓋壞邊'}, {'supercategory': '瓶蓋打旋', 'id': 4, 'name': '瓶蓋打旋'}, {'supercategory': '瓶身破損', 'id': 12, 'name': '瓶身破損'}, {'supercategory': '背景', 'id': 0, 'name': '背景'}, {'supercategory': '瓶蓋變形', 'id': 2, 'name': '瓶蓋變形'}, {'supercategory': '瓶身氣泡', 'id': 13, 'name': '瓶身氣泡'}, {'supercategory': '標貼氣泡', 'id': 8, 'name': '標貼氣泡'}, {'supercategory': '標貼歪斜', 'id': 6, 'name': '標貼歪斜'}, {'supercategory': '噴碼異常', 'id': 10, 'name': '噴碼異常'}, {'supercategory': '酒液雜質', 'id': 11, 'name': '酒液雜質'}, {'supercategory': '標貼起皺', 'id': 7, 'name': '標貼起皺'}] 類別數量： 14 訓練集圖片數量： 2668 訓練集標簽數量： 3658 長寬為(3000,4096)的圖片數量為： 2668 訓練集圖片數量: 2668 unique id 數量： 3658 unique image_id 數量 2668 標簽列表: dict_keys(['瓶蓋破損', '噴碼正常', '瓶蓋斷點', '瓶蓋壞邊', '瓶蓋打旋', '瓶身破損', '背景', '瓶蓋變形', '瓶身氣泡', '標貼氣泡', '標貼歪斜', '噴碼異常', '酒液雜質', '標貼起皺'])

[外鏈圖片轉存失敗,源站可能有防盜鏈機制,建議將圖片保存下來直接上傳(img-A3LPVMeu-1610175324782)(output_20_1.png)]

數據清洗

數據集的標注并不是標準的coco格式，因此需要額外開展數據清洗工作
參考數據清洗的簡單demo
PaddleDetection目前尚不支持解析標注文件中的中文字符，在infer階段會報錯(不影響訓練)，讀者可以思考，是否在該階段也替換中文標注內容

def generate_washed_anno(dataset_path):with open(os.path.join(dataset_path, 'annotations.json')) as f:json_file = json.load(f) print('所有圖片的數量：', len(json_file['images'])) print('所有標注的數量：', len(json_file['annotations']))bg_imgs = set() # 所有標注中包含背景的圖片 idfor c in json_file['annotations']:if c['category_id'] == 0:bg_imgs.add(c['image_id'])print('所有標注中包含背景的圖片數量：', len(bg_imgs))bg_only_imgs = set() # 只有背景的圖片的 idfor img_id in bg_imgs:co = 0for c in json_file['annotations']:if c['image_id'] == img_id:co += 1if co == 1:bg_only_imgs.add(img_id)print('只包含背景的圖片數量：', len(bg_only_imgs))images_to_be_deleted = []for img in json_file['images']:if img['id'] in bg_only_imgs:images_to_be_deleted.append(img)# 刪除的是只有一個標注，且為 background 的的圖片print('待刪除圖片的數量：', len(images_to_be_deleted))for img in images_to_be_deleted:json_file['images'].remove(img)print('處理之后圖片的數量：', len(json_file['images']))ann_to_be_deleted = []for c in json_file['annotations']:if c['category_id'] == 0:ann_to_be_deleted.append(c)print('待刪除標注的數量：', len(ann_to_be_deleted))for img in ann_to_be_deleted:json_file['annotations'].remove(img)print('處理之后標注的數量：', len(json_file['annotations']))bg_cate = {'supercategory': '背景', 'id': 0, 'name': '背景'}json_file['categories'].remove(bg_cate)json_file['categories']for idx in range(len(json_file['annotations'])):json_file['annotations'][idx]['id'] = idxwith open(os.path.join(dataset_path, 'annotations_washed.json'), 'w') as f:json.dump(json_file, f, ensure_ascii=False)f.close()print('清洗后標注文件已生成！') generate_washed_anno('data/chongqing1_round1_train1_20191223') 所有圖片的數量： 4516 所有標注的數量： 6945 所有標注中包含背景的圖片數量： 1168 只包含背景的圖片數量： 1168 待刪除圖片的數量： 1168 處理之后圖片的數量： 3348 待刪除標注的數量： 1170 處理之后標注的數量： 5775 清洗后標注文件已生成！ generate_washed_anno('data/chongqing1_round2_train_20200213') 所有圖片的數量： 2668 所有標注的數量： 3658 所有標注中包含背景的圖片數量： 985 只包含背景的圖片數量： 985 待刪除圖片的數量： 985 處理之后圖片的數量： 1683 待刪除標注的數量： 985 處理之后標注的數量： 2673 清洗后標注文件已生成！

合并多份訓練集與標注

!mkdir data/chongqing_train_merge !mkdir data/chongqing_train_merge/images !mv data/chongqing1_round1_train1_20191223/images/*.jpg data/chongqing_train_merge/images/ !mv data/chongqing1_round2_train_20200213/images/*.jpg data/chongqing_train_merge/images/ with open(os.path.join('data/chongqing1_round1_train1_20191223', 'annotations.json')) as f:round1_full = json.load(f) f.close()with open(os.path.join('data/chongqing1_round1_train1_20191223', 'annotations_washed.json')) as f:round1 = json.load(f) f.close()with open(os.path.join('data/chongqing1_round2_train_20200213', 'annotations_washed.json')) as f:round2 = json.load(f) f.close()data_0={} data_0['images'] = [] data_0['info'] = round1['info'] data_0['license'] = ['AIC IVI Created'] # 復賽數據集在初賽基礎上增加了3類缺陷 data_0['categories'] = round2['categories'] t1 = round1['images'] t2 = round1['annotations']for item in round2['images']: item['id'] += len(round1_full['images'])t1.append(item)for ann in round2['annotations']:ann['image_id'] += len(round1_full['images'])ann['id'] += len(round1_full['annotations'])t2.append(ann) data_0['images'] = t1 data_0['annotations'] = t2 # print(data_0) # 保存到新的JSON文件，便于查看數據特點 json.dump(data_0,open('data/chongqing_train_merge/annotations_washed.json','w'),indent=4) # indent=4 更加美觀顯示 # 添加segmentation字段 def add_seg(json_anno):new_json_anno = []for c_ann in json_anno:c_category_id = c_ann['category_id']if not c_category_id:continuebbox = c_ann['bbox']c_ann['segmentation'] = []seg = []#bbox[] is x,y,w,h#left_topseg.append(bbox[0])seg.append(bbox[1])#left_bottomseg.append(bbox[0])seg.append(bbox[1] + bbox[3])#right_bottomseg.append(bbox[0] + bbox[2])seg.append(bbox[1] + bbox[3])#right_topseg.append(bbox[0] + bbox[2])seg.append(bbox[1])c_ann['segmentation'].append(seg)new_json_anno.append(c_ann)return new_json_annojson_file = 'data/chongqing_train_merge/annotations_washed.json' with open(json_file) as f:a=json.load(f)a['annotations'] = add_seg(a['annotations'])f.close()with open("data/chongqing_train_merge/new_ann_file.json","w") as f:json.dump(a, f, ensure_ascii=False)f.close()

訓練集和驗證集劃分

這里用最簡單的邏輯，每5張圖分1張到驗證集，另外4張放訓練集
也可以參考mmdetection框架：清洗數據后將數據分為訓練集和測試集并形成相應的annotations.json文件

def generate_coco_dataset(dataset_path, coco_path, split_ratio):with open(os.path.join(dataset_path, 'new_ann_file.json')) as f:merge = json.load(f) data_0={}data_0['images'] = []data_0['info'] = merge['info']data_0['license'] = ['AIC IVI Created']# 復賽數據集在初賽基礎上增加了3類缺陷data_0['categories'] = merge['categories']t1 = []t2 = []t3 = []t4 = []# print(len(merge['images']))index = range(0, len(merge['images']))slide = int(split_ratio * len(merge['images']))train_index = random.sample(index, slide)for i,l in enumerate(merge['images']):if i in train_index:t1.append(l)for ann in merge['annotations']:if l['id'] == ann['image_id']:t2.append(ann)shutil.copy(os.path.join(dataset_path, 'images', l['file_name']), os.path.join(coco_path, 'train2017'))data_0['images'] = t1data_0['annotations'] = t2# 保存到新的JSON文件，便于查看數據特點json.dump(data_0,open(os.path.join(coco_path, 'annotations','instances_train2017.json'),'w'),indent=4) # indent=4 更加美觀顯示for i,l in enumerate(merge['images']):if i not in train_index:t3.append(l)for ann in merge['annotations']:if l['id'] == ann['image_id']:t4.append(ann)shutil.copy(os.path.join(dataset_path, 'images', l['file_name']), os.path.join(coco_path, 'val2017'))data_0['images'] = t3data_0['annotations'] = t4# 保存到新的JSON文件，便于查看數據特點json.dump(data_0,open(os.path.join(coco_path, 'annotations','instances_val2017.json'),'w'),indent=4) # indent=4 更加美觀顯示f.close() generate_coco_dataset('data/chongqing_train_merge', 'PaddleDetection/dataset/coco', 0.8)

生成分類訓練數據集

import os import numpy as np from tqdm import tqdm import cv2def get_annotations(datadir, mode="train"):"""獲取瑕疵標注信息"""if mode == "train":ann_file = 'instances_train2017.json'else:ann_file = 'instances_val2017.json'with open(os.path.join(datadir, 'annotations', ann_file) ) as f:json_file = json.load(f) records = []for objs in json_file['annotations']:# print(objs)box = []label = []bbox = objs['bbox']# print(bbox)# print(objs['image_id'])x1 = int(bbox[0])y1 = int(bbox[1])x2 = int(bbox[0] + bbox[2])y2 = int(bbox[1] + bbox[3])box.append([x1, y1, x2, y2])# 這里有個問題，因為分類訓練的label是從0開始算的，所以需要將原始值減去1objs['category_id'] = objs['category_id'] - 1label.append(objs['category_id'])for img in json_file['images']:if img['id'] == objs['image_id']:# print(img)img_file = img['file_name']# 缺陷id是唯一的，保證不會重復fid = objs['id']voc_rec = {'im_file': img_file,'im_id': fid,'gt_class': label,'gt_bbox': box}records.append(voc_rec)f.close()return recordsdef generate_data(datadir, save_dir, records, mode="train"):im_out = []if mode == "train":images_dir = 'train2017'else:images_dir = 'val2017'for record in tqdm(records):img = cv2.imread(os.path.join(datadir, images_dir, record["im_file"]))# img = imageio.imread(os.path.join(datadir, images_dir, record["im_file"]))ffile = record["im_file"][:-4]fid = record["im_id"]box = record["gt_bbox"][0]fl = record["gt_class"][0]# print(fl)# print(box)# print(img)# print(box[1])im = img[box[1]: box[3], box[0]: box[2]]fname = '{}/{}/{}_{}.jpg'.format(save_dir, mode, str(ffile), str(fid))cv2.imwrite(fname, im)outname = '{}/{}_{}.jpg'.format(mode, str(ffile), str(fid))im_out.append("{} {}".format(outname, fl))with open("{}/{}_list.txt".format(save_dir,mode), "w") as f:f.write("\n".join(im_out))f.close()

數據集相關計算

RGB通道的均值和標準差

# 先把測試集圖片都移到coco test dataset目錄下 !mv data/chongqing1_round1_testB_20200210/images/*.jpg PaddleDetection/dataset/coco/test2017/ !mv data/chongqing1_round1_testA_20191223/images/*.jpg PaddleDetection/dataset/coco/test2017/ """ 計算RGB通道的均值和標準差 """def compute(path):file_names = os.listdir(path)per_image_Rmean = []per_image_Gmean = []per_image_Bmean = []per_image_Rstd = []per_image_Gstd = []per_image_Bstd = []for file_name in file_names:img = cv2.imread(os.path.join(path, file_name), 1)per_image_Rmean.append(np.mean(img[:, :, 0]))per_image_Gmean.append(np.mean(img[:, :, 1]))per_image_Bmean.append(np.mean(img[:, :, 2]))per_image_Rstd.append(np.std(img[:, :, 0]))per_image_Gstd.append(np.std(img[:, :, 1]))per_image_Bstd.append(np.std(img[:, :, 2]))R_mean = np.mean(per_image_Rmean)/255.0G_mean = np.mean(per_image_Gmean)/255.0B_mean = np.mean(per_image_Bmean)/255.0R_std = np.mean(per_image_Rstd)/255.0G_std = np.mean(per_image_Gstd)/255.0B_std = np.mean(per_image_Bstd)/255.0image_mean = [R_mean, G_mean, B_mean]image_std = [R_std, G_std, B_std]return image_mean, image_std path = 'PaddleDetection/dataset/coco/test2017' image_mean, image_std = compute(path) print(image_mean, image_std) # [0.30180695179185196, 0.21959776452510077, 0.17572622616932373] [0.31354065666185776, 0.23364904437622586, 0.2054456915069016] [0.30180695179185196, 0.21959776452510077, 0.17572622616932373] [0.31354065666185776, 0.23364904437622588, 0.20544569150690165]

Kmeans聚類計算anchor boxes

如果讀者需要使用Yolo系列模型，可以參考Kmeans聚類結果設置anchors

def iou(box, clusters):"""Calculates the Intersection over Union (IoU) between a box and k clusters.:param box: tuple or array, shifted to the origin (i. e. width and height):param clusters: numpy array of shape (k, 2) where k is the number of clusters:return: numpy array of shape (k, 0) where k is the number of clusters"""x = np.minimum(clusters[:, 0], box[0])y = np.minimum(clusters[:, 1], box[1])if np.count_nonzero(x == 0) > 0 or np.count_nonzero(y == 0) > 0:raise ValueError("Box has no area")intersection = x * ybox_area = box[0] * box[1]cluster_area = clusters[:, 0] * clusters[:, 1]iou_ = intersection / (box_area + cluster_area - intersection)return iou_def avg_iou(boxes, clusters):"""Calculates the average Intersection over Union (IoU) between a numpy array of boxes and k clusters.:param boxes: numpy array of shape (r, 2), where r is the number of rows:param clusters: numpy array of shape (k, 2) where k is the number of clusters:return: average IoU as a single float"""return np.mean([np.max(iou(boxes[i], clusters)) for i in range(boxes.shape[0])])def translate_boxes(boxes):"""Translates all the boxes to the origin.:param boxes: numpy array of shape (r, 4):return: numpy array of shape (r, 2)"""new_boxes = boxes.copy()for row in range(new_boxes.shape[0]):new_boxes[row][2] = np.abs(new_boxes[row][2] - new_boxes[row][0])new_boxes[row][3] = np.abs(new_boxes[row][3] - new_boxes[row][1])return np.delete(new_boxes, [0, 1], axis=1)def kmeans(boxes, k, dist=np.median):"""Calculates k-means clustering with the Intersection over Union (IoU) metric.:param boxes: numpy array of shape (r, 2), where r is the number of rows:param k: number of clusters:param dist: distance function:return: numpy array of shape (k, 2)"""rows = boxes.shape[0]distances = np.empty((rows, k))last_clusters = np.zeros((rows,))np.random.seed()# the Forgy method will fail if the whole array contains the same rowsclusters = boxes[np.random.choice(rows, k, replace=False)]while True:for row in range(rows):distances[row] = 1 - iou(boxes[row], clusters)nearest_clusters = np.argmin(distances, axis=1)if (last_clusters == nearest_clusters).all():breakfor cluster in range(k):clusters[cluster] = dist(boxes[nearest_clusters == cluster], axis=0)last_clusters = nearest_clustersreturn clusters CLUSTERS = 9def load_dataset(path):dataset = []with open(os.path.join(path, 'annotations', 'instances_train2017.json')) as f:json_file = json.load(f) for item in json_file['images']: height = item['height']width = item['width']for objs in json_file['annotations']:if(objs['image_id'] == item['id']):bbox = objs['bbox']xmin = int(bbox[0]) / widthymin = int(bbox[1]) / heightxmax = int(bbox[0] + bbox[2]) / widthymax = int(bbox[1] + bbox[3]) / heightxmin = np.float64(xmin)ymin = np.float64(ymin)xmax = np.float64(xmax)ymax = np.float64(ymax)if xmax == xmin or ymax == ymin:print(item['file_name'])dataset.append([xmax - xmin, ymax - ymin])f.close()return np.array(dataset)#print(__file__) data = load_dataset('PaddleDetection/dataset/coco/')out = kmeans(data, k=CLUSTERS)*608 #clusters = [[10,13],[16,30],[33,23],[30,61],[62,45],[59,119],[116,90],[156,198],[373,326]] #out= np.array(clusters)/416.0 print(out) print("Accuracy: {:.2f}%".format(avg_iou(data, out/608) * 100)) print("Boxes:\n {}-{}".format(out[:, 0], out[:, 1]))ratios = np.around(out[:, 0] / out[:, 1], decimals=2).tolist() print("Ratios:\n {}".format(sorted(ratios))) [[ 57.2887538 44.48780488][329.87234043 133.46341463][ 23.10030395 56.84552846][329.6796875 337.44 ][332.6443769 148.29268293][ 7.8671875 11.552 ][145.99392097 49.43089431][ 4.3046875 5.67466667][ 16.63221884 21.00813008]] Accuracy: 64.98% Boxes:[ 57.2887538 329.87234043 23.10030395 329.6796875 332.64437697.8671875 145.99392097 4.3046875 16.63221884]-[ 44.48780488 133.46341463 56.84552846 337.44 148.2926829311.552 49.43089431 5.67466667 21.00813008] Ratios:[0.41, 0.68, 0.76, 0.79, 0.98, 1.29, 2.24, 2.47, 2.95]

遷移學習

遷移學習為利用已有知識，對新知識進行學習。例如利用ImageNet分類預訓練模型做初始化來訓練檢測模型，利用在COCO數據集上的檢測模型做初始化來訓練基于PascalVOC數據集的檢測模型。

在進行遷移學習時，由于會使用不同的數據集，數據類別數與COCO/VOC數據類別不同，導致在加載PaddlePaddle開源模型時，與類別數相關的權重（例如分類模塊的fc層）會出現維度不匹配的問題；另外，如果需要結構更加復雜的模型，需要對已有開源模型結構進行調整，對應權重也需要選擇性加載。因此，需要檢測庫能夠指定參數字段，在加載模型時不加載匹配的權重。

PaddleDetection進行遷移學習

加載預訓練模型

在進行遷移學習時，由于會使用不同的數據集，數據類別數與COCO/VOC數據類別不同，導致在加載開源模型(如COCO預訓練模型)時，與類別數相關的權重（例如分類模塊的fc層）會出現維度不匹配的問題；另外，如果需要結構更加復雜的模型，需要對已有開源模型結構進行調整，對應權重也需要選擇性加載。因此，需要在加載模型時不加載不能匹配的權重。

在遷移學習中，對預訓練模型進行選擇性加載，支持如下兩種遷移學習方式：

直接加載預訓練權重（推薦方式）

模型中和預訓練模型中對應參數形狀不同的參數將自動被忽略，例如：

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -u tools/train.py -c configs/faster_rcnn_r50_1x.yml \-o pretrain_weights=https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_r50_1x.tar

使用finetune_exclude_pretrained_params參數控制忽略參數名

可以顯示的指定訓練過程中忽略參數的名字，任何參數名均可加入finetune_exclude_pretrained_params中，為實現這一目的，可通過如下方式實現：

在 YMAL 配置文件中通過設置finetune_exclude_pretrained_params字段。可參考配置文件

在 train.py的啟動參數中設置finetune_exclude_pretrained_params。例如：

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -u tools/train.py -c configs/faster_rcnn_r50_1x.yml \-o pretrain_weights=https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_r50_1x.tar \finetune_exclude_pretrained_params=['cls_score','bbox_pred'] \

說明：

pretrain_weights的路徑為COCO數據集上開源的faster RCNN模型鏈接，完整模型鏈接可參考MODEL_ZOO

finetune_exclude_pretrained_params中設置參數字段，如果參數名能夠匹配以上參數字段（通配符匹配方式），則在模型加載時忽略該參數。

如果用戶需要利用自己的數據進行finetune，模型結構不變，只需要忽略與類別數相關的參數，不同模型類型所對應的忽略參數字段如下表所示：

模型類型忽略參數字段

Faster RCNN	cls_score, bbox_pred
Cascade RCNN	cls_score, bbox_pred
Mask RCNN	cls_score, bbox_pred, mask_fcn_logits
Cascade-Mask RCNN	cls_score, bbox_pred, mask_fcn_logits
RetinaNet	retnet_cls_pred_fpn
SSD	^conv2d_
YOLOv3	yolo_output

示例：遷移學習

在模型庫找到想要的預訓練模型，獲取下載鏈接

Faster & Mask R-CNN

骨架網絡網絡類型每張GPU圖片個數學習率策略推理時間(fps)Box APMask AP下載

ResNet101-vd-FPN

CascadeClsAware Faster

44.7(softnms)

下載鏈接

參考配置文件

configs/faster_fpn_reader.yml

請注意，在TestReader中，anno_path為None，否則會因為加載的標注文件有中文字符，預測時會報錯

TrainReader:inputs_def:fields: ['image', 'im_info', 'im_id', 'gt_bbox', 'gt_class', 'is_crowd']dataset:!COCODataSetimage_dir: train2017anno_path: annotations/instances_train2017.jsondataset_dir: dataset/cocosample_transforms:- !DecodeImageto_rgb: true- !RandomFlipImageprob: 0.5- !NormalizeImageis_channel_first: falseis_scale: truemean: [0.30180695179185196, 0.21959776452510077, 0.17572622616932373] std: [0.31354065666185776, 0.23364904437622586, 0.2054456915069016]- !ResizeImagetarget_size: 800max_size: 1333interp: 1use_cv2: true- !Permuteto_bgr: falsechannel_first: truebatch_transforms:- !PadBatchpad_to_stride: 32use_padded_im_info: falsebatch_size: 1shuffle: trueworker_num: 2use_process: falseEvalReader:inputs_def:fields: ['image', 'im_info', 'im_id', 'im_shape']# for voc#fields: ['image', 'im_info', 'im_id', 'im_shape', 'gt_bbox', 'gt_class', 'is_difficult']dataset:!COCODataSetimage_dir: val2017anno_path: annotations/instances_val2017.jsondataset_dir: dataset/cocosample_transforms:- !DecodeImageto_rgb: truewith_mixup: false- !NormalizeImageis_channel_first: falseis_scale: truemean: [0.30180695179185196, 0.21959776452510077, 0.17572622616932373]std: [0.31354065666185776, 0.23364904437622586, 0.2054456915069016]- !ResizeImageinterp: 1max_size: 1333target_size: 800use_cv2: true- !Permutechannel_first: trueto_bgr: falsebatch_transforms:- !PadBatchpad_to_stride: 32use_padded_im_info: truebatch_size: 1shuffle: falsedrop_empty: falseworker_num: 2TestReader:inputs_def:# set image_shape if neededfields: ['image', 'im_info', 'im_id', 'im_shape']dataset:!ImageFolderanno_path: Nonesample_transforms:- !DecodeImageto_rgb: truewith_mixup: false- !NormalizeImageis_channel_first: falseis_scale: truemean: [0.30180695179185196, 0.21959776452510077, 0.17572622616932373]std: [0.31354065666185776, 0.23364904437622586, 0.2054456915069016]- !ResizeImageinterp: 1max_size: 1333target_size: 800use_cv2: true- !Permutechannel_first: trueto_bgr: falsebatch_transforms:- !PadBatchpad_to_stride: 32use_padded_im_info: truebatch_size: 1shuffle: false

configs/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms.yml

architecture: CascadeRCNNClsAware max_iters: 90000 snapshot_iter: 10000 use_gpu: true log_iter: 200 save_dir: output pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_vd_pretrained.tar weights: output/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms/model_final metric: COCO num_classes: 11CascadeRCNNClsAware:backbone: ResNetfpn: FPNrpn_head: FPNRPNHeadroi_extractor: FPNRoIAlignbbox_head: CascadeBBoxHeadbbox_assigner: CascadeBBoxAssignerResNet:norm_type: bndepth: 101feature_maps: [2, 3, 4, 5]freeze_at: 2variant: dFPN:min_level: 2max_level: 6num_chan: 256spatial_scale: [0.03125, 0.0625, 0.125, 0.25]FPNRPNHead:anchor_generator:anchor_sizes: [32, 64, 128, 256, 512]aspect_ratios: [0.5, 1.0, 2.0]stride: [16.0, 16.0]variance: [1.0, 1.0, 1.0, 1.0]anchor_start_size: 32min_level: 2max_level: 6num_chan: 256rpn_target_assign:rpn_batch_size_per_im: 256rpn_fg_fraction: 0.5rpn_positive_overlap: 0.7rpn_negative_overlap: 0.3rpn_straddle_thresh: 0.0train_proposal:min_size: 0.0nms_thresh: 0.7pre_nms_top_n: 2000post_nms_top_n: 2000test_proposal:min_size: 0.0nms_thresh: 0.7pre_nms_top_n: 1000post_nms_top_n: 1000FPNRoIAlign:canconical_level: 4canonical_size: 224min_level: 2max_level: 5box_resolution: 14sampling_ratio: 2CascadeBBoxAssigner:batch_size_per_im: 512bbox_reg_weights: [10, 20, 30]bg_thresh_lo: [0.0, 0.0, 0.0]bg_thresh_hi: [0.5, 0.6, 0.7]fg_thresh: [0.5, 0.6, 0.7]fg_fraction: 0.25class_aware: TrueCascadeBBoxHead:head: CascadeTwoFCHeadnms: MultiClassSoftNMSCascadeTwoFCHead:mlp_dim: 1024MultiClassSoftNMS:score_threshold: 0.01keep_top_k: 300softnms_sigma: 0.5LearningRate:base_lr: 0.0005schedulers:- !PiecewiseDecaygamma: 0.1milestones: [60000, 80000]- !LinearWarmupstart_factor: 0.0steps: 2000OptimizerBuilder:optimizer:momentum: 0.9type: Momentumregularizer:factor: 0.0001type: L2_READER_: 'faster_fpn_reader.yml' TrainReader:batch_size: 4

訓練之前要再次確認Numpy版本

如果出現下列報錯，是因為Numpy版本過高導致的，因為在前面下依賴的時候將Numpy版本升級，與PaddlePaddle2.0.0rc不適配

2021-01-05 09:31:37,618-INFO: Start evaluate... Loading and preparing results... DONE (t=2.99s) creating index... index created! Traceback (most recent call last):File "tools/train.py", line 399, in <module>main()File "tools/train.py", line 320, in maincfg['EvalReader']['dataset'])File "/home/aistudio/PaddleDetection/ppdet/utils/eval_utils.py", line 241, in eval_resultssave_only=save_only)File "/home/aistudio/PaddleDetection/ppdet/utils/coco_eval.py", line 102, in bbox_evalmap_stats = cocoapi_eval(outfile, 'bbox', coco_gt=coco_gt)File "/home/aistudio/PaddleDetection/ppdet/utils/coco_eval.py", line 244, in cocoapi_evalcoco_eval = COCOeval(coco_gt, coco_dt, style)File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/pycocotools-2.0-py3.7-linux-x86_64.egg/pycocotools/cocoeval.py", line 75, in __init__self.params = Params(iouType=iouType) # parametersFile "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/pycocotools-2.0-py3.7-linux-x86_64.egg/pycocotools/cocoeval.py", line 527, in __init__self.setDetParams()File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/pycocotools-2.0-py3.7-linux-x86_64.egg/pycocotools/cocoeval.py", line 506, in setDetParamsself.iouThrs = np.linspace(.5, 0.95, np.round((0.95 - .5) / .05) + 1, endpoint=True)File "<__array_function__ internals>", line 6, in linspaceFile "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/numpy/core/function_base.py", line 113, in linspacenum = operator.index(num) TypeError: 'numpy.float64' object cannot be interpreted as an integer terminate called without an active exception-------------------------------------- C++ Traceback (most recent call last): -------------------------------------- 0 paddle::framework::SignalHandle(char const*, int) 1 paddle::platform::GetCurrentTraceBackString[abi:cxx11]()---------------------- Error Message Summary: ---------------------- FatalError: `Process abort signal` is detected by the operating system.[TimeInfo: *** Aborted at 1609810302 (unix time) try "date -d @1609810302" if you are using GNU date ***][SignalInfo: *** SIGABRT (@0x3e800005683) received by PID 22147 (TID 0x7ff15d3e1700) from PID 22147 ***]Aborted (core dumped) # 調整Numpy版本 !pip install -U numpy==1.17.0 Looking in indexes: https://mirror.baidu.com/pypi/simple/ Collecting numpy==1.17.0 [?25l Downloading https://mirror.baidu.com/pypi/packages/05/4b/55cfbfd3e5e85016eeef9f21c0ec809d978706a0d60b62cc28aeec8c792f/numpy-1.17.0-cp37-cp37m-manylinux1_x86_64.whl (20.3MB) [K |████████████████████████████████| 20.3MB 9.6MB/s eta 0:00:011 [31mERROR: xarray 0.16.2 has requirement pandas>=0.25, but you'll have pandas 0.23.4 which is incompatible.[0m [31mERROR: parl 1.3.2 has requirement pyarrow==0.13.0, but you'll have pyarrow 2.0.0 which is incompatible.[0m [?25hInstalling collected packages: numpyFound existing installation: numpy 1.19.4Uninstalling numpy-1.19.4:Successfully uninstalled numpy-1.19.4 Successfully installed numpy-1.17.0 # 開始訓練并打開visualdl，此處只訓練了3萬輪，訓練時間在6小時左右，讀者可以嘗試完整跑完整個訓練過程，mAP會進一步提升 !cd PaddleDetection && python -u tools/train.py -c configs/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms.yml \-o pretrain_weights=https://paddlemodels.bj.bcebos.com/object_detection/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms.tar \finetune_exclude_pretrained_params=['cls_score','bbox_pred'] \--eval -o use_gpu=true --use_vdl=True --vdl_log_dir=vdl_dir/scalar You are using Paddle compiled with TensorRT, but TensorRT dynamic library is not found. Ignore this if TensorRT is not needed. /home/aistudio/PaddleDetection/ppdet/utils/voc_utils.py:70: DeprecationWarning: invalid escape sequence \.elif re.match('test\.txt', fname): /home/aistudio/PaddleDetection/ppdet/utils/voc_utils.py:68: DeprecationWarning: invalid escape sequence \.if re.match('trainval\.txt', fname): /home/aistudio/PaddleDetection/ppdet/core/workspace.py:118: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop workingisinstance(merge_dct[k], collections.Mapping)): /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:297: UserWarning: /home/aistudio/PaddleDetection/ppdet/modeling/backbones/fpn.py:108 The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.op_type, op_type, EXPRESSION_MAP[method_name])) 2021-01-05 20:57:14,933-INFO: If regularizer of a Parameter has been set by 'fluid.ParamAttr' or 'fluid.WeightNormParamAttr' already. The Regularization[L2Decay, regularization_coeff=0.000100] in Optimizer will not take effect, and it will only be applied to other Parameters! /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:297: UserWarning: /home/aistudio/PaddleDetection/ppdet/modeling/roi_heads/cascade_head.py:245 The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.op_type, op_type, EXPRESSION_MAP[method_name])) /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:297: UserWarning: /home/aistudio/PaddleDetection/ppdet/modeling/roi_heads/cascade_head.py:249 The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.op_type, op_type, EXPRESSION_MAP[method_name])) /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:297: UserWarning: /home/aistudio/PaddleDetection/ppdet/modeling/roi_heads/cascade_head.py:255 The behavior of expression A / B has been unified with elementwise_div(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_div(X, Y, axis=0) instead of A / B. This transitional warning will be dropped in the future.op_type, op_type, EXPRESSION_MAP[method_name])) /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/nn.py:13362: DeprecationWarning: inspect.getargspec() is deprecated since Python 3.0, use inspect.signature() or inspect.getfullargspec()args = inspect.getargspec(self._func) /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/__init__.py:107: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop workingfrom collections import MutableMapping /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/colors.py:53: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop workingfrom collections import Sized loading annotations into memory... Done (t=0.01s) creating index... index created! W0105 20:57:17.060937 1092 device_context.cc:320] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.0, Runtime API Version: 10.1 W0105 20:57:17.135076 1092 device_context.cc:330] device: 0, cuDNN Version: 7.6. 2021-01-05 20:57:21,404-INFO: Downloading ResNet101_vd_pretrained.tar from https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_vd_pretrained.tar 100%|████████████████████████████████| 175040/175040 [00:02<00:00, 61135.79KB/s] 2021-01-05 20:57:24,826-INFO: Decompressing /home/aistudio/.cache/paddle/weights/ResNet101_vd_pretrained.tar... /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/io.py:2216: UserWarning: This list is not set, Because of Paramerter not found in program. There are: fc_0.b_0 fc_0.w_0format(" ".join(unused_para_list))) loading annotations into memory... Done (t=0.04s) creating index... index created! /home/aistudio/PaddleDetection/ppdet/data/reader.py:89: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop workingif isinstance(item, collections.Sequence) and len(item) == 0: 2021-01-05 20:57:27,623-INFO: iter: 0, lr: 0.000000, 'loss_cls_0': '2.453082', 'loss_loc_0': '0.000000', 'loss_cls_1': '1.257472', 'loss_loc_1': '0.000003', 'loss_cls_2': '0.568658', 'loss_loc_2': '0.000003', 'loss_rpn_cls': '0.689106', 'loss_rpn_bbox': '0.037696', 'loss': '5.006020', eta: 0:44:10, batch_cost: 0.08836 sec, ips: 45.26822 images/sec 2021-01-05 21:00:12,238-INFO: iter: 200, lr: 0.000050, 'loss_cls_0': '0.140615', 'loss_loc_0': '0.000022', 'loss_cls_1': '0.083600', 'loss_loc_1': '0.000143', 'loss_cls_2': '0.040301', 'loss_loc_2': '0.000162', 'loss_rpn_cls': '0.641392', 'loss_rpn_bbox': '0.098699', 'loss': '1.053611', eta: 6:50:00, batch_cost: 0.82552 sec, ips: 4.84544 images/sec 2021-01-05 21:03:09,020-INFO: iter: 400, lr: 0.000100, 'loss_cls_0': '0.030720', 'loss_loc_0': '0.001567', 'loss_cls_1': '0.022762', 'loss_loc_1': '0.000066', 'loss_cls_2': '0.014929', 'loss_loc_2': '0.000068', 'loss_rpn_cls': '0.180909', 'loss_rpn_bbox': '0.056140', 'loss': '0.312434', eta: 7:16:10, batch_cost: 0.88413 sec, ips: 4.52421 images/sec 2021-01-05 21:06:02,715-INFO: iter: 600, lr: 0.000150, 'loss_cls_0': '0.024992', 'loss_loc_0': '0.003722', 'loss_cls_1': '0.019258', 'loss_loc_1': '0.001492', 'loss_cls_2': '0.011919', 'loss_loc_2': '0.000084', 'loss_rpn_cls': '0.112500', 'loss_rpn_bbox': '0.043896', 'loss': '0.217339', eta: 7:06:32, batch_cost: 0.87050 sec, ips: 4.59509 images/sec 2021-01-05 21:08:59,673-INFO: iter: 800, lr: 0.000200, 'loss_cls_0': '0.026494', 'loss_loc_0': '0.007774', 'loss_cls_1': '0.016447', 'loss_loc_1': '0.002729', 'loss_cls_2': '0.010184', 'loss_loc_2': '0.000176', 'loss_rpn_cls': '0.094266', 'loss_rpn_bbox': '0.042120', 'loss': '0.224630', eta: 7:10:24, batch_cost: 0.88441 sec, ips: 4.52277 images/sec 2021-01-05 21:11:34,220-INFO: iter: 1000, lr: 0.000250, 'loss_cls_0': '0.030684', 'loss_loc_0': '0.009519', 'loss_cls_1': '0.017671', 'loss_loc_1': '0.004104', 'loss_cls_2': '0.010640', 'loss_loc_2': '0.000898', 'loss_rpn_cls': '0.076730', 'loss_rpn_bbox': '0.035291', 'loss': '0.190934', eta: 6:11:27, batch_cost: 0.76854 sec, ips: 5.20470 images/sec 2021-01-05 21:14:13,594-INFO: iter: 1200, lr: 0.000300, 'loss_cls_0': '0.026984', 'loss_loc_0': '0.008838', 'loss_cls_1': '0.016142', 'loss_loc_1': '0.003650', 'loss_cls_2': '0.009341', 'loss_loc_2': '0.000609', 'loss_rpn_cls': '0.068042', 'loss_rpn_bbox': '0.037983', 'loss': '0.184763', eta: 6:24:34, batch_cost: 0.80121 sec, ips: 4.99242 images/sec 2021-01-05 21:16:57,206-INFO: iter: 1400, lr: 0.000350, 'loss_cls_0': '0.036046', 'loss_loc_0': '0.013957', 'loss_cls_1': '0.019316', 'loss_loc_1': '0.006035', 'loss_cls_2': '0.010939', 'loss_loc_2': '0.001333', 'loss_rpn_cls': '0.061063', 'loss_rpn_bbox': '0.034824', 'loss': '0.199633', eta: 6:27:57, batch_cost: 0.81388 sec, ips: 4.91471 images/sec 2021-01-05 21:19:33,242-INFO: iter: 1600, lr: 0.000400, 'loss_cls_0': '0.032256', 'loss_loc_0': '0.014501', 'loss_cls_1': '0.020583', 'loss_loc_1': '0.007055', 'loss_cls_2': '0.011019', 'loss_loc_2': '0.001614', 'loss_rpn_cls': '0.062161', 'loss_rpn_bbox': '0.033353', 'loss': '0.190411', eta: 6:11:19, batch_cost: 0.78450 sec, ips: 5.09879 images/sec 2021-01-05 21:22:23,286-INFO: iter: 1800, lr: 0.000450, 'loss_cls_0': '0.032514', 'loss_loc_0': '0.018158', 'loss_cls_1': '0.018047', 'loss_loc_1': '0.009745', 'loss_cls_2': '0.010193', 'loss_loc_2': '0.002037', 'loss_rpn_cls': '0.053048', 'loss_rpn_bbox': '0.035240', 'loss': '0.196530', eta: 6:39:22, batch_cost: 0.84975 sec, ips: 4.70728 images/sec 2021-01-05 21:24:59,835-INFO: iter: 2000, lr: 0.000500, 'loss_cls_0': '0.043271', 'loss_loc_0': '0.021829', 'loss_cls_1': '0.021004', 'loss_loc_1': '0.010664', 'loss_cls_2': '0.012062', 'loss_loc_2': '0.002792', 'loss_rpn_cls': '0.048725', 'loss_rpn_bbox': '0.031871', 'loss': '0.202684', eta: 6:05:18, batch_cost: 0.78281 sec, ips: 5.10982 images/sec 2021-01-05 21:27:41,834-INFO: iter: 2200, lr: 0.000500, 'loss_cls_0': '0.041003', 'loss_loc_0': '0.021591', 'loss_cls_1': '0.020952', 'loss_loc_1': '0.011292', 'loss_cls_2': '0.011224', 'loss_loc_2': '0.002525', 'loss_rpn_cls': '0.043570', 'loss_rpn_bbox': '0.028753', 'loss': '0.197997', eta: 6:15:16, batch_cost: 0.80994 sec, ips: 4.93861 images/sec 2021-01-05 21:30:27,211-INFO: iter: 2400, lr: 0.000500, 'loss_cls_0': '0.043053', 'loss_loc_0': '0.024386', 'loss_cls_1': '0.019802', 'loss_loc_1': '0.013453', 'loss_cls_2': '0.010550', 'loss_loc_2': '0.003711', 'loss_rpn_cls': '0.037526', 'loss_rpn_bbox': '0.031488', 'loss': '0.199871', eta: 6:20:28, batch_cost: 0.82711 sec, ips: 4.83613 images/sec 2021-01-05 21:32:59,545-INFO: iter: 2600, lr: 0.000500, 'loss_cls_0': '0.053189', 'loss_loc_0': '0.029826', 'loss_cls_1': '0.024508', 'loss_loc_1': '0.013675', 'loss_cls_2': '0.012944', 'loss_loc_2': '0.004038', 'loss_rpn_cls': '0.038670', 'loss_rpn_bbox': '0.031219', 'loss': '0.233869', eta: 5:47:50, batch_cost: 0.76170 sec, ips: 5.25140 images/sec 2021-01-05 21:35:42,920-INFO: iter: 2800, lr: 0.000500, 'loss_cls_0': '0.047952', 'loss_loc_0': '0.026338', 'loss_cls_1': '0.021230', 'loss_loc_1': '0.014309', 'loss_cls_2': '0.011081', 'loss_loc_2': '0.003688', 'loss_rpn_cls': '0.043751', 'loss_rpn_bbox': '0.031890', 'loss': '0.208616', eta: 6:07:34, batch_cost: 0.81082 sec, ips: 4.93326 images/sec 2021-01-05 21:38:28,805-INFO: iter: 3000, lr: 0.000500, 'loss_cls_0': '0.053722', 'loss_loc_0': '0.022755', 'loss_cls_1': '0.022325', 'loss_loc_1': '0.012091', 'loss_cls_2': '0.012037', 'loss_loc_2': '0.003324', 'loss_rpn_cls': '0.045868', 'loss_rpn_bbox': '0.034881', 'loss': '0.226761', eta: 6:15:58, batch_cost: 0.83551 sec, ips: 4.78747 images/sec 2021-01-05 21:41:21,359-INFO: iter: 3200, lr: 0.000500, 'loss_cls_0': '0.047701', 'loss_loc_0': '0.024353', 'loss_cls_1': '0.017759', 'loss_loc_1': '0.011699', 'loss_cls_2': '0.009760', 'loss_loc_2': '0.002277', 'loss_rpn_cls': '0.038088', 'loss_rpn_bbox': '0.030854', 'loss': '0.197340', eta: 6:22:42, batch_cost: 0.85680 sec, ips: 4.66851 images/sec 2021-01-05 21:44:13,735-INFO: iter: 3400, lr: 0.000500, 'loss_cls_0': '0.050315', 'loss_loc_0': '0.030603', 'loss_cls_1': '0.020508', 'loss_loc_1': '0.015659', 'loss_cls_2': '0.010529', 'loss_loc_2': '0.004211', 'loss_rpn_cls': '0.032069', 'loss_rpn_bbox': '0.027820', 'loss': '0.197400', eta: 6:21:24, batch_cost: 0.86034 sec, ips: 4.64934 images/sec 2021-01-05 21:46:56,238-INFO: iter: 3600, lr: 0.000500, 'loss_cls_0': '0.049907', 'loss_loc_0': '0.034421', 'loss_cls_1': '0.020692', 'loss_loc_1': '0.017621', 'loss_cls_2': '0.010607', 'loss_loc_2': '0.004791', 'loss_rpn_cls': '0.029823', 'loss_rpn_bbox': '0.023192', 'loss': '0.210793', eta: 6:00:39, batch_cost: 0.81969 sec, ips: 4.87987 images/sec 2021-01-05 21:49:36,743-INFO: iter: 3800, lr: 0.000500, 'loss_cls_0': '0.057379', 'loss_loc_0': '0.037542', 'loss_cls_1': '0.023396', 'loss_loc_1': '0.019874', 'loss_cls_2': '0.010865', 'loss_loc_2': '0.005763', 'loss_rpn_cls': '0.031163', 'loss_rpn_bbox': '0.025450', 'loss': '0.222307', eta: 5:50:30, batch_cost: 0.80270 sec, ips: 4.98318 images/sec 2021-01-05 21:52:21,032-INFO: iter: 4000, lr: 0.000500, 'loss_cls_0': '0.056744', 'loss_loc_0': '0.039169', 'loss_cls_1': '0.023172', 'loss_loc_1': '0.021154', 'loss_cls_2': '0.010635', 'loss_loc_2': '0.005901', 'loss_rpn_cls': '0.028357', 'loss_rpn_bbox': '0.024548', 'loss': '0.224867', eta: 5:55:45, batch_cost: 0.82098 sec, ips: 4.87224 images/sec 2021-01-05 21:55:07,256-INFO: iter: 4200, lr: 0.000500, 'loss_cls_0': '0.055949', 'loss_loc_0': '0.038298', 'loss_cls_1': '0.024296', 'loss_loc_1': '0.024089', 'loss_cls_2': '0.010945', 'loss_loc_2': '0.007038', 'loss_rpn_cls': '0.027430', 'loss_rpn_bbox': '0.025825', 'loss': '0.228521', eta: 5:56:45, batch_cost: 0.82967 sec, ips: 4.82118 images/sec 2021-01-05 21:57:53,794-INFO: iter: 4400, lr: 0.000500, 'loss_cls_0': '0.058873', 'loss_loc_0': '0.039653', 'loss_cls_1': '0.027211', 'loss_loc_1': '0.028895', 'loss_cls_2': '0.011239', 'loss_loc_2': '0.008330', 'loss_rpn_cls': '0.023555', 'loss_rpn_bbox': '0.023535', 'loss': '0.239780', eta: 5:56:02, batch_cost: 0.83445 sec, ips: 4.79355 images/sec 2021-01-05 22:00:44,533-INFO: iter: 4600, lr: 0.000500, 'loss_cls_0': '0.061847', 'loss_loc_0': '0.042509', 'loss_cls_1': '0.030613', 'loss_loc_1': '0.035833', 'loss_cls_2': '0.012384', 'loss_loc_2': '0.009080', 'loss_rpn_cls': '0.022916', 'loss_rpn_bbox': '0.023494', 'loss': '0.259332', eta: 6:01:26, batch_cost: 0.85381 sec, ips: 4.68490 images/sec 2021-01-05 22:03:24,879-INFO: iter: 4800, lr: 0.000500, 'loss_cls_0': '0.063371', 'loss_loc_0': '0.040130', 'loss_cls_1': '0.030213', 'loss_loc_1': '0.033407', 'loss_cls_2': '0.012849', 'loss_loc_2': '0.009641', 'loss_rpn_cls': '0.022259', 'loss_rpn_bbox': '0.024018', 'loss': '0.239246', eta: 5:35:08, batch_cost: 0.79796 sec, ips: 5.01281 images/sec 2021-01-05 22:06:15,639-INFO: iter: 5000, lr: 0.000500, 'loss_cls_0': '0.051159', 'loss_loc_0': '0.036952', 'loss_cls_1': '0.026734', 'loss_loc_1': '0.032490', 'loss_cls_2': '0.011440', 'loss_loc_2': '0.010483', 'loss_rpn_cls': '0.023847', 'loss_rpn_bbox': '0.023062', 'loss': '0.228851', eta: 5:56:24, batch_cost: 0.85540 sec, ips: 4.67618 images/sec 2021-01-05 22:06:15,639-INFO: Save model to output/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms/5000. 2021-01-05 22:06:37,550-INFO: Test iter 0 2021-01-05 22:06:46,237-INFO: Test iter 100 2021-01-05 22:06:54,671-INFO: Test iter 200 2021-01-05 22:07:03,071-INFO: Test iter 300 2021-01-05 22:07:11,466-INFO: Test iter 400 2021-01-05 22:07:19,898-INFO: Test iter 500 2021-01-05 22:07:29,319-INFO: Test iter 600 2021-01-05 22:08:14,591-INFO: Test iter 700 2021-01-05 22:08:57,124-INFO: Test iter 800 2021-01-05 22:09:40,921-INFO: Test iter 900 2021-01-05 22:10:23,589-INFO: Test iter 1000 2021-01-05 22:10:26,063-INFO: Test finish iter 1007 2021-01-05 22:10:26,063-INFO: Total number of images: 1007, inference time: 4.399583983754978 fps. loading annotations into memory... Done (t=0.01s) creating index... index created! 2021-01-05 22:10:27,767-INFO: Start evaluate... Loading and preparing results... DONE (t=0.95s) creating index... index created! <string>:6: DeprecationWarning: object of type <class 'numpy.float64'> cannot be safely interpreted as an integer. Running per image evaluation... Evaluate annotation type *bbox* DONE (t=2.70s). Accumulating evaluation results... DONE (t=1.07s).Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.169Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.319Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.158Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.052Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.091Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.160Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.209Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.269Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.284Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.177Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.205Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.273 2021-01-05 22:10:32,699-INFO: Save model to output/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms/best_model. 2021-01-05 22:10:52,459-INFO: Best test box ap: 0.16946797762716523, in iter: 5000 2021-01-05 22:13:26,510-INFO: iter: 5200, lr: 0.000500, 'loss_cls_0': '0.054198', 'loss_loc_0': '0.041386', 'loss_cls_1': '0.029086', 'loss_loc_1': '0.035541', 'loss_cls_2': '0.012477', 'loss_loc_2': '0.012471', 'loss_rpn_cls': '0.019937', 'loss_rpn_bbox': '0.024147', 'loss': '0.253841', eta: 14:51:18, batch_cost: 2.15637 sec, ips: 1.85497 images/sec 2021-01-05 22:16:09,620-INFO: iter: 5400, lr: 0.000500, 'loss_cls_0': '0.065278', 'loss_loc_0': '0.044940', 'loss_cls_1': '0.032158', 'loss_loc_1': '0.040234', 'loss_cls_2': '0.013711', 'loss_loc_2': '0.013033', 'loss_rpn_cls': '0.020986', 'loss_rpn_bbox': '0.022511', 'loss': '0.261794', eta: 5:32:45, batch_cost: 0.81162 sec, ips: 4.92840 images/sec 2021-01-05 22:18:45,766-INFO: iter: 5600, lr: 0.000500, 'loss_cls_0': '0.057673', 'loss_loc_0': '0.040482', 'loss_cls_1': '0.030845', 'loss_loc_1': '0.041537', 'loss_cls_2': '0.013042', 'loss_loc_2': '0.014510', 'loss_rpn_cls': '0.018892', 'loss_rpn_bbox': '0.021275', 'loss': '0.248344', eta: 5:19:06, batch_cost: 0.78468 sec, ips: 5.09760 images/sec 2021-01-05 22:21:27,442-INFO: iter: 5800, lr: 0.000500, 'loss_cls_0': '0.062059', 'loss_loc_0': '0.043988', 'loss_cls_1': '0.034767', 'loss_loc_1': '0.046085', 'loss_cls_2': '0.014820', 'loss_loc_2': '0.016614', 'loss_rpn_cls': '0.020042', 'loss_rpn_bbox': '0.024015', 'loss': '0.276166', eta: 5:26:05, batch_cost: 0.80849 sec, ips: 4.94749 images/sec 2021-01-05 22:24:12,662-INFO: iter: 6000, lr: 0.000500, 'loss_cls_0': '0.062730', 'loss_loc_0': '0.041818', 'loss_cls_1': '0.035358', 'loss_loc_1': '0.045083', 'loss_cls_2': '0.014346', 'loss_loc_2': '0.017037', 'loss_rpn_cls': '0.019274', 'loss_rpn_bbox': '0.020919', 'loss': '0.270917', eta: 5:30:29, batch_cost: 0.82622 sec, ips: 4.84132 images/sec 2021-01-05 22:26:58,390-INFO: iter: 6200, lr: 0.000500, 'loss_cls_0': '0.060976', 'loss_loc_0': '0.044309', 'loss_cls_1': '0.034077', 'loss_loc_1': '0.045225', 'loss_cls_2': '0.015392', 'loss_loc_2': '0.018923', 'loss_rpn_cls': '0.019982', 'loss_rpn_bbox': '0.021688', 'loss': '0.274153', eta: 5:28:36, batch_cost: 0.82844 sec, ips: 4.82833 images/sec 2021-01-05 22:29:46,424-INFO: iter: 6400, lr: 0.000500, 'loss_cls_0': '0.053250', 'loss_loc_0': '0.039203', 'loss_cls_1': '0.032501', 'loss_loc_1': '0.041819', 'loss_cls_2': '0.014239', 'loss_loc_2': '0.016649', 'loss_rpn_cls': '0.019580', 'loss_rpn_bbox': '0.022341', 'loss': '0.252671', eta: 5:30:26, batch_cost: 0.84010 sec, ips: 4.76133 images/sec 2021-01-05 22:32:26,026-INFO: iter: 6600, lr: 0.000500, 'loss_cls_0': '0.063607', 'loss_loc_0': '0.043990', 'loss_cls_1': '0.037503', 'loss_loc_1': '0.050593', 'loss_cls_2': '0.016669', 'loss_loc_2': '0.022681', 'loss_rpn_cls': '0.016060', 'loss_rpn_bbox': '0.021256', 'loss': '0.286975', eta: 5:11:12, batch_cost: 0.79799 sec, ips: 5.01261 images/sec 2021-01-05 22:35:11,505-INFO: iter: 6800, lr: 0.000500, 'loss_cls_0': '0.064951', 'loss_loc_0': '0.046994', 'loss_cls_1': '0.038398', 'loss_loc_1': '0.056318', 'loss_cls_2': '0.018047', 'loss_loc_2': '0.025666', 'loss_rpn_cls': '0.016604', 'loss_rpn_bbox': '0.019919', 'loss': '0.296968', eta: 5:19:16, batch_cost: 0.82569 sec, ips: 4.84442 images/sec 2021-01-05 22:37:53,025-INFO: iter: 7000, lr: 0.000500, 'loss_cls_0': '0.073776', 'loss_loc_0': '0.041866', 'loss_cls_1': '0.033220', 'loss_loc_1': '0.039030', 'loss_cls_2': '0.016223', 'loss_loc_2': '0.014920', 'loss_rpn_cls': '0.023379', 'loss_rpn_bbox': '0.029821', 'loss': '0.302102', eta: 5:10:15, batch_cost: 0.80936 sec, ips: 4.94215 images/sec 2021-01-05 22:40:32,263-INFO: iter: 7200, lr: 0.000500, 'loss_cls_0': '0.079483', 'loss_loc_0': '0.041692', 'loss_cls_1': '0.032414', 'loss_loc_1': '0.038257', 'loss_cls_2': '0.016636', 'loss_loc_2': '0.020715', 'loss_rpn_cls': '0.020567', 'loss_rpn_bbox': '0.025319', 'loss': '0.290566', eta: 5:01:12, batch_cost: 0.79264 sec, ips: 5.04640 images/sec 2021-01-05 22:43:18,123-INFO: iter: 7400, lr: 0.000500, 'loss_cls_0': '0.062858', 'loss_loc_0': '0.043024', 'loss_cls_1': '0.033079', 'loss_loc_1': '0.045922', 'loss_cls_2': '0.016507', 'loss_loc_2': '0.021034', 'loss_rpn_cls': '0.016285', 'loss_rpn_bbox': '0.020960', 'loss': '0.286527', eta: 5:13:41, batch_cost: 0.83280 sec, ips: 4.80310 images/sec 2021-01-05 22:46:00,214-INFO: iter: 7600, lr: 0.000500, 'loss_cls_0': '0.066365', 'loss_loc_0': '0.048007', 'loss_cls_1': '0.039830', 'loss_loc_1': '0.050740', 'loss_cls_2': '0.018847', 'loss_loc_2': '0.024220', 'loss_rpn_cls': '0.018933', 'loss_rpn_bbox': '0.018931', 'loss': '0.316787', eta: 5:02:38, batch_cost: 0.81063 sec, ips: 4.93446 images/sec 2021-01-05 22:48:36,822-INFO: iter: 7800, lr: 0.000500, 'loss_cls_0': '0.063402', 'loss_loc_0': '0.043144', 'loss_cls_1': '0.036604', 'loss_loc_1': '0.050970', 'loss_cls_2': '0.017042', 'loss_loc_2': '0.025316', 'loss_rpn_cls': '0.018096', 'loss_rpn_bbox': '0.019829', 'loss': '0.291184', eta: 4:48:02, batch_cost: 0.77849 sec, ips: 5.13817 images/sec 2021-01-05 22:51:24,624-INFO: iter: 8000, lr: 0.000500, 'loss_cls_0': '0.065256', 'loss_loc_0': '0.046761', 'loss_cls_1': '0.037826', 'loss_loc_1': '0.051794', 'loss_cls_2': '0.019198', 'loss_loc_2': '0.026519', 'loss_rpn_cls': '0.017717', 'loss_rpn_bbox': '0.019827', 'loss': '0.298254', eta: 5:09:19, batch_cost: 0.84360 sec, ips: 4.74156 images/sec 2021-01-05 22:54:04,868-INFO: iter: 8200, lr: 0.000500, 'loss_cls_0': '0.064572', 'loss_loc_0': '0.045726', 'loss_cls_1': '0.040799', 'loss_loc_1': '0.060172', 'loss_cls_2': '0.020329', 'loss_loc_2': '0.033177', 'loss_rpn_cls': '0.015725', 'loss_rpn_bbox': '0.016582', 'loss': '0.308055', eta: 4:51:08, batch_cost: 0.80133 sec, ips: 4.99171 images/sec 2021-01-05 22:56:49,353-INFO: iter: 8400, lr: 0.000500, 'loss_cls_0': '0.067996', 'loss_loc_0': '0.046675', 'loss_cls_1': '0.042896', 'loss_loc_1': '0.062049', 'loss_cls_2': '0.022289', 'loss_loc_2': '0.033982', 'loss_rpn_cls': '0.014479', 'loss_rpn_bbox': '0.019303', 'loss': '0.331155', eta: 4:56:00, batch_cost: 0.82223 sec, ips: 4.86483 images/sec 2021-01-05 22:59:22,608-INFO: iter: 8600, lr: 0.000500, 'loss_cls_0': '0.062519', 'loss_loc_0': '0.044342', 'loss_cls_1': '0.037895', 'loss_loc_1': '0.058148', 'loss_cls_2': '0.020016', 'loss_loc_2': '0.033121', 'loss_rpn_cls': '0.015533', 'loss_rpn_bbox': '0.018067', 'loss': '0.309335', eta: 4:33:20, batch_cost: 0.76639 sec, ips: 5.21930 images/sec 2021-01-05 23:02:04,273-INFO: iter: 8800, lr: 0.000500, 'loss_cls_0': '0.064332', 'loss_loc_0': '0.046233', 'loss_cls_1': '0.041988', 'loss_loc_1': '0.058609', 'loss_cls_2': '0.021166', 'loss_loc_2': '0.032041', 'loss_rpn_cls': '0.014322', 'loss_rpn_bbox': '0.020238', 'loss': '0.321560', eta: 4:45:40, batch_cost: 0.80850 sec, ips: 4.94743 images/sec 2021-01-05 23:04:50,075-INFO: iter: 9000, lr: 0.000500, 'loss_cls_0': '0.062631', 'loss_loc_0': '0.045678', 'loss_cls_1': '0.040696', 'loss_loc_1': '0.062479', 'loss_cls_2': '0.020717', 'loss_loc_2': '0.033202', 'loss_rpn_cls': '0.014819', 'loss_rpn_bbox': '0.020391', 'loss': '0.306398', eta: 4:50:03, batch_cost: 0.82875 sec, ips: 4.82657 images/sec 2021-01-05 23:07:35,760-INFO: iter: 9200, lr: 0.000500, 'loss_cls_0': '0.065762', 'loss_loc_0': '0.048450', 'loss_cls_1': '0.042558', 'loss_loc_1': '0.064597', 'loss_cls_2': '0.022751', 'loss_loc_2': '0.037364', 'loss_rpn_cls': '0.014959', 'loss_rpn_bbox': '0.019045', 'loss': '0.333959', eta: 4:47:09, batch_cost: 0.82836 sec, ips: 4.82884 images/sec 2021-01-05 23:10:19,010-INFO: iter: 9400, lr: 0.000500, 'loss_cls_0': '0.067505', 'loss_loc_0': '0.049369', 'loss_cls_1': '0.044104', 'loss_loc_1': '0.068064', 'loss_cls_2': '0.023161', 'loss_loc_2': '0.037313', 'loss_rpn_cls': '0.013166', 'loss_rpn_bbox': '0.015955', 'loss': '0.344013', eta: 4:40:12, batch_cost: 0.81612 sec, ips: 4.90126 images/sec 2021-01-05 23:12:54,915-INFO: iter: 9600, lr: 0.000500, 'loss_cls_0': '0.063142', 'loss_loc_0': '0.048759', 'loss_cls_1': '0.043983', 'loss_loc_1': '0.066892', 'loss_cls_2': '0.021171', 'loss_loc_2': '0.035976', 'loss_rpn_cls': '0.013273', 'loss_rpn_bbox': '0.020711', 'loss': '0.335024', eta: 4:25:13, batch_cost: 0.78009 sec, ips: 5.12763 images/sec 2021-01-05 23:15:43,873-INFO: iter: 9800, lr: 0.000500, 'loss_cls_0': '0.062264', 'loss_loc_0': '0.046465', 'loss_cls_1': '0.040825', 'loss_loc_1': '0.061097', 'loss_cls_2': '0.021357', 'loss_loc_2': '0.034278', 'loss_rpn_cls': '0.018188', 'loss_rpn_bbox': '0.022862', 'loss': '0.337288', eta: 4:44:13, batch_cost: 0.84422 sec, ips: 4.73809 images/sec 2021-01-05 23:18:12,544-INFO: iter: 10000, lr: 0.000500, 'loss_cls_0': '0.068592', 'loss_loc_0': '0.045686', 'loss_cls_1': '0.041851', 'loss_loc_1': '0.056290', 'loss_cls_2': '0.021169', 'loss_loc_2': '0.034735', 'loss_rpn_cls': '0.024130', 'loss_rpn_bbox': '0.022144', 'loss': '0.348690', eta: 4:07:15, batch_cost: 0.74179 sec, ips: 5.39233 images/sec 2021-01-05 23:18:12,545-INFO: Save model to output/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms/10000. 2021-01-05 23:18:32,878-INFO: Test iter 0 2021-01-05 23:18:42,559-INFO: Test iter 100 2021-01-05 23:18:51,977-INFO: Test iter 200 2021-01-05 23:19:01,319-INFO: Test iter 300 2021-01-05 23:19:10,298-INFO: Test iter 400 2021-01-05 23:19:19,694-INFO: Test iter 500 2021-01-05 23:19:30,167-INFO: Test iter 600 2021-01-05 23:20:12,973-INFO: Test iter 700 2021-01-05 23:20:56,846-INFO: Test iter 800 2021-01-05 23:21:41,043-INFO: Test iter 900 2021-01-05 23:22:24,846-INFO: Test iter 1000 2021-01-05 23:22:27,424-INFO: Test finish iter 1007 2021-01-05 23:22:27,424-INFO: Total number of images: 1007, inference time: 4.291148035091142 fps. loading annotations into memory... Done (t=0.01s) creating index... index created! 2021-01-05 23:22:30,228-INFO: Start evaluate... Loading and preparing results... DONE (t=1.66s) creating index... index created! Running per image evaluation... Evaluate annotation type *bbox* DONE (t=4.36s). Accumulating evaluation results... DONE (t=1.82s).Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.182Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.349Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.179Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.109Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.122Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.153Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.215Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.283Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.293Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.249Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.235Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.247 2021-01-05 23:22:38,453-INFO: Save model to output/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms/best_model. 2021-01-05 23:23:14,508-INFO: Best test box ap: 0.18210595131337717, in iter: 10000 2021-01-05 23:25:50,456-INFO: iter: 10200, lr: 0.000500, 'loss_cls_0': '0.067151', 'loss_loc_0': '0.044671', 'loss_cls_1': '0.035398', 'loss_loc_1': '0.045706', 'loss_cls_2': '0.018710', 'loss_loc_2': '0.026910', 'loss_rpn_cls': '0.021107', 'loss_rpn_bbox': '0.023018', 'loss': '0.306695', eta: 12:35:50, batch_cost: 2.29043 sec, ips: 1.74640 images/sec 2021-01-05 23:28:37,358-INFO: iter: 10400, lr: 0.000500, 'loss_cls_0': '0.062583', 'loss_loc_0': '0.044021', 'loss_cls_1': '0.038340', 'loss_loc_1': '0.060625', 'loss_cls_2': '0.019099', 'loss_loc_2': '0.034739', 'loss_rpn_cls': '0.017975', 'loss_rpn_bbox': '0.019496', 'loss': '0.312906', eta: 4:32:48, batch_cost: 0.83514 sec, ips: 4.78961 images/sec 2021-01-05 23:31:20,758-INFO: iter: 10600, lr: 0.000500, 'loss_cls_0': '0.067678', 'loss_loc_0': '0.046534', 'loss_cls_1': '0.039705', 'loss_loc_1': '0.063273', 'loss_cls_2': '0.020378', 'loss_loc_2': '0.034880', 'loss_rpn_cls': '0.016508', 'loss_rpn_bbox': '0.019341', 'loss': '0.338241', eta: 4:24:12, batch_cost: 0.81712 sec, ips: 4.89527 images/sec 2021-01-05 23:34:14,119-INFO: iter: 10800, lr: 0.000500, 'loss_cls_0': '0.064181', 'loss_loc_0': '0.048763', 'loss_cls_1': '0.040945', 'loss_loc_1': '0.063098', 'loss_cls_2': '0.022183', 'loss_loc_2': '0.041388', 'loss_rpn_cls': '0.014780', 'loss_rpn_bbox': '0.018421', 'loss': '0.335949', eta: 4:34:48, batch_cost: 0.85880 sec, ips: 4.65767 images/sec 2021-01-05 23:36:52,551-INFO: iter: 11000, lr: 0.000500, 'loss_cls_0': '0.064664', 'loss_loc_0': '0.048685', 'loss_cls_1': '0.042219', 'loss_loc_1': '0.072353', 'loss_cls_2': '0.021658', 'loss_loc_2': '0.040372', 'loss_rpn_cls': '0.016160', 'loss_rpn_bbox': '0.016732', 'loss': '0.329565', eta: 4:13:21, batch_cost: 0.80006 sec, ips: 4.99962 images/sec 2021-01-05 23:39:39,306-INFO: iter: 11200, lr: 0.000500, 'loss_cls_0': '0.069030', 'loss_loc_0': '0.052823', 'loss_cls_1': '0.043170', 'loss_loc_1': '0.069212', 'loss_cls_2': '0.022755', 'loss_loc_2': '0.042808', 'loss_rpn_cls': '0.014426', 'loss_rpn_bbox': '0.015931', 'loss': '0.363643', eta: 4:21:04, batch_cost: 0.83319 sec, ips: 4.80080 images/sec 2021-01-05 23:42:23,095-INFO: iter: 11400, lr: 0.000500, 'loss_cls_0': '0.064709', 'loss_loc_0': '0.043366', 'loss_cls_1': '0.038690', 'loss_loc_1': '0.051932', 'loss_cls_2': '0.020410', 'loss_loc_2': '0.026836', 'loss_rpn_cls': '0.032153', 'loss_rpn_bbox': '0.026254', 'loss': '0.333374', eta: 4:14:10, batch_cost: 0.81990 sec, ips: 4.87863 images/sec 2021-01-05 23:44:59,455-INFO: iter: 11600, lr: 0.000500, 'loss_cls_0': '0.069346', 'loss_loc_0': '0.049519', 'loss_cls_1': '0.042725', 'loss_loc_1': '0.065981', 'loss_cls_2': '0.021750', 'loss_loc_2': '0.035443', 'loss_rpn_cls': '0.020221', 'loss_rpn_bbox': '0.020773', 'loss': '0.334121', eta: 3:59:38, batch_cost: 0.78146 sec, ips: 5.11865 images/sec 2021-01-05 23:47:43,903-INFO: iter: 11800, lr: 0.000500, 'loss_cls_0': '0.067058', 'loss_loc_0': '0.047519', 'loss_cls_1': '0.042720', 'loss_loc_1': '0.066260', 'loss_cls_2': '0.021873', 'loss_loc_2': '0.040856', 'loss_rpn_cls': '0.015860', 'loss_rpn_bbox': '0.020680', 'loss': '0.340207', eta: 4:09:25, batch_cost: 0.82229 sec, ips: 4.86446 images/sec 2021-01-05 23:50:37,185-INFO: iter: 12000, lr: 0.000500, 'loss_cls_0': '0.062646', 'loss_loc_0': '0.049676', 'loss_cls_1': '0.044488', 'loss_loc_1': '0.070180', 'loss_cls_2': '0.022801', 'loss_loc_2': '0.040405', 'loss_rpn_cls': '0.016876', 'loss_rpn_bbox': '0.021076', 'loss': '0.346093', eta: 4:19:48, batch_cost: 0.86605 sec, ips: 4.61869 images/sec 2021-01-05 23:53:25,579-INFO: iter: 12200, lr: 0.000500, 'loss_cls_0': '0.063689', 'loss_loc_0': '0.050207', 'loss_cls_1': '0.042780', 'loss_loc_1': '0.073127', 'loss_cls_2': '0.022725', 'loss_loc_2': '0.044891', 'loss_rpn_cls': '0.017205', 'loss_rpn_bbox': '0.019897', 'loss': '0.359676', eta: 4:09:58, batch_cost: 0.84262 sec, ips: 4.74712 images/sec 2021-01-05 23:56:05,269-INFO: iter: 12400, lr: 0.000500, 'loss_cls_0': '0.067415', 'loss_loc_0': '0.045460', 'loss_cls_1': '0.046120', 'loss_loc_1': '0.070760', 'loss_cls_2': '0.025582', 'loss_loc_2': '0.044430', 'loss_rpn_cls': '0.016760', 'loss_rpn_bbox': '0.019427', 'loss': '0.360593', eta: 3:51:16, batch_cost: 0.78842 sec, ips: 5.07343 images/sec 2021-01-05 23:58:56,848-INFO: iter: 12600, lr: 0.000500, 'loss_cls_0': '0.061238', 'loss_loc_0': '0.022800', 'loss_cls_1': '0.029173', 'loss_loc_1': '0.010960', 'loss_cls_2': '0.016174', 'loss_loc_2': '0.005824', 'loss_rpn_cls': '0.078040', 'loss_rpn_bbox': '0.035954', 'loss': '0.332935', eta: 4:11:37, batch_cost: 0.86767 sec, ips: 4.61006 images/sec 2021-01-06 00:01:35,524-INFO: iter: 12800, lr: 0.000500, 'loss_cls_0': '0.064901', 'loss_loc_0': '0.027350', 'loss_cls_1': '0.022874', 'loss_loc_1': '0.012730', 'loss_cls_2': '0.012261', 'loss_loc_2': '0.005126', 'loss_rpn_cls': '0.042158', 'loss_rpn_bbox': '0.030900', 'loss': '0.223862', eta: 3:46:49, batch_cost: 0.79125 sec, ips: 5.05532 images/sec 2021-01-06 00:04:26,768-INFO: iter: 13000, lr: 0.000500, 'loss_cls_0': '0.064326', 'loss_loc_0': '0.040274', 'loss_cls_1': '0.025074', 'loss_loc_1': '0.024208', 'loss_cls_2': '0.013382', 'loss_loc_2': '0.010759', 'loss_rpn_cls': '0.024951', 'loss_rpn_bbox': '0.022420', 'loss': '0.240248', eta: 4:03:20, batch_cost: 0.85883 sec, ips: 4.65750 images/sec 2021-01-06 00:07:10,062-INFO: iter: 13200, lr: 0.000500, 'loss_cls_0': '0.072644', 'loss_loc_0': '0.050717', 'loss_cls_1': '0.031190', 'loss_loc_1': '0.038650', 'loss_cls_2': '0.018082', 'loss_loc_2': '0.026189', 'loss_rpn_cls': '0.016587', 'loss_rpn_bbox': '0.019350', 'loss': '0.296472', eta: 3:48:29, batch_cost: 0.81602 sec, ips: 4.90182 images/sec 2021-01-06 00:09:47,111-INFO: iter: 13400, lr: 0.000500, 'loss_cls_0': '0.072115', 'loss_loc_0': '0.050292', 'loss_cls_1': '0.035664', 'loss_loc_1': '0.053717', 'loss_cls_2': '0.020093', 'loss_loc_2': '0.035742', 'loss_rpn_cls': '0.015603', 'loss_rpn_bbox': '0.020776', 'loss': '0.326792', eta: 3:37:13, batch_cost: 0.78515 sec, ips: 5.09457 images/sec 2021-01-06 00:12:34,818-INFO: iter: 13600, lr: 0.000500, 'loss_cls_0': '0.070923', 'loss_loc_0': '0.049481', 'loss_cls_1': '0.036676', 'loss_loc_1': '0.060173', 'loss_cls_2': '0.021140', 'loss_loc_2': '0.036982', 'loss_rpn_cls': '0.016696', 'loss_rpn_bbox': '0.018729', 'loss': '0.329280', eta: 3:49:18, batch_cost: 0.83893 sec, ips: 4.76799 images/sec 2021-01-06 00:15:23,952-INFO: iter: 13800, lr: 0.000500, 'loss_cls_0': '0.066262', 'loss_loc_0': '0.049082', 'loss_cls_1': '0.039771', 'loss_loc_1': '0.056706', 'loss_cls_2': '0.021089', 'loss_loc_2': '0.036825', 'loss_rpn_cls': '0.014515', 'loss_rpn_bbox': '0.018457', 'loss': '0.332186', eta: 3:48:17, batch_cost: 0.84553 sec, ips: 4.73075 images/sec 2021-01-06 00:18:08,104-INFO: iter: 14000, lr: 0.000500, 'loss_cls_0': '0.072276', 'loss_loc_0': '0.052077', 'loss_cls_1': '0.044840', 'loss_loc_1': '0.070187', 'loss_cls_2': '0.025765', 'loss_loc_2': '0.045588', 'loss_rpn_cls': '0.016084', 'loss_rpn_bbox': '0.018939', 'loss': '0.375673', eta: 3:38:54, batch_cost: 0.82090 sec, ips: 4.87269 images/sec 2021-01-06 00:21:08,665-INFO: iter: 14200, lr: 0.000500, 'loss_cls_0': '0.076326', 'loss_loc_0': '0.052754', 'loss_cls_1': '0.044431', 'loss_loc_1': '0.069688', 'loss_cls_2': '0.025825', 'loss_loc_2': '0.044177', 'loss_rpn_cls': '0.015096', 'loss_rpn_bbox': '0.017612', 'loss': '0.382994', eta: 3:57:35, batch_cost: 0.90225 sec, ips: 4.43336 images/sec 2021-01-06 00:24:06,606-INFO: iter: 14400, lr: 0.000500, 'loss_cls_0': '0.072874', 'loss_loc_0': '0.050279', 'loss_cls_1': '0.046469', 'loss_loc_1': '0.071076', 'loss_cls_2': '0.025007', 'loss_loc_2': '0.046304', 'loss_rpn_cls': '0.011439', 'loss_rpn_bbox': '0.015747', 'loss': '0.367624', eta: 3:51:26, batch_cost: 0.89013 sec, ips: 4.49373 images/sec 2021-01-06 00:26:59,892-INFO: iter: 14600, lr: 0.000500, 'loss_cls_0': '0.066495', 'loss_loc_0': '0.051622', 'loss_cls_1': '0.042681', 'loss_loc_1': '0.072888', 'loss_cls_2': '0.024463', 'loss_loc_2': '0.048027', 'loss_rpn_cls': '0.012560', 'loss_rpn_bbox': '0.018051', 'loss': '0.343523', eta: 3:42:23, batch_cost: 0.86643 sec, ips: 4.61664 images/sec 2021-01-06 00:29:51,288-INFO: iter: 14800, lr: 0.000500, 'loss_cls_0': '0.071515', 'loss_loc_0': '0.053212', 'loss_cls_1': '0.043183', 'loss_loc_1': '0.072248', 'loss_cls_2': '0.024531', 'loss_loc_2': '0.047225', 'loss_rpn_cls': '0.013225', 'loss_rpn_bbox': '0.016982', 'loss': '0.348721', eta: 3:37:07, batch_cost: 0.85709 sec, ips: 4.66697 images/sec 2021-01-06 00:32:36,621-INFO: iter: 15000, lr: 0.000500, 'loss_cls_0': '0.070571', 'loss_loc_0': '0.049246', 'loss_cls_1': '0.045589', 'loss_loc_1': '0.072676', 'loss_cls_2': '0.025107', 'loss_loc_2': '0.048486', 'loss_rpn_cls': '0.013074', 'loss_rpn_bbox': '0.015932', 'loss': '0.379738', eta: 3:26:14, batch_cost: 0.82495 sec, ips: 4.84878 images/sec 2021-01-06 00:32:36,621-INFO: Save model to output/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms/15000. 2021-01-06 00:32:56,975-INFO: Test iter 0 2021-01-06 00:33:05,701-INFO: Test iter 100 2021-01-06 00:33:14,292-INFO: Test iter 200 2021-01-06 00:33:22,707-INFO: Test iter 300 2021-01-06 00:33:31,037-INFO: Test iter 400 2021-01-06 00:33:39,479-INFO: Test iter 500 2021-01-06 00:33:49,083-INFO: Test iter 600 2021-01-06 00:34:32,669-INFO: Test iter 700 2021-01-06 00:35:16,346-INFO: Test iter 800 2021-01-06 00:36:01,073-INFO: Test iter 900 2021-01-06 00:36:46,427-INFO: Test iter 1000 2021-01-06 00:36:49,627-INFO: Test finish iter 1007 2021-01-06 00:36:49,627-INFO: Total number of images: 1007, inference time: 4.3263168012490025 fps. loading annotations into memory... Done (t=0.01s) creating index... index created! 2021-01-06 00:36:50,417-INFO: Start evaluate... Loading and preparing results... DONE (t=0.42s) creating index... index created! Running per image evaluation... Evaluate annotation type *bbox* DONE (t=1.97s). Accumulating evaluation results... DONE (t=0.66s).Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.261Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.462Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.256Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.189Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.185Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.237Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.287Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.363Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.370Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.315Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.296Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.348 2021-01-06 00:36:53,609-INFO: Save model to output/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms/best_model. 2021-01-06 00:37:29,535-INFO: Best test box ap: 0.2611445159639011, in iter: 15000 2021-01-06 00:40:04,188-INFO: iter: 15200, lr: 0.000500, 'loss_cls_0': '0.066876', 'loss_loc_0': '0.051639', 'loss_cls_1': '0.047556', 'loss_loc_1': '0.076481', 'loss_cls_2': '0.026373', 'loss_loc_2': '0.048205', 'loss_rpn_cls': '0.012475', 'loss_rpn_bbox': '0.016148', 'loss': '0.350458', eta: 9:11:52, batch_cost: 2.23731 sec, ips: 1.78786 images/sec 2021-01-06 00:42:46,245-INFO: iter: 15400, lr: 0.000500, 'loss_cls_0': '0.066060', 'loss_loc_0': '0.052031', 'loss_cls_1': '0.044632', 'loss_loc_1': '0.082567', 'loss_cls_2': '0.025446', 'loss_loc_2': '0.058891', 'loss_rpn_cls': '0.011484', 'loss_rpn_bbox': '0.015583', 'loss': '0.365930', eta: 3:17:26, batch_cost: 0.81138 sec, ips: 4.92989 images/sec 2021-01-06 00:45:37,729-INFO: iter: 15600, lr: 0.000500, 'loss_cls_0': '0.072315', 'loss_loc_0': '0.055202', 'loss_cls_1': '0.045002', 'loss_loc_1': '0.076404', 'loss_cls_2': '0.025505', 'loss_loc_2': '0.053287', 'loss_rpn_cls': '0.013081', 'loss_rpn_bbox': '0.016834', 'loss': '0.384814', eta: 3:26:00, batch_cost: 0.85834 sec, ips: 4.66018 images/sec 2021-01-06 00:48:23,624-INFO: iter: 15800, lr: 0.000500, 'loss_cls_0': '0.070336', 'loss_loc_0': '0.050973', 'loss_cls_1': '0.046747', 'loss_loc_1': '0.074327', 'loss_cls_2': '0.027804', 'loss_loc_2': '0.049670', 'loss_rpn_cls': '0.014746', 'loss_rpn_bbox': '0.017956', 'loss': '0.368038', eta: 3:16:18, batch_cost: 0.82950 sec, ips: 4.82218 images/sec 2021-01-06 00:51:11,615-INFO: iter: 16000, lr: 0.000500, 'loss_cls_0': '0.071403', 'loss_loc_0': '0.053685', 'loss_cls_1': '0.047917', 'loss_loc_1': '0.079114', 'loss_cls_2': '0.025545', 'loss_loc_2': '0.057177', 'loss_rpn_cls': '0.012593', 'loss_rpn_bbox': '0.015095', 'loss': '0.373875', eta: 3:16:01, batch_cost: 0.84012 sec, ips: 4.76123 images/sec 2021-01-06 00:54:04,151-INFO: iter: 16200, lr: 0.000500, 'loss_cls_0': '0.072183', 'loss_loc_0': '0.052932', 'loss_cls_1': '0.050209', 'loss_loc_1': '0.086099', 'loss_cls_2': '0.026967', 'loss_loc_2': '0.061982', 'loss_rpn_cls': '0.011397', 'loss_rpn_bbox': '0.017264', 'loss': '0.392682', eta: 3:18:21, batch_cost: 0.86246 sec, ips: 4.63791 images/sec 2021-01-06 00:56:49,088-INFO: iter: 16400, lr: 0.000500, 'loss_cls_0': '0.068158', 'loss_loc_0': '0.050173', 'loss_cls_1': '0.049182', 'loss_loc_1': '0.082593', 'loss_cls_2': '0.026114', 'loss_loc_2': '0.057491', 'loss_rpn_cls': '0.012618', 'loss_rpn_bbox': '0.016772', 'loss': '0.389515', eta: 3:06:54, batch_cost: 0.82457 sec, ips: 4.85099 images/sec 2021-01-06 00:59:49,505-INFO: iter: 16600, lr: 0.000500, 'loss_cls_0': '0.071238', 'loss_loc_0': '0.054461', 'loss_cls_1': '0.046717', 'loss_loc_1': '0.080054', 'loss_cls_2': '0.025314', 'loss_loc_2': '0.055334', 'loss_rpn_cls': '0.012206', 'loss_rpn_bbox': '0.016646', 'loss': '0.388371', eta: 3:21:28, batch_cost: 0.90211 sec, ips: 4.43407 images/sec 2021-01-06 01:02:51,205-INFO: iter: 16800, lr: 0.000500, 'loss_cls_0': '0.075484', 'loss_loc_0': '0.053586', 'loss_cls_1': '0.047874', 'loss_loc_1': '0.073687', 'loss_cls_2': '0.027131', 'loss_loc_2': '0.050229', 'loss_rpn_cls': '0.018113', 'loss_rpn_bbox': '0.023147', 'loss': '0.399621', eta: 3:19:57, batch_cost: 0.90888 sec, ips: 4.40101 images/sec 2021-01-06 01:05:45,816-INFO: iter: 17000, lr: 0.000500, 'loss_cls_0': '0.070221', 'loss_loc_0': '0.049896', 'loss_cls_1': '0.047129', 'loss_loc_1': '0.077793', 'loss_cls_2': '0.026747', 'loss_loc_2': '0.055582', 'loss_rpn_cls': '0.014799', 'loss_rpn_bbox': '0.017906', 'loss': '0.388288', eta: 3:09:03, batch_cost: 0.87256 sec, ips: 4.58419 images/sec 2021-01-06 01:08:47,833-INFO: iter: 17200, lr: 0.000500, 'loss_cls_0': '0.069130', 'loss_loc_0': '0.046456', 'loss_cls_1': '0.038361', 'loss_loc_1': '0.063351', 'loss_cls_2': '0.021034', 'loss_loc_2': '0.042539', 'loss_rpn_cls': '0.023127', 'loss_rpn_bbox': '0.022972', 'loss': '0.354758', eta: 3:13:27, batch_cost: 0.90684 sec, ips: 4.41090 images/sec 2021-01-06 01:11:47,992-INFO: iter: 17400, lr: 0.000500, 'loss_cls_0': '0.071986', 'loss_loc_0': '0.049988', 'loss_cls_1': '0.040766', 'loss_loc_1': '0.069570', 'loss_cls_2': '0.022108', 'loss_loc_2': '0.045157', 'loss_rpn_cls': '0.019879', 'loss_rpn_bbox': '0.021197', 'loss': '0.354816', eta: 3:09:47, batch_cost: 0.90374 sec, ips: 4.42607 images/sec 2021-01-06 01:14:39,452-INFO: iter: 17600, lr: 0.000500, 'loss_cls_0': '0.077550', 'loss_loc_0': '0.051988', 'loss_cls_1': '0.047692', 'loss_loc_1': '0.078635', 'loss_cls_2': '0.025124', 'loss_loc_2': '0.053583', 'loss_rpn_cls': '0.018500', 'loss_rpn_bbox': '0.019003', 'loss': '0.388664', eta: 2:57:18, batch_cost: 0.85794 sec, ips: 4.66232 images/sec 2021-01-06 01:17:37,300-INFO: iter: 17800, lr: 0.000500, 'loss_cls_0': '0.068284', 'loss_loc_0': '0.052294', 'loss_cls_1': '0.049155', 'loss_loc_1': '0.081252', 'loss_cls_2': '0.026631', 'loss_loc_2': '0.060193', 'loss_rpn_cls': '0.014038', 'loss_rpn_bbox': '0.018192', 'loss': '0.393549', eta: 2:59:26, batch_cost: 0.88249 sec, ips: 4.53262 images/sec 2021-01-06 01:20:39,974-INFO: iter: 18000, lr: 0.000500, 'loss_cls_0': '0.068271', 'loss_loc_0': '0.051573', 'loss_cls_1': '0.045028', 'loss_loc_1': '0.072750', 'loss_cls_2': '0.026686', 'loss_loc_2': '0.050858', 'loss_rpn_cls': '0.015933', 'loss_rpn_bbox': '0.022426', 'loss': '0.371616', eta: 3:02:44, batch_cost: 0.91372 sec, ips: 4.37772 images/sec 2021-01-06 01:23:44,231-INFO: iter: 18200, lr: 0.000500, 'loss_cls_0': '0.071103', 'loss_loc_0': '0.051483', 'loss_cls_1': '0.048356', 'loss_loc_1': '0.084477', 'loss_cls_2': '0.026666', 'loss_loc_2': '0.054075', 'loss_rpn_cls': '0.014952', 'loss_rpn_bbox': '0.019183', 'loss': '0.388800', eta: 3:01:36, batch_cost: 0.92343 sec, ips: 4.33166 images/sec 2021-01-06 01:26:35,871-INFO: iter: 18400, lr: 0.000500, 'loss_cls_0': '0.096464', 'loss_loc_0': '0.043241', 'loss_cls_1': '0.031824', 'loss_loc_1': '0.029978', 'loss_cls_2': '0.016243', 'loss_loc_2': '0.017683', 'loss_rpn_cls': '0.034154', 'loss_rpn_bbox': '0.024555', 'loss': '0.323143', eta: 2:46:43, batch_cost: 0.86236 sec, ips: 4.63845 images/sec 2021-01-06 01:29:28,909-INFO: iter: 18600, lr: 0.000500, 'loss_cls_0': '0.081579', 'loss_loc_0': '0.053368', 'loss_cls_1': '0.042914', 'loss_loc_1': '0.061944', 'loss_cls_2': '0.024832', 'loss_loc_2': '0.048310', 'loss_rpn_cls': '0.017314', 'loss_rpn_bbox': '0.017131', 'loss': '0.360574', eta: 2:44:22, batch_cost: 0.86517 sec, ips: 4.62339 images/sec 2021-01-06 01:32:22,152-INFO: iter: 18800, lr: 0.000500, 'loss_cls_0': '0.076844', 'loss_loc_0': '0.053850', 'loss_cls_1': '0.046370', 'loss_loc_1': '0.074772', 'loss_cls_2': '0.024707', 'loss_loc_2': '0.054410', 'loss_rpn_cls': '0.011939', 'loss_rpn_bbox': '0.016113', 'loss': '0.374962', eta: 2:41:45, batch_cost: 0.86658 sec, ips: 4.61583 images/sec 2021-01-06 01:35:21,893-INFO: iter: 19000, lr: 0.000500, 'loss_cls_0': '0.075113', 'loss_loc_0': '0.052251', 'loss_cls_1': '0.043262', 'loss_loc_1': '0.068422', 'loss_cls_2': '0.026201', 'loss_loc_2': '0.048790', 'loss_rpn_cls': '0.012312', 'loss_rpn_bbox': '0.016343', 'loss': '0.344422', eta: 2:44:41, batch_cost: 0.89833 sec, ips: 4.45271 images/sec 2021-01-06 01:38:31,768-INFO: iter: 19200, lr: 0.000500, 'loss_cls_0': '0.068320', 'loss_loc_0': '0.050021', 'loss_cls_1': '0.038670', 'loss_loc_1': '0.070499', 'loss_cls_2': '0.022054', 'loss_loc_2': '0.051290', 'loss_rpn_cls': '0.010597', 'loss_rpn_bbox': '0.015990', 'loss': '0.340088', eta: 2:50:55, batch_cost: 0.94955 sec, ips: 4.21252 images/sec 2021-01-06 01:41:39,515-INFO: iter: 19400, lr: 0.000500, 'loss_cls_0': '0.068600', 'loss_loc_0': '0.053063', 'loss_cls_1': '0.039759', 'loss_loc_1': '0.074037', 'loss_cls_2': '0.023851', 'loss_loc_2': '0.055952', 'loss_rpn_cls': '0.010131', 'loss_rpn_bbox': '0.016156', 'loss': '0.368476', eta: 2:45:51, batch_cost: 0.93886 sec, ips: 4.26047 images/sec 2021-01-06 01:44:30,105-INFO: iter: 19600, lr: 0.000500, 'loss_cls_0': '0.072594', 'loss_loc_0': '0.052328', 'loss_cls_1': '0.044232', 'loss_loc_1': '0.083698', 'loss_cls_2': '0.025185', 'loss_loc_2': '0.060894', 'loss_rpn_cls': '0.009301', 'loss_rpn_bbox': '0.014536', 'loss': '0.387677', eta: 2:27:40, batch_cost: 0.85193 sec, ips: 4.69521 images/sec 2021-01-06 01:47:30,750-INFO: iter: 19800, lr: 0.000500, 'loss_cls_0': '0.072003', 'loss_loc_0': '0.053069', 'loss_cls_1': '0.045005', 'loss_loc_1': '0.081133', 'loss_cls_2': '0.025797', 'loss_loc_2': '0.056887', 'loss_rpn_cls': '0.010997', 'loss_rpn_bbox': '0.015304', 'loss': '0.369631', eta: 2:33:38, batch_cost: 0.90382 sec, ips: 4.42567 images/sec 2021-01-06 01:50:27,758-INFO: iter: 20000, lr: 0.000500, 'loss_cls_0': '0.072654', 'loss_loc_0': '0.057977', 'loss_cls_1': '0.050105', 'loss_loc_1': '0.094602', 'loss_cls_2': '0.028801', 'loss_loc_2': '0.064558', 'loss_rpn_cls': '0.010585', 'loss_rpn_bbox': '0.014488', 'loss': '0.403120', eta: 2:27:31, batch_cost: 0.88511 sec, ips: 4.51919 images/sec 2021-01-06 01:50:27,758-INFO: Save model to output/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms/20000. 2021-01-06 01:50:50,078-INFO: Test iter 0 2021-01-06 01:50:59,272-INFO: Test iter 100 2021-01-06 01:51:07,717-INFO: Test iter 200 2021-01-06 01:51:15,987-INFO: Test iter 300 2021-01-06 01:51:24,364-INFO: Test iter 400 2021-01-06 01:51:32,664-INFO: Test iter 500 2021-01-06 01:51:43,373-INFO: Test iter 600 2021-01-06 01:52:31,232-INFO: Test iter 700 2021-01-06 01:53:16,968-INFO: Test iter 800 2021-01-06 01:54:02,546-INFO: Test iter 900 2021-01-06 01:54:49,043-INFO: Test iter 1000 2021-01-06 01:54:51,716-INFO: Test finish iter 1007 2021-01-06 01:54:51,716-INFO: Total number of images: 1007, inference time: 4.1655359606568245 fps. loading annotations into memory... Done (t=0.01s) creating index... index created! 2021-01-06 01:54:51,998-INFO: Start evaluate... Loading and preparing results... DONE (t=0.07s) creating index... index created! Running per image evaluation... Evaluate annotation type *bbox* DONE (t=1.40s). Accumulating evaluation results... DONE (t=0.39s).Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.254Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.485Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.227Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.209Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.198Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.263Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.276Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.365Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.371Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.337Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.315Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.368 2021-01-06 01:54:53,939-INFO: Best test box ap: 0.2611445159639011, in iter: 15000 2021-01-06 01:57:35,774-INFO: iter: 20200, lr: 0.000500, 'loss_cls_0': '0.067996', 'loss_loc_0': '0.051438', 'loss_cls_1': '0.046367', 'loss_loc_1': '0.088071', 'loss_cls_2': '0.024312', 'loss_loc_2': '0.059267', 'loss_rpn_cls': '0.011154', 'loss_rpn_bbox': '0.015090', 'loss': '0.382032', eta: 5:49:33, batch_cost: 2.14013 sec, ips: 1.86905 images/sec 2021-01-06 02:00:33,355-INFO: iter: 20400, lr: 0.000500, 'loss_cls_0': '0.069697', 'loss_loc_0': '0.047153', 'loss_cls_1': '0.047188', 'loss_loc_1': '0.077697', 'loss_cls_2': '0.025462', 'loss_loc_2': '0.057405', 'loss_rpn_cls': '0.010917', 'loss_rpn_bbox': '0.015248', 'loss': '0.380167', eta: 2:21:46, batch_cost: 0.88607 sec, ips: 4.51429 images/sec 2021-01-06 02:03:28,568-INFO: iter: 20600, lr: 0.000500, 'loss_cls_0': '0.070637', 'loss_loc_0': '0.047554', 'loss_cls_1': '0.047443', 'loss_loc_1': '0.078691', 'loss_cls_2': '0.026572', 'loss_loc_2': '0.054043', 'loss_rpn_cls': '0.011957', 'loss_rpn_bbox': '0.014491', 'loss': '0.379912', eta: 2:17:32, batch_cost: 0.87787 sec, ips: 4.55647 images/sec 2021-01-06 02:06:23,888-INFO: iter: 20800, lr: 0.000500, 'loss_cls_0': '0.076565', 'loss_loc_0': '0.048938', 'loss_cls_1': '0.046933', 'loss_loc_1': '0.075766', 'loss_cls_2': '0.026646', 'loss_loc_2': '0.045944', 'loss_rpn_cls': '0.014892', 'loss_rpn_bbox': '0.016397', 'loss': '0.384557', eta: 2:13:50, batch_cost: 0.87291 sec, ips: 4.58236 images/sec 2021-01-06 02:09:14,410-INFO: iter: 21000, lr: 0.000050, 'loss_cls_0': '0.066035', 'loss_loc_0': '0.048458', 'loss_cls_1': '0.041120', 'loss_loc_1': '0.076421', 'loss_cls_2': '0.024078', 'loss_loc_2': '0.052341', 'loss_rpn_cls': '0.015237', 'loss_rpn_bbox': '0.015883', 'loss': '0.352647', eta: 2:08:21, batch_cost: 0.85568 sec, ips: 4.67463 images/sec 2021-01-06 02:12:12,149-INFO: iter: 21200, lr: 0.000050, 'loss_cls_0': '0.062485', 'loss_loc_0': '0.046402', 'loss_cls_1': '0.040620', 'loss_loc_1': '0.077234', 'loss_cls_2': '0.024188', 'loss_loc_2': '0.060511', 'loss_rpn_cls': '0.012216', 'loss_rpn_bbox': '0.013590', 'loss': '0.346763', eta: 2:10:26, batch_cost: 0.88933 sec, ips: 4.49778 images/sec 2021-01-06 02:15:09,523-INFO: iter: 21400, lr: 0.000050, 'loss_cls_0': '0.063643', 'loss_loc_0': '0.048332', 'loss_cls_1': '0.040845', 'loss_loc_1': '0.073241', 'loss_cls_2': '0.023997', 'loss_loc_2': '0.059742', 'loss_rpn_cls': '0.011540', 'loss_rpn_bbox': '0.014379', 'loss': '0.334844', eta: 2:06:56, batch_cost: 0.88569 sec, ips: 4.51624 images/sec 2021-01-06 02:18:20,909-INFO: iter: 21600, lr: 0.000050, 'loss_cls_0': '0.061014', 'loss_loc_0': '0.047951', 'loss_cls_1': '0.040056', 'loss_loc_1': '0.083119', 'loss_cls_2': '0.023942', 'loss_loc_2': '0.062724', 'loss_rpn_cls': '0.011713', 'loss_rpn_bbox': '0.012024', 'loss': '0.355165', eta: 2:14:11, batch_cost: 0.95854 sec, ips: 4.17302 images/sec 2021-01-06 02:21:20,262-INFO: iter: 21800, lr: 0.000050, 'loss_cls_0': '0.061330', 'loss_loc_0': '0.044765', 'loss_cls_1': '0.044894', 'loss_loc_1': '0.079589', 'loss_cls_2': '0.025026', 'loss_loc_2': '0.058431', 'loss_rpn_cls': '0.010151', 'loss_rpn_bbox': '0.012448', 'loss': '0.358363', eta: 2:02:29, batch_cost: 0.89623 sec, ips: 4.46312 images/sec 2021-01-06 02:24:16,538-INFO: iter: 22000, lr: 0.000050, 'loss_cls_0': '0.055050', 'loss_loc_0': '0.042753', 'loss_cls_1': '0.041686', 'loss_loc_1': '0.074214', 'loss_cls_2': '0.022362', 'loss_loc_2': '0.055138', 'loss_rpn_cls': '0.009634', 'loss_rpn_bbox': '0.012673', 'loss': '0.328087', eta: 1:56:53, batch_cost: 0.87674 sec, ips: 4.56238 images/sec 2021-01-06 02:27:20,307-INFO: iter: 22200, lr: 0.000050, 'loss_cls_0': '0.061450', 'loss_loc_0': '0.048885', 'loss_cls_1': '0.038686', 'loss_loc_1': '0.073773', 'loss_cls_2': '0.022901', 'loss_loc_2': '0.060185', 'loss_rpn_cls': '0.009489', 'loss_rpn_bbox': '0.011620', 'loss': '0.338412', eta: 2:00:06, batch_cost: 0.92396 sec, ips: 4.32918 images/sec 2021-01-06 02:30:15,198-INFO: iter: 22400, lr: 0.000050, 'loss_cls_0': '0.058797', 'loss_loc_0': '0.044692', 'loss_cls_1': '0.040251', 'loss_loc_1': '0.076499', 'loss_cls_2': '0.022954', 'loss_loc_2': '0.059674', 'loss_rpn_cls': '0.009335', 'loss_rpn_bbox': '0.011529', 'loss': '0.340508', eta: 1:50:42, batch_cost: 0.87397 sec, ips: 4.57680 images/sec 2021-01-06 02:33:18,128-INFO: iter: 22600, lr: 0.000050, 'loss_cls_0': '0.059648', 'loss_loc_0': '0.047960', 'loss_cls_1': '0.040750', 'loss_loc_1': '0.077109', 'loss_cls_2': '0.023749', 'loss_loc_2': '0.061010', 'loss_rpn_cls': '0.009789', 'loss_rpn_bbox': '0.013983', 'loss': '0.347845', eta: 1:52:51, batch_cost: 0.91500 sec, ips: 4.37156 images/sec 2021-01-06 02:36:12,537-INFO: iter: 22800, lr: 0.000050, 'loss_cls_0': '0.053772', 'loss_loc_0': '0.045236', 'loss_cls_1': '0.036966', 'loss_loc_1': '0.073296', 'loss_cls_2': '0.021989', 'loss_loc_2': '0.060257', 'loss_rpn_cls': '0.010123', 'loss_rpn_bbox': '0.013911', 'loss': '0.321833', eta: 1:44:33, batch_cost: 0.87130 sec, ips: 4.59086 images/sec 2021-01-06 02:39:24,288-INFO: iter: 23000, lr: 0.000050, 'loss_cls_0': '0.062987', 'loss_loc_0': '0.046052', 'loss_cls_1': '0.039021', 'loss_loc_1': '0.076655', 'loss_cls_2': '0.023612', 'loss_loc_2': '0.055976', 'loss_rpn_cls': '0.012197', 'loss_rpn_bbox': '0.018384', 'loss': '0.356799', eta: 1:50:58, batch_cost: 0.95122 sec, ips: 4.20514 images/sec 2021-01-06 02:42:19,372-INFO: iter: 23200, lr: 0.000050, 'loss_cls_0': '0.256711', 'loss_loc_0': '0.047560', 'loss_cls_1': '0.028792', 'loss_loc_1': '0.022843', 'loss_cls_2': '0.016318', 'loss_loc_2': '0.010051', 'loss_rpn_cls': '0.056537', 'loss_rpn_bbox': '0.025825', 'loss': '0.506591', eta: 1:40:08, batch_cost: 0.88360 sec, ips: 4.52693 images/sec 2021-01-06 02:45:24,678-INFO: iter: 23400, lr: 0.000050, 'loss_cls_0': '0.140686', 'loss_loc_0': '0.051835', 'loss_cls_1': '0.025448', 'loss_loc_1': '0.021671', 'loss_cls_2': '0.013546', 'loss_loc_2': '0.009832', 'loss_rpn_cls': '0.033941', 'loss_rpn_bbox': '0.020519', 'loss': '0.338006', eta: 1:41:53, batch_cost: 0.92636 sec, ips: 4.31799 images/sec 2021-01-06 02:48:26,613-INFO: iter: 23600, lr: 0.000050, 'loss_cls_0': '0.124520', 'loss_loc_0': '0.060364', 'loss_cls_1': '0.024334', 'loss_loc_1': '0.025063', 'loss_cls_2': '0.013598', 'loss_loc_2': '0.013101', 'loss_rpn_cls': '0.023419', 'loss_rpn_bbox': '0.019172', 'loss': '0.307840', eta: 1:37:03, batch_cost: 0.90993 sec, ips: 4.39593 images/sec 2021-01-06 02:51:28,226-INFO: iter: 23800, lr: 0.000050, 'loss_cls_0': '0.121752', 'loss_loc_0': '0.058775', 'loss_cls_1': '0.023275', 'loss_loc_1': '0.028408', 'loss_cls_2': '0.014992', 'loss_loc_2': '0.019810', 'loss_rpn_cls': '0.018241', 'loss_rpn_bbox': '0.018349', 'loss': '0.313363', eta: 1:33:22, batch_cost: 0.90365 sec, ips: 4.42648 images/sec 2021-01-06 02:54:26,409-INFO: iter: 24000, lr: 0.000050, 'loss_cls_0': '0.110970', 'loss_loc_0': '0.064095', 'loss_cls_1': '0.024390', 'loss_loc_1': '0.031322', 'loss_cls_2': '0.015712', 'loss_loc_2': '0.022760', 'loss_rpn_cls': '0.015053', 'loss_rpn_bbox': '0.017782', 'loss': '0.331016', eta: 1:29:32, batch_cost: 0.89539 sec, ips: 4.46732 images/sec 2021-01-06 02:57:37,351-INFO: iter: 24200, lr: 0.000050, 'loss_cls_0': '0.126855', 'loss_loc_0': '0.063092', 'loss_cls_1': '0.025268', 'loss_loc_1': '0.031403', 'loss_cls_2': '0.017470', 'loss_loc_2': '0.028340', 'loss_rpn_cls': '0.013807', 'loss_rpn_bbox': '0.014591', 'loss': '0.346434', eta: 1:32:16, batch_cost: 0.95459 sec, ips: 4.19029 images/sec 2021-01-06 03:00:33,449-INFO: iter: 24400, lr: 0.000050, 'loss_cls_0': '0.102293', 'loss_loc_0': '0.066853', 'loss_cls_1': '0.026317', 'loss_loc_1': '0.035164', 'loss_cls_2': '0.018177', 'loss_loc_2': '0.032626', 'loss_rpn_cls': '0.011967', 'loss_rpn_bbox': '0.014041', 'loss': '0.325386', eta: 1:21:56, batch_cost: 0.87794 sec, ips: 4.55612 images/sec 2021-01-06 03:03:41,112-INFO: iter: 24600, lr: 0.000050, 'loss_cls_0': '0.100577', 'loss_loc_0': '0.068350', 'loss_cls_1': '0.026801', 'loss_loc_1': '0.034967', 'loss_cls_2': '0.018418', 'loss_loc_2': '0.033548', 'loss_rpn_cls': '0.013025', 'loss_rpn_bbox': '0.016182', 'loss': '0.332211', eta: 1:24:41, batch_cost: 0.94096 sec, ips: 4.25096 images/sec 2021-01-06 03:06:45,225-INFO: iter: 24800, lr: 0.000050, 'loss_cls_0': '0.088150', 'loss_loc_0': '0.064762', 'loss_cls_1': '0.026388', 'loss_loc_1': '0.035140', 'loss_cls_2': '0.019332', 'loss_loc_2': '0.036130', 'loss_rpn_cls': '0.010777', 'loss_rpn_bbox': '0.013363', 'loss': '0.326085', eta: 1:19:37, batch_cost: 0.91877 sec, ips: 4.35364 images/sec 2021-01-06 03:09:35,441-INFO: iter: 25000, lr: 0.000050, 'loss_cls_0': '0.071250', 'loss_loc_0': '0.057493', 'loss_cls_1': '0.028099', 'loss_loc_1': '0.038818', 'loss_cls_2': '0.018760', 'loss_loc_2': '0.040656', 'loss_rpn_cls': '0.009161', 'loss_rpn_bbox': '0.012405', 'loss': '0.303553', eta: 1:11:02, batch_cost: 0.85253 sec, ips: 4.69189 images/sec 2021-01-06 03:09:35,441-INFO: Save model to output/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms/25000. 2021-01-06 03:09:55,626-INFO: Test iter 0 2021-01-06 03:10:04,432-INFO: Test iter 100 2021-01-06 03:10:12,805-INFO: Test iter 200 2021-01-06 03:10:21,323-INFO: Test iter 300 2021-01-06 03:10:29,730-INFO: Test iter 400 2021-01-06 03:10:38,390-INFO: Test iter 500 2021-01-06 03:10:48,368-INFO: Test iter 600 2021-01-06 03:11:33,851-INFO: Test iter 700 2021-01-06 03:12:20,907-INFO: Test iter 800 2021-01-06 03:13:06,163-INFO: Test iter 900 2021-01-06 03:13:53,028-INFO: Test iter 1000 2021-01-06 03:13:56,224-INFO: Test finish iter 1007 2021-01-06 03:13:56,224-INFO: Total number of images: 1007, inference time: 4.183414368835553 fps. loading annotations into memory... Done (t=0.01s) creating index... index created! 2021-01-06 03:13:56,516-INFO: Start evaluate... Loading and preparing results... DONE (t=0.18s) creating index... index created! Running per image evaluation... Evaluate annotation type *bbox* DONE (t=1.35s). Accumulating evaluation results... DONE (t=0.42s).Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.287Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.497Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.296Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.238Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.210Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.285Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.315Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.396Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.402Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.361Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.339Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.389 2021-01-06 03:13:58,548-INFO: Save model to output/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms/best_model. 2021-01-06 03:14:34,654-INFO: Best test box ap: 0.28740127380447306, in iter: 25000 2021-01-06 03:17:11,639-INFO: iter: 25200, lr: 0.000050, 'loss_cls_0': '0.077958', 'loss_loc_0': '0.065383', 'loss_cls_1': '0.033764', 'loss_loc_1': '0.052715', 'loss_cls_2': '0.021716', 'loss_loc_2': '0.050493', 'loss_rpn_cls': '0.009516', 'loss_rpn_bbox': '0.012694', 'loss': '0.335838', eta: 3:02:20, batch_cost: 2.27917 sec, ips: 1.75502 images/sec 2021-01-06 03:20:18,550-INFO: iter: 25400, lr: 0.000050, 'loss_cls_0': '0.064321', 'loss_loc_0': '0.056466', 'loss_cls_1': '0.031016', 'loss_loc_1': '0.045898', 'loss_cls_2': '0.019417', 'loss_loc_2': '0.045549', 'loss_rpn_cls': '0.010067', 'loss_rpn_bbox': '0.012768', 'loss': '0.301518', eta: 1:11:48, batch_cost: 0.93663 sec, ips: 4.27065 images/sec 2021-01-06 03:23:14,542-INFO: iter: 25600, lr: 0.000050, 'loss_cls_0': '0.066452', 'loss_loc_0': '0.052236', 'loss_cls_1': '0.030797', 'loss_loc_1': '0.049198', 'loss_cls_2': '0.018396', 'loss_loc_2': '0.044125', 'loss_rpn_cls': '0.008846', 'loss_rpn_bbox': '0.013270', 'loss': '0.297378', eta: 1:04:31, batch_cost: 0.87984 sec, ips: 4.54630 images/sec 2021-01-06 03:26:16,916-INFO: iter: 25800, lr: 0.000050, 'loss_cls_0': '0.067813', 'loss_loc_0': '0.054272', 'loss_cls_1': '0.030983', 'loss_loc_1': '0.052423', 'loss_cls_2': '0.018430', 'loss_loc_2': '0.044596', 'loss_rpn_cls': '0.008323', 'loss_rpn_bbox': '0.011949', 'loss': '0.306619', eta: 1:03:49, batch_cost: 0.91176 sec, ips: 4.38710 images/sec 2021-01-06 03:29:17,814-INFO: iter: 26000, lr: 0.000050, 'loss_cls_0': '0.069245', 'loss_loc_0': '0.053925', 'loss_cls_1': '0.032926', 'loss_loc_1': '0.055293', 'loss_cls_2': '0.021025', 'loss_loc_2': '0.046128', 'loss_rpn_cls': '0.008191', 'loss_rpn_bbox': '0.012866', 'loss': '0.311819', eta: 1:00:11, batch_cost: 0.90289 sec, ips: 4.43020 images/sec 2021-01-06 03:32:10,291-INFO: iter: 26200, lr: 0.000050, 'loss_cls_0': '0.064675', 'loss_loc_0': '0.051550', 'loss_cls_1': '0.034429', 'loss_loc_1': '0.063346', 'loss_cls_2': '0.022219', 'loss_loc_2': '0.055209', 'loss_rpn_cls': '0.008815', 'loss_rpn_bbox': '0.012215', 'loss': '0.341209', eta: 0:54:43, batch_cost: 0.86412 sec, ips: 4.62896 images/sec 2021-01-06 03:35:12,410-INFO: iter: 26400, lr: 0.000050, 'loss_cls_0': '0.063754', 'loss_loc_0': '0.047027', 'loss_cls_1': '0.031622', 'loss_loc_1': '0.059396', 'loss_cls_2': '0.018286', 'loss_loc_2': '0.050568', 'loss_rpn_cls': '0.007994', 'loss_rpn_bbox': '0.011996', 'loss': '0.298561', eta: 0:54:37, batch_cost: 0.91040 sec, ips: 4.39368 images/sec 2021-01-06 03:38:10,189-INFO: iter: 26600, lr: 0.000050, 'loss_cls_0': '0.059881', 'loss_loc_0': '0.050625', 'loss_cls_1': '0.035852', 'loss_loc_1': '0.063599', 'loss_cls_2': '0.021136', 'loss_loc_2': '0.049686', 'loss_rpn_cls': '0.008624', 'loss_rpn_bbox': '0.013539', 'loss': '0.327920', eta: 0:49:49, batch_cost: 0.87930 sec, ips: 4.54910 images/sec 2021-01-06 03:41:05,971-INFO: iter: 26800, lr: 0.000050, 'loss_cls_0': '0.064217', 'loss_loc_0': '0.049913', 'loss_cls_1': '0.033870', 'loss_loc_1': '0.061922', 'loss_cls_2': '0.019468', 'loss_loc_2': '0.050536', 'loss_rpn_cls': '0.008149', 'loss_rpn_bbox': '0.012345', 'loss': '0.325155', eta: 0:47:22, batch_cost: 0.88819 sec, ips: 4.50357 images/sec 2021-01-06 03:44:26,746-INFO: iter: 27000, lr: 0.000005, 'loss_cls_0': '0.066008', 'loss_loc_0': '0.049365', 'loss_cls_1': '0.033829', 'loss_loc_1': '0.065227', 'loss_cls_2': '0.021358', 'loss_loc_2': '0.050954', 'loss_rpn_cls': '0.007439', 'loss_rpn_bbox': '0.014543', 'loss': '0.332869', eta: 0:49:48, batch_cost: 0.99613 sec, ips: 4.01553 images/sec 2021-01-06 03:47:21,968-INFO: iter: 27200, lr: 0.000005, 'loss_cls_0': '0.066441', 'loss_loc_0': '0.045323', 'loss_cls_1': '0.035435', 'loss_loc_1': '0.064383', 'loss_cls_2': '0.021325', 'loss_loc_2': '0.052081', 'loss_rpn_cls': '0.007020', 'loss_rpn_bbox': '0.012593', 'loss': '0.314068', eta: 0:41:16, batch_cost: 0.88442 sec, ips: 4.52273 images/sec 2021-01-06 03:50:22,494-INFO: iter: 27400, lr: 0.000005, 'loss_cls_0': '0.065435', 'loss_loc_0': '0.043514', 'loss_cls_1': '0.034542', 'loss_loc_1': '0.061237', 'loss_cls_2': '0.021234', 'loss_loc_2': '0.049846', 'loss_rpn_cls': '0.008164', 'loss_rpn_bbox': '0.013268', 'loss': '0.310637', eta: 0:39:02, batch_cost: 0.90096 sec, ips: 4.43970 images/sec 2021-01-06 03:53:18,987-INFO: iter: 27600, lr: 0.000005, 'loss_cls_0': '0.065536', 'loss_loc_0': '0.048888', 'loss_cls_1': '0.035785', 'loss_loc_1': '0.064248', 'loss_cls_2': '0.020574', 'loss_loc_2': '0.050526', 'loss_rpn_cls': '0.007891', 'loss_rpn_bbox': '0.012462', 'loss': '0.315502', eta: 0:35:21, batch_cost: 0.88392 sec, ips: 4.52528 images/sec 2021-01-06 03:56:23,407-INFO: iter: 27800, lr: 0.000005, 'loss_cls_0': '0.059900', 'loss_loc_0': '0.048005', 'loss_cls_1': '0.032008', 'loss_loc_1': '0.057013', 'loss_cls_2': '0.017983', 'loss_loc_2': '0.045468', 'loss_rpn_cls': '0.007717', 'loss_rpn_bbox': '0.012207', 'loss': '0.293149', eta: 0:33:49, batch_cost: 0.92230 sec, ips: 4.33699 images/sec 2021-01-06 03:59:17,031-INFO: iter: 28000, lr: 0.000005, 'loss_cls_0': '0.063061', 'loss_loc_0': '0.047308', 'loss_cls_1': '0.036785', 'loss_loc_1': '0.064119', 'loss_cls_2': '0.020220', 'loss_loc_2': '0.052347', 'loss_rpn_cls': '0.007812', 'loss_rpn_bbox': '0.010560', 'loss': '0.323311', eta: 0:28:51, batch_cost: 0.86571 sec, ips: 4.62048 images/sec 2021-01-06 04:02:24,763-INFO: iter: 28200, lr: 0.000005, 'loss_cls_0': '0.060541', 'loss_loc_0': '0.050029', 'loss_cls_1': '0.033646', 'loss_loc_1': '0.058373', 'loss_cls_2': '0.019470', 'loss_loc_2': '0.045367', 'loss_rpn_cls': '0.008726', 'loss_rpn_bbox': '0.013255', 'loss': '0.305056', eta: 0:28:13, batch_cost: 0.94104 sec, ips: 4.25060 images/sec 2021-01-06 04:05:19,099-INFO: iter: 28400, lr: 0.000005, 'loss_cls_0': '0.061692', 'loss_loc_0': '0.046110', 'loss_cls_1': '0.033161', 'loss_loc_1': '0.057631', 'loss_cls_2': '0.019117', 'loss_loc_2': '0.049664', 'loss_rpn_cls': '0.008495', 'loss_rpn_bbox': '0.010714', 'loss': '0.295149', eta: 0:23:13, batch_cost: 0.87109 sec, ips: 4.59194 images/sec 2021-01-06 04:08:14,912-INFO: iter: 28600, lr: 0.000005, 'loss_cls_0': '0.061621', 'loss_loc_0': '0.047172', 'loss_cls_1': '0.033504', 'loss_loc_1': '0.064611', 'loss_cls_2': '0.018819', 'loss_loc_2': '0.049447', 'loss_rpn_cls': '0.007575', 'loss_rpn_bbox': '0.013647', 'loss': '0.306952', eta: 0:20:31, batch_cost: 0.87949 sec, ips: 4.54808 images/sec 2021-01-06 04:11:09,393-INFO: iter: 28800, lr: 0.000005, 'loss_cls_0': '0.068867', 'loss_loc_0': '0.048824', 'loss_cls_1': '0.037000', 'loss_loc_1': '0.066814', 'loss_cls_2': '0.020599', 'loss_loc_2': '0.050471', 'loss_rpn_cls': '0.007435', 'loss_rpn_bbox': '0.011747', 'loss': '0.322600', eta: 0:17:27, batch_cost: 0.87280 sec, ips: 4.58297 images/sec 2021-01-06 04:14:08,814-INFO: iter: 29000, lr: 0.000005, 'loss_cls_0': '0.064433', 'loss_loc_0': '0.046462', 'loss_cls_1': '0.035133', 'loss_loc_1': '0.059508', 'loss_cls_2': '0.019597', 'loss_loc_2': '0.048002', 'loss_rpn_cls': '0.009046', 'loss_rpn_bbox': '0.012375', 'loss': '0.314372', eta: 0:14:56, batch_cost: 0.89672 sec, ips: 4.46070 images/sec 2021-01-06 04:17:07,957-INFO: iter: 29200, lr: 0.000005, 'loss_cls_0': '0.064912', 'loss_loc_0': '0.049047', 'loss_cls_1': '0.036925', 'loss_loc_1': '0.064885', 'loss_cls_2': '0.021077', 'loss_loc_2': '0.055657', 'loss_rpn_cls': '0.008692', 'loss_rpn_bbox': '0.012543', 'loss': '0.316626', eta: 0:11:56, batch_cost: 0.89571 sec, ips: 4.46574 images/sec 2021-01-06 04:20:11,313-INFO: iter: 29400, lr: 0.000005, 'loss_cls_0': '0.060650', 'loss_loc_0': '0.047701', 'loss_cls_1': '0.033714', 'loss_loc_1': '0.061954', 'loss_cls_2': '0.020238', 'loss_loc_2': '0.050219', 'loss_rpn_cls': '0.008307', 'loss_rpn_bbox': '0.012106', 'loss': '0.313071', eta: 0:09:10, batch_cost: 0.91675 sec, ips: 4.36323 images/sec 2021-01-06 04:22:56,240-INFO: iter: 29600, lr: 0.000005, 'loss_cls_0': '0.065240', 'loss_loc_0': '0.046613', 'loss_cls_1': '0.038239', 'loss_loc_1': '0.066724', 'loss_cls_2': '0.022705', 'loss_loc_2': '0.054745', 'loss_rpn_cls': '0.007720', 'loss_rpn_bbox': '0.012038', 'loss': '0.325763', eta: 0:05:29, batch_cost: 0.82278 sec, ips: 4.86157 images/sec 2021-01-06 04:26:01,484-INFO: iter: 29800, lr: 0.000005, 'loss_cls_0': '0.061378', 'loss_loc_0': '0.046471', 'loss_cls_1': '0.034178', 'loss_loc_1': '0.061592', 'loss_cls_2': '0.019224', 'loss_loc_2': '0.046671', 'loss_rpn_cls': '0.008376', 'loss_rpn_bbox': '0.012916', 'loss': '0.307270', eta: 0:03:05, batch_cost: 0.92832 sec, ips: 4.30885 images/sec 2021-01-06 04:29:07,247-INFO: Save model to output/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms/model_final. 2021-01-06 04:29:28,271-INFO: Test iter 0 2021-01-06 04:29:36,504-INFO: Test iter 100 2021-01-06 04:29:44,746-INFO: Test iter 200 2021-01-06 04:29:52,972-INFO: Test iter 300 2021-01-06 04:30:01,234-INFO: Test iter 400 2021-01-06 04:30:09,611-INFO: Test iter 500 2021-01-06 04:30:19,865-INFO: Test iter 600 2021-01-06 04:31:04,422-INFO: Test iter 700 2021-01-06 04:31:49,465-INFO: Test iter 800 2021-01-06 04:32:34,277-INFO: Test iter 900 2021-01-06 04:33:21,236-INFO: Test iter 1000 2021-01-06 04:33:24,177-INFO: Test finish iter 1007 2021-01-06 04:33:24,177-INFO: Total number of images: 1007, inference time: 4.266227654170191 fps. loading annotations into memory... Done (t=0.15s) creating index... index created! 2021-01-06 04:33:24,610-INFO: Start evaluate... Loading and preparing results... DONE (t=0.19s) creating index... index created! Running per image evaluation... Evaluate annotation type *bbox* DONE (t=1.24s). Accumulating evaluation results... DONE (t=0.38s).Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.291Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.498Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.289Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.243Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.218Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.302Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.315Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.397Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.403Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.370Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.320Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.420 2021-01-06 04:33:26,502-INFO: Save model to output/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms/best_model. 2021-01-06 04:34:02,993-INFO: Best test box ap: 0.2914904588209843, in iter: 29999

在可視化欄目中查看訓練過程

查看預測效果

這里有個問題，如果預測時使用中文標簽會報錯，那么繪制結果的時候如何替換默認的coco標簽？
本文的做法是直接修改PaddleDetection/ppdet/utils/coco_eval.py文件內容，這里只將映射字典替換成數字。

# 替換默認標簽字典 p9djY-1610175324787)]```python # 替換默認標簽字典 !cp ./coco_eval.py PaddleDetection/ppdet/utils/coco_eval.py # 配置文件在測試時加載標注文件，中文標注會有報錯，當然可以在前面數據清洗的時候改成拼音 !cd PaddleDetection && python -u tools/infer.py -c configs/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms.yml \--infer_img=dataset/coco/test2017/img_0043740.jpg \--output_dir=infer_output/ \--draw_threshold=0.5 \-o weights=output/cascade_rcnn_cls_aware_r101_vd_fpn_1x_softnms/model_final You are using Paddle compiled with TensorRT, but TensorRT dynamic library is not found. Ignore this if TensorRT is not needed. /home/aistudio/PaddleDetection/ppdet/core/workspace.py:118: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop workingisinstance(merge_dct[k], collections.Mapping)): /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:297: UserWarning: /home/aistudio/PaddleDetection/ppdet/modeling/backbones/fpn.py:108 The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.op_type, op_type, EXPRESSION_MAP[method_name])) /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:297: UserWarning: /home/aistudio/PaddleDetection/ppdet/modeling/roi_heads/cascade_head.py:245 The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.op_type, op_type, EXPRESSION_MAP[method_name])) /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:297: UserWarning: /home/aistudio/PaddleDetection/ppdet/modeling/roi_heads/cascade_head.py:249 The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.op_type, op_type, EXPRESSION_MAP[method_name])) /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:297: UserWarning: /home/aistudio/PaddleDetection/ppdet/modeling/roi_heads/cascade_head.py:255 The behavior of expression A / B has been unified with elementwise_div(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_div(X, Y, axis=0) instead of A / B. This transitional warning will be dropped in the future.op_type, op_type, EXPRESSION_MAP[method_name])) /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/nn.py:13362: DeprecationWarning: inspect.getargspec() is deprecated since Python 3.0, use inspect.signature() or inspect.getfullargspec()args = inspect.getargspec(self._func) W0106 06:30:28.188877 6074 device_context.cc:320] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.0, Runtime API Version: 10.1 W0106 06:30:28.194259 6074 device_context.cc:330] device: 0, cuDNN Version: 7.6. 2021-01-06 06:30:40,608-INFO: Not found annotation file None, load coco17 categories. 2021-01-06 06:30:41,781-INFO: Infer iter 0 2021-01-06 06:30:41,974-INFO: Detection bbox results save in infer_output/img_0043740.jpg

小結

什么最難？數據清洗最難——主要是這一步其實非常套路且機械，沒有多少技術含量，沒有價值創造；同樣地，當我們能夠對數據集的標注過程進行管控時，要盡可能使用通用工具，即可以直接送入PaddleDetection進行訓練。
什么最重要？數據分析最重要——其實本文的數據分析還有非常大提升空間，感興趣的讀者可以看看比賽復盤，嘗試更多的分析方法。該數據集的目標變化非常大，對模型調優提出了很大挑戰。
模型還能調優嗎？當然可以，本文只是個拋磚引玉的入門。

總結

以上是生活随笔為你收集整理的飞桨2.0 PaddleDetection：瓶装酒瑕疵检测迁移学习教程的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。

上一篇： python语言特点多模型_利用CRF模
下一篇： kakfa从入门到放弃(四): 分区和副