Dataset - Tencent ML-Images 腾讯多标签数据集

Author： AIHGF
发布时间：November 15, 2018
342views
No comments
18755 words
Categories：数据分析

论文：Tencent ML-Images: A Large-Scale Multi-Label Image Database for Visual Representation Learning - 2019
Github - Tencent ML-Images

腾讯开源的多标签数据集，主要包括两部分：

[1] - ML-Images - 最大的开源多标签图s片数据集，包括 17,609,752 张训练图片URLs，88, 739 张验证图片URLs，共标注了多达 11,166 个类别标签.

[2] - Resnet-101 模型 - 基于 ML-Images 的预训练模型，通过迁移学习，其在 ImageNet上的 top-1 准确率为 80.73% .

依赖环境：

Linux
Python2.7
Tensorflow >= 1.6.0

1. 数据集

图片 URLs 和对应的标注标签，数据集下载：

train_urls.txt - 下载地址1 - GoogleDrive, 下载地址2 - 百度网盘
var_urls.txt - 下载地址1 - GoogleDrive，下载地址2 - 百度网盘

train_urls.txt 的内容格式：

...
https://c4.staticflickr.com/8/7239/6997334729_e5fb3938b1_o.jpg  3:1  5193:0.9  5851:0.9 9413:1 9416:1
https://farm8.staticflickr.com/3796/19447434033_7820f80b7b_o.jpg    9290:0.8    1043:0.8
https://c7.staticflickr.com/1/42/74229536_063452179d_o.jpg    9416:1.0    9413:1.0    3:1.0
https://farm7.staticflickr.com/2378/2572916531_38735defe1_o.jpg    4063:0.9    1837:0.9    4132:1    4097:1    4089:1    1054:1    1041:1    865:1    2:1
...

每一行对应一张图片及其标注. 第一项为图片 URL，其后是由空格分割的标注标签. 例如，5193:0.9 表示类别索引5193及其置信度0.9.

类别索引由 0 开始计数，具体类别名可见：data/dictionary_and_semantic_hierarchy.txt.

1.1 图片源

ML-Images 中的图片 URLs 是根据 ImageNet 和 Open Images 收集而来的.

具体地，

[1] - Part 1 - 基于 ImageNet 整个数据库，采用了 10,706,941 张训练图片URLs 和 50,000 张验证图片 URLs，覆盖了 10, 032 个类别.

[2] - Part 2 - 基于 OpenImages 数据集，采用了额 6,902,811 张训练图片 URLs 和 38,739 张验证图片 URLs，覆盖了 1,134 个独立的类别(有些类别与 ImageNet 相似类进行了合并).

1.2 语义层次

根据 WordNet 对数据集的 11,166 个类别建立了语义层次.

每个类别的直接父类可以参见 data/dictionary_and_semantic_hierarchy.txt.

整个语义层次包含 4 个独立的树，其根节点分别为 thing, matter, object, physical objectand 和 atmospheric phenomenon. 从根节点到叶节点的最长语义路径的长度为 16，平均长度为 7.47.

1.3 标注

由于 ML-Images 中的图片 URLs 是根据 ImageNet 和 Open Images 收集而来的，因此，标注也是基于 ImageNet 和 OpenImages 的原始标注的.

具体地，

[1] - 对于 OpenImages 的 6,902,811 张训练图片URLs，移除不在保留的 1,134 个类别中的其它标注标签tags.

[2] - 根据构建的 11,166 个类别的语义层次，对 ML-Images 所有图片 URLs 进行了增强标注，即，如果一个 URL 被标注为类别 i，则其父类(ancestor categories) 也都被标注到该 URL.

[3] - 基于 OpenImages 的 6,902,811 张训练图片URLs 训练 ResNet-101 模型，输出为 1,134 个类别. 基于该 ResNet-101 模型，对 ImageNet 的10,756,941 个单标注的图片进行预测，得到其在 1,134 个类别的标签. 然后，得到 ImageNet 的 10,032 个类别和 OpenImages 的 1,134 个类别之间的归一化共存矩阵(normalized co-occurrence matrix). 例如，如果类别 i 和类别 j 是强共存的，则，如果一张图片被标注为类别 i，则类别 j 也应该被标注.

1.4 统计

# 训练图片	# 验证图片	# 类别	# 训练类别	# 平均每张图片的标签数	# 平均每个类别的图片数
17,609,752	88,739	11,166	10,505	8	1447.2

每个类别的图片数和训练集中标注数的直方图，如图：

2. 数据集多线程下载

下载测试的 tiny 数据集：train_urls_tiny.txt

cd data
./download_urls_multithreading.sh

download_urls_multithreading.sh:

python2.7 ./download_urls_multithreading.py 
            --url_list=train_urls_tiny.txt 
            --im_list=train_im_list_tiny.txt 
            --num_threads=20 --save_dir='./images'

download_urls_multithreading.py :

#!/usr/bin/env python
import os
import urllib
import argparse
import threading,signal
import time
import socket
socket.setdefaulttimeout(10.0) 

def downloadImg(start, end, url_list, save_dir):
    global record,count,count_invalid,is_exit
    im_names = []
    with open(url_list, 'r')  as url_f:
        for line in url_f.readlines()[start:end]:
            sp = line.rstrip('\n').split('\t')
            url = sp[0]
            im_name = url.split('/')[-1]
            try:
                urllib.urlretrieve(url, os.path.join(save_dir, im_name))
                record += 1
                im_file_Record.write(im_name + '\t' + '\t'.join(sp[1:]) + '\n')
                print('url = {} is finished and {} imgs have been '
                      'downloaded of all {} imgs'.format(url, record, count))
            except IOError as e:
                print ("The url:{} is ***INVALID***".format(url))
                invalid_file.write(url + '\n')
                count_invalid += 1

if __name__ == "__main__":  
    parser = argparse.ArgumentParser()
    parser.add_argument('--url_list', type=str, help='the url list file')
    parser.add_argument('--im_list', type=str, 
                        default='img.txt',help='the image list file')
    parser.add_argument('--num_threads', type=int, 
                        default=8, help='the num of processing')
    parser.add_argument('--save_dir', type=str, 
                        default='./images', help='the directory to save images')
    args = parser.parse_args()

    url_list = args.url_list
    im_list = args.im_list
    num_threads = args.num_threads
    save_dir = args.save_dir
    # create savedir
    if not os.path.isdir(save_dir):
        os.mkdir(save_dir)

    count = 0 # 图片 urls 数量
    count_invalid = 0 # 无效的图片 urls 数量
    record = 0
    with open(url_list,'r') as f:
        for line in f:
            count += 1
    part = int(count/num_threads)  
    with open(im_list, 'w') as im_file_Record,open('invalid_url.txt','w') as invalid_file:
        # 记录下载的 imgs 
        thread_list = []
        for i in range(num_threads):
            if(i == num_threads-1):
                t = threading.Thread(target = downloadImg, 
                                     kwargs={'start':i*part, 
                                             'end':count, 
                                             'url_list':url_list, 
                                             'save_dir':save_dir})
            else:
                t = threading.Thread(target = downloadImg, 
                                     kwargs={'start':i*part, 
                                             'end':(i+1)*part, 
                                             'url_list':url_list, 
                                             'save_dir':save_dir})
            t.setDaemon(True)
            # Thread.setDaemon(true)设置为守护线程
            # Thread.setDaemon(false)设置为用户线程
            thread_list.append(t)
            t.start()
        
        for i in range(num_threads):
            try:
                while thread_list[i].isAlive():
                    pass
            except KeyboardInterrupt:
                break

        if count_invalid==0:
            print ("all {} imgs have been downloaded!".format(count))
        else:
            print("{}/{} imgs have been downloaded, {} URLs are invalid".
                  format(count-count_invalid, count, count_invalid))

3. 数据集转换为 TFRecord 文件

采用多线程模块生成 tfrecords. 需要首先将 train_im_list_tiny.txt 分为多个小文件，然后保存到子目录 data/images_lists/, 如 data/images_list/train_im_list_tiny_1.txt 和 data/images_list/train_im_list_tiny_2.txt.

生成的多个 tfrecords 保存在 data/tfrecords/x.tfrecords.

cd data
./tfrecord.sh

tfrecord.sh:

echo "Generating tfrecords ..."

./tfrecord.py -idx image_lists/ 
              -tfs tfrecords/ 
              -im images/ 
              -cls 11166 
              -one True

tfrecord.py:

#!/usr/bin/python
import sys
import os
import tensorflow as tf
import numpy as np
import imghdr
import threading
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("-idx","--indexs", type=str, default="", 
                    help="dirs contains train index files")
parser.add_argument("-tfs", "--tfrecords", type=str, default="", 
                    help="dirs contains train tfrecords")
parser.add_argument("-im", "--images", type=str, default="", 
                    help="the path contains the raw images")
parser.add_argument("-cls", "--num_class", type=int, default=0, 
                    help="class label number")
parser.add_argument("-one", "--one_hot", type=bool, default=True, 
                    help="indicates the format of label fields in tfrecords")
parser.add_argument("-sidx", "--start_index", type=int, default=0, 
                    help="the start number of train tfrecord files")
args = parser.parse_args()

def _int64_feature(value):
  """
  Wrapper，用于插入 int64 特征到 Example proto.
  """
  if not isinstance(value, list):
    value = [value]
  return tf.train.Feature(int64_list=tf.train.Int64List(value=value))

def _float_feature(value):
  """
  Wrapper，用于插入 float 特征到 Example proto.
  """
  if not isinstance(value, list):
    value = [value]
  return tf.train.Feature(float_list=tf.train.FloatList(value=value))

def _bytes_feature(value):
  """
  Wrapper，用于插入 bytes 特征到 Example proto.
  """
  return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))

class ImageCoder(object):
  """
  TF 图像编码的辅助函数类.
  """
  def __init__(self):
    # 创建单个会话(Session)，以运行所有的图片编码调用.
    self._sess = tf.Session()
    
    # 初始化函数，用于转换 PNG 到 JPEG 数据.
    self._png_data = tf.placeholder(dtype=tf.string)
    image = tf.image.decode_png(self._png_data, channels=3)
    self._png_to_jpeg = tf.image.encode_jpeg(image, format='rgb', quality=100)

    # 初始化函数，用于解码 RGB JPEG 数据.
    self._decode_jpeg_data = tf.placeholder(dtype=tf.string)
    self._decode_jpeg = tf.image.decode_jpeg(self._decode_jpeg_data, channels=3)

  def png_to_jpeg(self, image_data):
    return self._sess.run(self._png_to_jpeg,
                          feed_dict={self._png_data: image_data})
  
  def decode_jpeg(self, image_data):
    image = self._sess.run(self._decode_jpeg,
                           feed_dict={self._decode_jpeg_data: image_data})
    assert len(image.shape) == 3
    assert image.shape[2] == 3
    return image

def _is_png(filename):
  return (imghdr.what(filename)=='png')

def _is_jpeg(filename):
  return (imghdr.what(filename)=='jpeg')

def _process_image(filename, coder):
  """
  单个图像文件处理.
  """
  with tf.gfile.FastGFile(filename, 'rb') as f:
    image_data = f.read()
    if not _is_jpeg(filename):
      if _is_png(filename):
        print('Converting PNG to JPEG for %s' % filename)
        image_data = coder.png_to_jpeg(image_data)
      else:
        try:
            image = coder.decode_jpeg(image_data)
            assert len(image.shape) == 3
            height = image.shape[0]
            width = image.shape[1]
            assert image.shape[2] == 3
            return image_data, height, width
        except:
            print('Cannot converted type %s' % imghdr.what(filename))
            return [], 0, 0

    image = coder.decode_jpeg(image_data)
    assert len(image.shape) == 3
    height = image.shape[0]
    width = image.shape[1]
    assert image.shape[2] == 3

    return image_data, height, width

def _save_one(train_txt, tfrecord_name, label_num, one_hot):
  writer = tf.python_io.TFRecordWriter(tfrecord_name)
  with tf.Session() as  sess:
    coder = ImageCoder()
    with open(train_txt, 'r') as lines:
      for line in lines:
        sp = line.rstrip("\n").split()
        imgf = os.path.join(args.images, sp[0])
        print(imgf)
        img, height, width = _process_image(imgf, coder)
        if height*width==0:
          continue

        if one_hot:
          label = np.zeros([label_num,], dtype=np.float32)
          for i in range(1, len(sp)):
            if len(sp[i].split(":"))==2:
              label[int(sp[i].split(":")[0])] = float(sp[i].split(":")[1])
            else:
              label[int(sp[i].split(":")[0])] = 1.0
          example = tf.train.Example(features=tf.train.Features(feature={
            'width': _int64_feature(width),
            'height': _int64_feature(height),
            'image': _bytes_feature(tf.compat.as_bytes(img)),
            'label': _bytes_feature(tf.compat.as_bytes(label.tostring())),
            'name': _bytes_feature(sp[0])
          }))
          writer.write(example.SerializeToString())
  
        else:
          label = int(sp[1])
          example = tf.train.Example(features=tf.train.Features(feature={
            'width': _int64_feature(width),
            'height': _int64_feature(height),
            'image': _bytes_feature(tf.compat.as_bytes(img)),
            'label': _int64_feature(label),
            'name': _bytes_feature(sp[0])
          }))
          writer.write(example.SerializeToString())
  writer.close()

def _save():
  files = os.listdir(args.indexs)
  coord = tf.train.Coordinator()
  threads = []

  i = args.start_index
  for idxf in files:
    threads.append( 
      threading.Thread(target=_save_one, 
                       args=(os.path.join(args.indexs, idxf), 
                             os.path.join(args.tfrecords, str(i) + ".tfrecords"), 
                             args.num_class, args.one_hot)
                      )
    )
    i = i+1

  i=0
  thread = []
  for t in threads:
    if i==32:
      for ct in thread: 
        ct.start()
      coord.join(thread)
      i = 0
      thread = [t]
    else:
      thread.append(t)
      i += 1

  for ct in thread:
    ct.start()
  coord.join(thread)

if __name__=='__main__':
  _save()

4. 训练/Finetune/分类/特征提取

TFRecords 文件处理完以后即可进行训练，源码中给出了基于 ResNet101 的模型训练与 Finetune，训练好的 ResNet101 模型，单标签图像分类，特征提取.

4.1 ResNet101 训练/Finetune

train.sh
train.py

./example/train.sh

注：这里提供的训练代码是单节点单 GPU 的.

finetune.sh
finetune.py
[ImageNet PreTrained ResNet101 模型]()

./example/finetune.sh

4.2 ResNet101 模型断点

[1] - ckpt-resnet101-mlimages - 下载地址1：GoogleDrive 下载地址2：百度网盘 - 在 ML-Images 数据训练的模型

[2] - ckpt-resnet101-mlimages-imagenet - 下载地址1：GoogleDrive 下载地址2：百度网盘 - 在 ML-Images 数据集预训练，并在 ImageNet(ILSVRC2012) 数据集 Finetune 的模型.

4.3 单标签图像分类

基于 ckpt-resnet101-mlimages-imagenet 模型，

image_classification.sh
image_classification.py

./example/image_classification.sh

./example/image_classification.sh:

python2.7 image_classification.py \
            --images=data/im_list_for_classification.txt \
            --top_k_pred=5 \
            --model_dir=checkpoints/resnet.ckpt \
            --dictionary=data/imagenet2012_dictionary.txt

image_classification.py:

#!/usr/bin/python

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import sys
import numpy as np
import cv2 as cv
import tensorflow as tf
from models import resnet as resnet
from flags import FLAGS

"""
采用保存的断点模型，进行单标签(single-label)图像分类.
"""

tf.app.flags.DEFINE_string("result", "label_pred.txt",
                           "file name to save predictions")
tf.app.flags.DEFINE_string("images", "",
                           "contains image path per line per image")
tf.app.flags.DEFINE_integer("top_k_pred", 5,
                            "the top-k predictions")
tf.app.flags.DEFINE_string("dictionary", "",
                           "the class dictionary of imagenet-2012")

def _load_dictionary(dict_file):
    dictionary = dict()
    with open(dict_file, 'r') as lines:
        for line in lines:
            sp = line.rstrip('\n').split('\t')
            idx, name = sp[0], sp[1]
            dictionary[idx] = name
    return dictionary

def preprocess(img):
    rawH = float(img.shape[0])
    rawW = float(img.shape[1])
    newH = 256.0
    newW = 256.0
    test_crop = 224.0 

    if rawH <= rawW:
        newW = (rawW/rawH) * newH
    else:
        newH = (rawH/rawW) * newW
    img = cv.resize(img, (int(newW), int(newH)))
    img = img[int((newH-test_crop)/2):int((newH-test_crop)/2)+int(test_crop),
              int((newW-test_crop)/2):int((newW-test_crop)/2)+int(test_crop)]
    img = ((img/255.0) - 0.5) * 2.0
    img = img[...,::-1]
    return img

# build model
images = tf.placeholder(dtype=tf.float32, shape=[None, 224, 224, 3])
net = resnet.ResNet(images, is_training=False)
net.build_model()

logit = net.logit
prob = tf.nn.softmax(logit)
prob_topk, pred_topk = tf.nn.top_k(prob, k=FLAGS.top_k_pred)

# restore model
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(FLAGS.visiable_gpu)
config.log_device_placement=False
sess = tf.Session(config=config)
saver = tf.train.Saver(tf.global_variables())
saver.restore(sess, FLAGS.model_dir)

dictionary = _load_dictionary(FLAGS.dictionary)

# inference
types= 'center'#'10crop'
orig_stdout = sys.stdout
f = open(FLAGS.result, 'w')
sys.stdout = f
with open(FLAGS.images, 'r') as lines:
  for line in lines:
    sp = line.rstrip('\n').split('\t')
    raw_img = cv.imread(sp[0])
    if type(raw_img)==None or raw_img.data==None :
      print("open pic " + sp[0] + " failed")
      continue
    #imgs = preprocess(raw_img, types)
    img = preprocess(raw_img)
    logits, probs_topk, preds_topk = sess.run([logit, prob_topk, pred_topk], 
                                              {images:np.expand_dims(img, axis=0)}
                                             )
    probs_topk = np.squeeze(probs_topk)
    preds_topk = np.squeeze(preds_topk)
    names_topk = [dictionary[str(i)] for i in preds_topk]
    print('+++ the predictions of {} is:'.format(sp[0]))
    for i, pred in enumerate(preds_topk):
        print('%d %s: %.3f' % (pred, names_topk[i], probs_topk[i]))    
sys.stdout = orig_stdout
f.close()

4.4 特征提取

extract_feature.sh
extract_feature.py

./example/extract_feature.sh

./example/extract_feature.sh:

#!/usr/bin/bash
set -x

PYTHON=/path/to/you/python
RESNET=101
DATA_FORMAT='NCHW'
GPUID=0
CKPT="./ckpts"

$PYTHON extract_feature.py --resnet_size=$RESNET \
                           --data_format=$DATA_FORMAT \
                           --visiable_gpu=${GPUID} \
                           --pretrain_ckpt=$CKPT \
                           --result=test.txt \
                           --images=imglist.txt

extract_feature.sh.py:

#!/usr/bin/python
"""
采用预训练的模型提取图像特征
Use pre-trained model extract image feature
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import sys
import numpy as np
import cv2 as cv
import tensorflow as tf
from models import resnet as resnet
from flags import FLAGS

tf.app.flags.DEFINE_string("result", "",
                           "file name to save features")
tf.app.flags.DEFINE_string("images", "",
                           "contains image path per line per image")

"""Crop Image To 224*224
 Args:
   img: an 3-D numpy array (H,W,C)
   type: crop method support [ center | 10crop ]
"""
def preprocess(img, type="center"):
  # resize image with smallest side to be 256
  rawH = float(img.shape[0])
  rawW = float(img.shape[1])
  newH = 256.0
  newW = 256.0
  if rawH <= rawW:
    newW = (rawW/rawH) * newH
  else:
    newH = (rawH/rawW) * newW
  img = cv.resize(img, (int(newW), int(newH)))
  imgs = None
  if type=='center':
    imgs = np.zeros((1, 224, 224, 3))
    imgs[0,...] = img[int((newH-224)/2):int((newH-224)/2)+224,
                      int((newW-224)/2):int((newW-224)/2)+224]
  elif type=='10crop':
    imgs = np.zeros((10, 224, 224, 3))
    offset = [(0, 0),
              (0, int(newW-224)),
              (int(newH-224), 0),
              (int(newH-224), int(newW-224)),
              (int((newH-224)/2), int((newW-224)/2))]
    for i in range(0, 5):
      imgs[i,...] = img[offset[i][0]:offset[i][0]+224, 
                        offset[i][1]:offset[i][1]+224]
    img = cv.flip(img, 1)
    for i in range(0, 5):
      imgs[i+5,...] = img[offset[i][0]:offset[i][0]+224, 
                          offset[i][1]:offset[i][1]+224]
  else:
    raise ValueError("Type not support")
  imgs = ((imgs/255.0) - 0.5) * 2.0
  imgs = imgs[...,::-1]
  return imgs

# build model
images = tf.placeholder(dtype=tf.float32, shape=[None, 224, 224, 3])
net = resnet.ResNet(images, is_training=False)
net.build_model()

logits = net.logit
feat = net.feat

# restore model
saver = tf.train.Saver(tf.global_variables())
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(FLAGS.visiable_gpu)
config.log_device_placement=False
sess = tf.Session(config=config)

# load trained model
saver.restore(sess, FLAGS.pretrain_ckpt)

# inference on net
types='center'
ffeat = open(FLAGS.result, 'w')
with open(FLAGS.images, 'r') as lines:
  for line in lines:
    sp = line.rstrip('\n').split(' ')
    raw_img = cv.imread(sp[0])
    if type(raw_img)==None or raw_img.data==None :
      print("open pic " + sp[0] + " failed")
      continue
    imgs = preprocess(raw_img, types)
    feats = sess.run(feat, {images:imgs})
    feats = np.squeeze(feats[0])
    if types=='10crop':
      feats = np.mean(feats, axis=0)
    print('feature-length:{}, feature={}'.format(len(feats), feats))
    ffeat.write(sp[0] + "\t" + sp[1] + "\t" + " ".
                join([str(x) for x in list(feats)]) + '\n')
ffeat.close()

5. Results

不同的 ResNet101 模型在 ImageNet(ILSVRC2012) 验证数据集上的结果如下：

Checkpoints	Train and finetune setting	Top-1 acc on Val 224	Top-5 acc on Val 224	Top-1 acc on Val 299	Top-5 accuracy on Val 299
MSRA ResNet-101	train on ImageNet	76.4	92.9	--	--
Google ResNet-101 ckpt1	train on ImageNet, 299 x 299	--	--	77.5	93.9
Our ResNet-101 ckpt1	train on ImageNet	77.8	93.9	79.0	94.5
Google ResNet-101 ckpt2	Pretrain on JFT-300M, finetune on ImageNet, 299 x 299	--	--	79.2	94.7
Our ResNet-101 ckpt2	Pretrain on ML-Images, finetune on ImageNet	78.8	94.5	79.5	94.9
Our ResNet-101 ckpt3	Pretrain on ML-Images, finetune on ImageNet 224 to 299	78.3	94.2	80.73	95.5
Our ResNet-101 ckpt4	Pretrain on ML-Images, finetune on ImageNet 299 x 299	75.8	92.7	79.6	94.6

注：

[1] - 如果没有说明，则默认训练时图像尺寸为 224x224.

[2] - 在 ImageNet 由 224 到 299 进行 finetune，表示在靠前的 epochs 的 finetune 采用 224x224，然后在后面的 epochs 采用 299x299.

[3] - Top-1 acc on Val 224 表示是 224x2224 验证图片数据集的 top-1 准确率.

Last modification：January 11th, 2019 at 02:06 pm

Dataset - Tencent ML-Images 腾讯多标签数据集

AIHGF • 2018 年 11 月 15 日

论文：Tencent ML-Images: A Large-Scale Multi-Label Image Database for Visual Representation Learning - 2019
Github - Tencent ML-Images

腾讯开源的多标签数据集，主要包括两部分：

[1] - ML-Images - 最大的开源多标签图s片数据集，包括 17,609,752 张训练图片URLs，88, 739 张验证图片URLs，共标注了多达 11,166 个类别标签.

[2] - Resnet-101 模型 - 基于 ML-Images 的预训练模型，通过迁移学习，其在 ImageNet上的 top-1 准确率为 80.73% .

依赖环境：

Linux
Python2.7
Tensorflow >= 1.6.0

1. 数据集

图片 URLs 和对应的标注标签，数据集下载：

train_urls.txt - 下载地址1 - GoogleDrive, 下载地址2 - 百度网盘
var_urls.txt - 下载地址1 - GoogleDrive，下载地址2 - 百度网盘

train_urls.txt 的内容格式：

...
https://c4.staticflickr.com/8/7239/6997334729_e5fb3938b1_o.jpg  3:1  5193:0.9  5851:0.9 9413:1 9416:1
https://farm8.staticflickr.com/3796/19447434033_7820f80b7b_o.jpg    9290:0.8    1043:0.8
https://c7.staticflickr.com/1/42/74229536_063452179d_o.jpg    9416:1.0    9413:1.0    3:1.0
https://farm7.staticflickr.com/2378/2572916531_38735defe1_o.jpg    4063:0.9    1837:0.9    4132:1    4097:1    4089:1    1054:1    1041:1    865:1    2:1
...

每一行对应一张图片及其标注. 第一项为图片 URL，其后是由空格分割的标注标签. 例如，5193:0.9 表示类别索引5193及其置信度0.9.

类别索引由 0 开始计数，具体类别名可见：data/dictionary_and_semantic_hierarchy.txt.

1.1 图片源

ML-Images 中的图片 URLs 是根据 ImageNet 和 Open Images 收集而来的.

具体地，

[1] - Part 1 - 基于 ImageNet 整个数据库，采用了 10,706,941 张训练图片URLs 和 50,000 张验证图片 URLs，覆盖了 10, 032 个类别.

1.2 语义层次

根据 WordNet 对数据集的 11,166 个类别建立了语义层次.

每个类别的直接父类可以参见 data/dictionary_and_semantic_hierarchy.txt.

1.3 标注

由于 ML-Images 中的图片 URLs 是根据 ImageNet 和 Open Images 收集而来的，因此，标注也是基于 ImageNet 和 OpenImages 的原始标注的.

具体地，

[1] - 对于 OpenImages 的 6,902,811 张训练图片URLs，移除不在保留的 1,134 个类别中的其它标注标签tags.

1.4 统计

# 训练图片	# 验证图片	# 类别	# 训练类别	# 平均每张图片的标签数	# 平均每个类别的图片数
17,609,752	88,739	11,166	10,505	8	1447.2

每个类别的图片数和训练集中标注数的直方图，如图：

2. 数据集多线程下载

下载测试的 tiny 数据集：train_urls_tiny.txt

cd data
./download_urls_multithreading.sh

download_urls_multithreading.sh:

python2.7 ./download_urls_multithreading.py 
            --url_list=train_urls_tiny.txt 
            --im_list=train_im_list_tiny.txt 
            --num_threads=20 --save_dir='./images'

download_urls_multithreading.py :

#!/usr/bin/env python
import os
import urllib
import argparse
import threading,signal
import time
import socket
socket.setdefaulttimeout(10.0) 

def downloadImg(start, end, url_list, save_dir):
    global record,count,count_invalid,is_exit
    im_names = []
    with open(url_list, 'r')  as url_f:
        for line in url_f.readlines()[start:end]:
            sp = line.rstrip('\n').split('\t')
            url = sp[0]
            im_name = url.split('/')[-1]
            try:
                urllib.urlretrieve(url, os.path.join(save_dir, im_name))
                record += 1
                im_file_Record.write(im_name + '\t' + '\t'.join(sp[1:]) + '\n')
                print('url = {} is finished and {} imgs have been '
                      'downloaded of all {} imgs'.format(url, record, count))
            except IOError as e:
                print ("The url:{} is ***INVALID***".format(url))
                invalid_file.write(url + '\n')
                count_invalid += 1

if __name__ == "__main__":  
    parser = argparse.ArgumentParser()
    parser.add_argument('--url_list', type=str, help='the url list file')
    parser.add_argument('--im_list', type=str, 
                        default='img.txt',help='the image list file')
    parser.add_argument('--num_threads', type=int, 
                        default=8, help='the num of processing')
    parser.add_argument('--save_dir', type=str, 
                        default='./images', help='the directory to save images')
    args = parser.parse_args()

    url_list = args.url_list
    im_list = args.im_list
    num_threads = args.num_threads
    save_dir = args.save_dir
    # create savedir
    if not os.path.isdir(save_dir):
        os.mkdir(save_dir)

    count = 0 # 图片 urls 数量
    count_invalid = 0 # 无效的图片 urls 数量
    record = 0
    with open(url_list,'r') as f:
        for line in f:
            count += 1
    part = int(count/num_threads)  
    with open(im_list, 'w') as im_file_Record,open('invalid_url.txt','w') as invalid_file:
        # 记录下载的 imgs 
        thread_list = []
        for i in range(num_threads):
            if(i == num_threads-1):
                t = threading.Thread(target = downloadImg, 
                                     kwargs={'start':i*part, 
                                             'end':count, 
                                             'url_list':url_list, 
                                             'save_dir':save_dir})
            else:
                t = threading.Thread(target = downloadImg, 
                                     kwargs={'start':i*part, 
                                             'end':(i+1)*part, 
                                             'url_list':url_list, 
                                             'save_dir':save_dir})
            t.setDaemon(True)
            # Thread.setDaemon(true)设置为守护线程
            # Thread.setDaemon(false)设置为用户线程
            thread_list.append(t)
            t.start()
        
        for i in range(num_threads):
            try:
                while thread_list[i].isAlive():
                    pass
            except KeyboardInterrupt:
                break

        if count_invalid==0:
            print ("all {} imgs have been downloaded!".format(count))
        else:
            print("{}/{} imgs have been downloaded, {} URLs are invalid".
                  format(count-count_invalid, count, count_invalid))

3. 数据集转换为 TFRecord 文件

生成的多个 tfrecords 保存在 data/tfrecords/x.tfrecords.

cd data
./tfrecord.sh

tfrecord.sh:

echo "Generating tfrecords ..."

./tfrecord.py -idx image_lists/ 
              -tfs tfrecords/ 
              -im images/ 
              -cls 11166 
              -one True

tfrecord.py:

#!/usr/bin/python
import sys
import os
import tensorflow as tf
import numpy as np
import imghdr
import threading
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("-idx","--indexs", type=str, default="", 
                    help="dirs contains train index files")
parser.add_argument("-tfs", "--tfrecords", type=str, default="", 
                    help="dirs contains train tfrecords")
parser.add_argument("-im", "--images", type=str, default="", 
                    help="the path contains the raw images")
parser.add_argument("-cls", "--num_class", type=int, default=0, 
                    help="class label number")
parser.add_argument("-one", "--one_hot", type=bool, default=True, 
                    help="indicates the format of label fields in tfrecords")
parser.add_argument("-sidx", "--start_index", type=int, default=0, 
                    help="the start number of train tfrecord files")
args = parser.parse_args()

def _int64_feature(value):
  """
  Wrapper，用于插入 int64 特征到 Example proto.
  """
  if not isinstance(value, list):
    value = [value]
  return tf.train.Feature(int64_list=tf.train.Int64List(value=value))

def _float_feature(value):
  """
  Wrapper，用于插入 float 特征到 Example proto.
  """
  if not isinstance(value, list):
    value = [value]
  return tf.train.Feature(float_list=tf.train.FloatList(value=value))

def _bytes_feature(value):
  """
  Wrapper，用于插入 bytes 特征到 Example proto.
  """
  return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))

class ImageCoder(object):
  """
  TF 图像编码的辅助函数类.
  """
  def __init__(self):
    # 创建单个会话(Session)，以运行所有的图片编码调用.
    self._sess = tf.Session()
    
    # 初始化函数，用于转换 PNG 到 JPEG 数据.
    self._png_data = tf.placeholder(dtype=tf.string)
    image = tf.image.decode_png(self._png_data, channels=3)
    self._png_to_jpeg = tf.image.encode_jpeg(image, format='rgb', quality=100)

    # 初始化函数，用于解码 RGB JPEG 数据.
    self._decode_jpeg_data = tf.placeholder(dtype=tf.string)
    self._decode_jpeg = tf.image.decode_jpeg(self._decode_jpeg_data, channels=3)

  def png_to_jpeg(self, image_data):
    return self._sess.run(self._png_to_jpeg,
                          feed_dict={self._png_data: image_data})
  
  def decode_jpeg(self, image_data):
    image = self._sess.run(self._decode_jpeg,
                           feed_dict={self._decode_jpeg_data: image_data})
    assert len(image.shape) == 3
    assert image.shape[2] == 3
    return image

def _is_png(filename):
  return (imghdr.what(filename)=='png')

def _is_jpeg(filename):
  return (imghdr.what(filename)=='jpeg')

def _process_image(filename, coder):
  """
  单个图像文件处理.
  """
  with tf.gfile.FastGFile(filename, 'rb') as f:
    image_data = f.read()
    if not _is_jpeg(filename):
      if _is_png(filename):
        print('Converting PNG to JPEG for %s' % filename)
        image_data = coder.png_to_jpeg(image_data)
      else:
        try:
            image = coder.decode_jpeg(image_data)
            assert len(image.shape) == 3
            height = image.shape[0]
            width = image.shape[1]
            assert image.shape[2] == 3
            return image_data, height, width
        except:
            print('Cannot converted type %s' % imghdr.what(filename))
            return [], 0, 0

    image = coder.decode_jpeg(image_data)
    assert len(image.shape) == 3
    height = image.shape[0]
    width = image.shape[1]
    assert image.shape[2] == 3

    return image_data, height, width

def _save_one(train_txt, tfrecord_name, label_num, one_hot):
  writer = tf.python_io.TFRecordWriter(tfrecord_name)
  with tf.Session() as  sess:
    coder = ImageCoder()
    with open(train_txt, 'r') as lines:
      for line in lines:
        sp = line.rstrip("\n").split()
        imgf = os.path.join(args.images, sp[0])
        print(imgf)
        img, height, width = _process_image(imgf, coder)
        if height*width==0:
          continue

        if one_hot:
          label = np.zeros([label_num,], dtype=np.float32)
          for i in range(1, len(sp)):
            if len(sp[i].split(":"))==2:
              label[int(sp[i].split(":")[0])] = float(sp[i].split(":")[1])
            else:
              label[int(sp[i].split(":")[0])] = 1.0
          example = tf.train.Example(features=tf.train.Features(feature={
            'width': _int64_feature(width),
            'height': _int64_feature(height),
            'image': _bytes_feature(tf.compat.as_bytes(img)),
            'label': _bytes_feature(tf.compat.as_bytes(label.tostring())),
            'name': _bytes_feature(sp[0])
          }))
          writer.write(example.SerializeToString())
  
        else:
          label = int(sp[1])
          example = tf.train.Example(features=tf.train.Features(feature={
            'width': _int64_feature(width),
            'height': _int64_feature(height),
            'image': _bytes_feature(tf.compat.as_bytes(img)),
            'label': _int64_feature(label),
            'name': _bytes_feature(sp[0])
          }))
          writer.write(example.SerializeToString())
  writer.close()

def _save():
  files = os.listdir(args.indexs)
  coord = tf.train.Coordinator()
  threads = []

  i = args.start_index
  for idxf in files:
    threads.append( 
      threading.Thread(target=_save_one, 
                       args=(os.path.join(args.indexs, idxf), 
                             os.path.join(args.tfrecords, str(i) + ".tfrecords"), 
                             args.num_class, args.one_hot)
                      )
    )
    i = i+1

  i=0
  thread = []
  for t in threads:
    if i==32:
      for ct in thread: 
        ct.start()
      coord.join(thread)
      i = 0
      thread = [t]
    else:
      thread.append(t)
      i += 1

  for ct in thread:
    ct.start()
  coord.join(thread)

if __name__=='__main__':
  _save()

4. 训练/Finetune/分类/特征提取

TFRecords 文件处理完以后即可进行训练，源码中给出了基于 ResNet101 的模型训练与 Finetune，训练好的 ResNet101 模型，单标签图像分类，特征提取.

4.1 ResNet101 训练/Finetune

train.sh
train.py

./example/train.sh

注：这里提供的训练代码是单节点单 GPU 的.

finetune.sh
finetune.py
[ImageNet PreTrained ResNet101 模型]()

./example/finetune.sh

4.2 ResNet101 模型断点

[1] - ckpt-resnet101-mlimages - 下载地址1：GoogleDrive 下载地址2：百度网盘 - 在 ML-Images 数据训练的模型

[2] - ckpt-resnet101-mlimages-imagenet - 下载地址1：GoogleDrive 下载地址2：百度网盘 - 在 ML-Images 数据集预训练，并在 ImageNet(ILSVRC2012) 数据集 Finetune 的模型.

4.3 单标签图像分类

基于 ckpt-resnet101-mlimages-imagenet 模型，

image_classification.sh
image_classification.py

./example/image_classification.sh

./example/image_classification.sh:

python2.7 image_classification.py \
            --images=data/im_list_for_classification.txt \
            --top_k_pred=5 \
            --model_dir=checkpoints/resnet.ckpt \
            --dictionary=data/imagenet2012_dictionary.txt

image_classification.py:

#!/usr/bin/python

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import sys
import numpy as np
import cv2 as cv
import tensorflow as tf
from models import resnet as resnet
from flags import FLAGS

"""
采用保存的断点模型，进行单标签(single-label)图像分类.
"""

tf.app.flags.DEFINE_string("result", "label_pred.txt",
                           "file name to save predictions")
tf.app.flags.DEFINE_string("images", "",
                           "contains image path per line per image")
tf.app.flags.DEFINE_integer("top_k_pred", 5,
                            "the top-k predictions")
tf.app.flags.DEFINE_string("dictionary", "",
                           "the class dictionary of imagenet-2012")

def _load_dictionary(dict_file):
    dictionary = dict()
    with open(dict_file, 'r') as lines:
        for line in lines:
            sp = line.rstrip('\n').split('\t')
            idx, name = sp[0], sp[1]
            dictionary[idx] = name
    return dictionary

def preprocess(img):
    rawH = float(img.shape[0])
    rawW = float(img.shape[1])
    newH = 256.0
    newW = 256.0
    test_crop = 224.0 

    if rawH <= rawW:
        newW = (rawW/rawH) * newH
    else:
        newH = (rawH/rawW) * newW
    img = cv.resize(img, (int(newW), int(newH)))
    img = img[int((newH-test_crop)/2):int((newH-test_crop)/2)+int(test_crop),
              int((newW-test_crop)/2):int((newW-test_crop)/2)+int(test_crop)]
    img = ((img/255.0) - 0.5) * 2.0
    img = img[...,::-1]
    return img

# build model
images = tf.placeholder(dtype=tf.float32, shape=[None, 224, 224, 3])
net = resnet.ResNet(images, is_training=False)
net.build_model()

logit = net.logit
prob = tf.nn.softmax(logit)
prob_topk, pred_topk = tf.nn.top_k(prob, k=FLAGS.top_k_pred)

# restore model
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(FLAGS.visiable_gpu)
config.log_device_placement=False
sess = tf.Session(config=config)
saver = tf.train.Saver(tf.global_variables())
saver.restore(sess, FLAGS.model_dir)

dictionary = _load_dictionary(FLAGS.dictionary)

# inference
types= 'center'#'10crop'
orig_stdout = sys.stdout
f = open(FLAGS.result, 'w')
sys.stdout = f
with open(FLAGS.images, 'r') as lines:
  for line in lines:
    sp = line.rstrip('\n').split('\t')
    raw_img = cv.imread(sp[0])
    if type(raw_img)==None or raw_img.data==None :
      print("open pic " + sp[0] + " failed")
      continue
    #imgs = preprocess(raw_img, types)
    img = preprocess(raw_img)
    logits, probs_topk, preds_topk = sess.run([logit, prob_topk, pred_topk], 
                                              {images:np.expand_dims(img, axis=0)}
                                             )
    probs_topk = np.squeeze(probs_topk)
    preds_topk = np.squeeze(preds_topk)
    names_topk = [dictionary[str(i)] for i in preds_topk]
    print('+++ the predictions of {} is:'.format(sp[0]))
    for i, pred in enumerate(preds_topk):
        print('%d %s: %.3f' % (pred, names_topk[i], probs_topk[i]))    
sys.stdout = orig_stdout
f.close()

4.4 特征提取

extract_feature.sh
extract_feature.py

./example/extract_feature.sh

./example/extract_feature.sh:

#!/usr/bin/bash
set -x

PYTHON=/path/to/you/python
RESNET=101
DATA_FORMAT='NCHW'
GPUID=0
CKPT="./ckpts"

$PYTHON extract_feature.py --resnet_size=$RESNET \
                           --data_format=$DATA_FORMAT \
                           --visiable_gpu=${GPUID} \
                           --pretrain_ckpt=$CKPT \
                           --result=test.txt \
                           --images=imglist.txt

extract_feature.sh.py:

#!/usr/bin/python
"""
采用预训练的模型提取图像特征
Use pre-trained model extract image feature
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import sys
import numpy as np
import cv2 as cv
import tensorflow as tf
from models import resnet as resnet
from flags import FLAGS

tf.app.flags.DEFINE_string("result", "",
                           "file name to save features")
tf.app.flags.DEFINE_string("images", "",
                           "contains image path per line per image")

"""Crop Image To 224*224
 Args:
   img: an 3-D numpy array (H,W,C)
   type: crop method support [ center | 10crop ]
"""
def preprocess(img, type="center"):
  # resize image with smallest side to be 256
  rawH = float(img.shape[0])
  rawW = float(img.shape[1])
  newH = 256.0
  newW = 256.0
  if rawH <= rawW:
    newW = (rawW/rawH) * newH
  else:
    newH = (rawH/rawW) * newW
  img = cv.resize(img, (int(newW), int(newH)))
  imgs = None
  if type=='center':
    imgs = np.zeros((1, 224, 224, 3))
    imgs[0,...] = img[int((newH-224)/2):int((newH-224)/2)+224,
                      int((newW-224)/2):int((newW-224)/2)+224]
  elif type=='10crop':
    imgs = np.zeros((10, 224, 224, 3))
    offset = [(0, 0),
              (0, int(newW-224)),
              (int(newH-224), 0),
              (int(newH-224), int(newW-224)),
              (int((newH-224)/2), int((newW-224)/2))]
    for i in range(0, 5):
      imgs[i,...] = img[offset[i][0]:offset[i][0]+224, 
                        offset[i][1]:offset[i][1]+224]
    img = cv.flip(img, 1)
    for i in range(0, 5):
      imgs[i+5,...] = img[offset[i][0]:offset[i][0]+224, 
                          offset[i][1]:offset[i][1]+224]
  else:
    raise ValueError("Type not support")
  imgs = ((imgs/255.0) - 0.5) * 2.0
  imgs = imgs[...,::-1]
  return imgs

# build model
images = tf.placeholder(dtype=tf.float32, shape=[None, 224, 224, 3])
net = resnet.ResNet(images, is_training=False)
net.build_model()

logits = net.logit
feat = net.feat

# restore model
saver = tf.train.Saver(tf.global_variables())
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(FLAGS.visiable_gpu)
config.log_device_placement=False
sess = tf.Session(config=config)

# load trained model
saver.restore(sess, FLAGS.pretrain_ckpt)

# inference on net
types='center'
ffeat = open(FLAGS.result, 'w')
with open(FLAGS.images, 'r') as lines:
  for line in lines:
    sp = line.rstrip('\n').split(' ')
    raw_img = cv.imread(sp[0])
    if type(raw_img)==None or raw_img.data==None :
      print("open pic " + sp[0] + " failed")
      continue
    imgs = preprocess(raw_img, types)
    feats = sess.run(feat, {images:imgs})
    feats = np.squeeze(feats[0])
    if types=='10crop':
      feats = np.mean(feats, axis=0)
    print('feature-length:{}, feature={}'.format(len(feats), feats))
    ffeat.write(sp[0] + "\t" + sp[1] + "\t" + " ".
                join([str(x) for x in list(feats)]) + '\n')
ffeat.close()

5. Results

不同的 ResNet101 模型在 ImageNet(ILSVRC2012) 验证数据集上的结果如下：

Checkpoints	Train and finetune setting	Top-1 acc on Val 224	Top-5 acc on Val 224	Top-1 acc on Val 299	Top-5 accuracy on Val 299
MSRA ResNet-101	train on ImageNet	76.4	92.9	--	--
Google ResNet-101 ckpt1	train on ImageNet, 299 x 299	--	--	77.5	93.9
Our ResNet-101 ckpt1	train on ImageNet	77.8	93.9	79.0	94.5
Google ResNet-101 ckpt2	Pretrain on JFT-300M, finetune on ImageNet, 299 x 299	--	--	79.2	94.7
Our ResNet-101 ckpt2	Pretrain on ML-Images, finetune on ImageNet	78.8	94.5	79.5	94.9
Our ResNet-101 ckpt3	Pretrain on ML-Images, finetune on ImageNet 224 to 299	78.3	94.2	80.73	95.5
Our ResNet-101 ckpt4	Pretrain on ML-Images, finetune on ImageNet 299 x 299	75.8	92.7	79.6	94.6

注：

[1] - 如果没有说明，则默认训练时图像尺寸为 224x224.

[2] - 在 ImageNet 由 224 到 299 进行 finetune，表示在靠前的 epochs 的 finetune 采用 224x224，然后在后面的 epochs 采用 299x299.

[3] - Top-1 acc on Val 224 表示是 224x2224 验证图片数据集的 top-1 准确率.

Dataset - Tencent ML-Images 腾讯多标签数据集

1. 数据集

1.1 图片源

1.2 语义层次

1.3 标注

1.4 统计

2. 数据集多线程下载

3. 数据集转换为 TFRecord 文件

4. 训练/Finetune/分类/特征提取

4.1 ResNet101 训练/Finetune

4.2 ResNet101 模型断点

4.3 单标签图像分类

4.4 特征提取

5. Results

※相关文章推荐※

※最新文章推荐※

Leave a Comment Cancel reply

Dataset - Tencent ML-Images 腾讯多标签数据集

1. 数据集

1.1 图片源

1.2 语义层次

1.3 标注

1.4 统计

2. 数据集多线程下载

3. 数据集转换为 TFRecord 文件

4. 训练/Finetune/分类/特征提取

4.1 ResNet101 训练/Finetune

4.2 ResNet101 模型断点

4.3 单标签图像分类

4.4 特征提取

5. Results