Google Open Images Dataset V4 圖檔資料集詳解2-分類快速下載下傳

上節我們介紹了open image v4資料集的結構資訊，這節裡我們來嘗試來真正下載下傳相應的圖檔，整個資料集很大有561GB,這麼大的資料量，對于學習者，傳輸和存儲都是個問題。其實我最常用的方式是下載下傳某些(某個)對象的圖檔就夠了，根據上節我們講的關系，以對象檢測為例，我們可以寫一個腳本，單獨的擷取某些對象圖檔。這節我們講述如何快速下載下傳一個烏龜的圖像集，我們先在v4的官網上浏覽Tortoise,差不多是這樣:

我推薦的其它深度學習相關文章： 利用搜尋引擎和ffmpeg生成深度學習圖檔資料集 深度學習tensorflow object detect 之識别社會人小豬佩奇 滴滴雲GPU伺服器做深度學習之tensorflow object detect 對象檢測 實踐 如何利用YOLO訓練自己的模型 YOLO圖像快速标注工具 視訊教程：

tensorflow object detection api 檢測自己的資料集

YOLO實戰

一、安裝tensorflow object detect Api

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

#在根目錄下建立一個output目錄 mkdir /output cd /output/ #下載下傳舊版本的tensorflow model（object api 包含在model裡）,最新版本的api存在問題（目前2018.4.20） wget  https: //github .com /tensorflow/models/archive/dcfe009a024854207c9067d785c105f5ebf5a01b .zip unzip dcfe009a024854207c9067d785c105f5ebf5a01b.zip  mv models-dcfe009a024854207c9067d785c105f5ebf5a01b models rm dcfe009a024854207c9067d785c105f5ebf5a01b.zip  #安裝依賴項 pip  install Cython pip  install pillow pip  install lxml pip  install jupyter pip  install matplotlib pip  install opencv-python pip  install pycocotools #安裝object detection api 并驗證 cd /output/models/research/ protoc object_detection /protos/ *.proto --python_out=. export PYTHONPATH=$PYTHONPATH:` pwd `:` pwd ` /slim python object_detection /builders/model_builder_test .py

下載下傳代碼github

二、根據關鍵字生成tfrecord

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170

import pandas as pd import numpy as np import os import tensorflow as tf import io import logging import random import sys import PIL.Image import hashlib from urllib  import request sys.path.append( "/output/models/research/" ) from object_detection.utils  import dataset_util class open_image_dataset: classname_csv = "https://storage.googleapis.com/openimages/2018_04/class-descriptions-boxable.csv" test_image_csv = "https://storage.googleapis.com/openimages/2018_04/test/test-images-with-rotation.csv" test_box_csv = "https://storage.googleapis.com/openimages/2018_04/test/test-annotations-bbox.csv" val_image_csv = "https://storage.googleapis.com/openimages/2018_04/validation/validation-images-with-rotation.csv" val_box_csv = "https://storage.googleapis.com/openimages/2018_04/validation/validation-annotations-bbox.csv" train_image_csv = "https://storage.googleapis.com/openimages/2018_04/train/train-images-boxable-with-rotation.csv" train_box_csv = "https://storage.googleapis.com/openimages/2018_04/train/train-annotations-bbox.csv" def download_test( self ): print ( "start download test info" ) folder = "test" if os.path.exists(folder,)  is False : os.makedirs(folder) image_csv_path = folder + "/image.csv" box_csv_path = folder + "/box.csv" classname_csv_path = folder + "/classname.csv" if os.path.exists(image_csv_path)  is False : request.urlretrieve( self .test_image_csv,image_csv_path) if os.path.exists(box_csv_path)  is False : request.urlretrieve( self .test_box_csv,box_csv_path ) if os.path.exists(classname_csv_path)  is False : request.urlretrieve( self .classname_csv,classname_csv_path ) print ( "download test complete" ) def download_val( self ): folder = "val" if os.path.exists(folder,)  is False : os.makedirs(folder) image_csv_path = folder + "/image.csv" box_csv_path = folder + "/box.csv" classname_csv_path = folder + "/classname.csv" if os.path.exists(image_csv_path)  is False : request.urlretrieve( self .val_image_csv,image_csv_path) if os.path.exists(box_csv_path)  is False : request.urlretrieve( self .val_box_csv,box_csv_path ) if os.path.exists(classname_csv_path)  is False : request.urlretrieve( self .classname_csv,classname_csv_path) print ( "download val complete" ) def download_train( self ): folder = "train" if os.path.exists(folder,)  is False : os.makedirs(folder) image_csv_path = folder + "/image.csv" box_csv_path = folder + "/box.csv" classname_csv_path = folder + "/classname.csv" if os.path.exists(image_csv_path)  is False : request.urlretrieve( self .train_image_csv,image_csv_path) if os.path.exists(box_csv_path)  is False : request.urlretrieve( self .train_box_csv,box_csv_path ) if os.path.exists(classname_csv_path)  is False : request.urlretrieve( self .classname_csv,classname_csv_path ) print ( "download train complete" ) def create_tfrecord( self ,folder,keywords):   image_csv_path = folder + "/image.csv" box_csv_path = folder + "/box.csv" classname_csv_path = folder + "/classname.csv" df_image  = pd.read_csv(image_csv_path) df_box  = pd.read_csv(box_csv_path) df_classname  = pd.read_csv(classname_csv_path,names = [ 'labelID' , 'LabelName' ]) data = df_classname[df_classname[ 'LabelName' ] = = keywords] data = pd.merge(data, df_box, left_on  = 'labelID' , right_on  = 'LabelName' , how = 'right' ) data = pd.merge(data, df_image, left_on  = 'ImageID' , right_on  = 'ImageID' , how = 'right' ) data = data[data[ 'labelID' ].notna() & data[ 'ImageID' ].notna()] folder_path = keywords + "/" + folder + "/" if os.path.exists(folder_path)  is False : os.makedirs(folder_path) tfrecord_file = folder_path + keywords + ".tfrecord" writer  = tf.python_io.TFRecordWriter(tfrecord_file) for index,row  in data.iterrows(): file_name = row[ 'ImageID' ] + ".jpg" file_path = folder_path + file_name if os.path.exists(file_path)  is False : request.urlretrieve(row[ 'OriginalURL' ],file_path)         with tf.gfile.GFile(file_path,  'rb' ) as fid: encoded_jpg  = fid.read() encoded_jpg_io  = io.BytesIO(encoded_jpg) image  = PIL.Image. open (encoded_jpg_io) if image. format ! = 'JPEG' : print ( "file format error " + file_path) os.remove(file_path) continue image.close()   key  = hashlib.sha256(encoded_jpg).hexdigest()     xmin  = [] ymin  = [] xmax  = [] ymax  = [] classes  = [] classes_text  = [] width = image.width height = image.height xmin.append( float (row[ 'XMin' ])) xmax.append( float (row[ 'XMax' ])) ymin.append( float (row[ 'YMin' ])) ymax.append( float (row[ 'YMax' ])) classes.append( int ( 1 )) classes_text.append(keywords.encode( 'utf8' )) example  = tf.train.Example(features = tf.train.Features(feature = { 'image/height' : dataset_util.int64_feature( int (height)), 'image/width' : dataset_util.int64_feature( int (width)), 'image/filename' : dataset_util.bytes_feature(file_name.encode( 'utf8' )), 'image/source_id' : dataset_util.bytes_feature(file_name.encode( 'utf8' )), 'image/key/sha256' : dataset_util.bytes_feature(key.encode( 'utf8' )), 'image/encoded' : dataset_util.bytes_feature(encoded_jpg), 'image/format' : dataset_util.bytes_feature( 'jpeg' .encode( 'utf8' )), 'image/object/bbox/xmin' : dataset_util.float_list_feature(xmin), 'image/object/bbox/xmax' : dataset_util.float_list_feature(xmax), 'image/object/bbox/ymin' : dataset_util.float_list_feature(ymin), 'image/object/bbox/ymax' : dataset_util.float_list_feature(ymax), 'image/object/class/text' : dataset_util.bytes_list_feature(classes_text), 'image/object/class/label' : dataset_util.int64_list_feature(classes), })) writer.write(example.SerializeToString()) os.remove(file_path) print ( "file " + file_path) writer.close()  print ( "create " + tfrecord_file + " success!" ) def create_train_tfrecord( self ,keywords):   self .download_train() self .create_tfrecord( "train" ,keywords) def create_val_tfrecord( self ,keywords):   self .download_val() self .create_tfrecord( "val" ,keywords)  def create_test_tfrecord( self ,keywords):   self .download_test() self .create_tfrecord( "test" ,keywords) def create_all_tfrecord( self ,keywords): self .create_train_tfrecord(keywords) self .create_val_tfrecord(keywords) dataset = open_image_dataset() dataset.download_test() dataset.create_tfrecord( "test" , "Tortoise" ) #下載下傳關鍵字為"Tortoise"的測試資料集 #dataset.download_val() #dataset.create_tfrecord("val","Tortoise")#下載下傳關鍵字為"Tortoise"的驗證資料集 #dataset.download_train() #dataset.create_tfrecord("train","Tortoise")#下載下傳關鍵字為"Tortoise"的訓練資料集 # dataset.create_all_tfrecord("Tortoise") #下載下傳所有關鍵字為"Tortoise"的資料集

三、對生成的tfrecord進行驗證

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

import tensorflow as tf import numpy as np import os import skimage.io as io import cv2 tfrecords_filename  = "Tortoise/test/Tortoise.tfrecord" filename_queue  = tf.train.string_input_producer([tfrecords_filename])  reader  = tf.TFRecordReader() _, serialized_example  = reader.read(filename_queue)  features  = tf.parse_single_example(serialized_example, features = { 'image/width' :tf.FixedLenFeature([], tf.int64), 'image/height' : tf.FixedLenFeature([], tf.int64), 'image/filename' :  tf.FixedLenFeature([], tf.string), 'image/source_id' : tf.FixedLenFeature([], tf.string), 'image/key/sha256' :  tf.FixedLenFeature([], tf.string), 'image/encoded' : tf.FixedLenFeature([], tf.string), 'image/format' :  tf.FixedLenFeature([], tf.string), 'image/object/bbox/xmin' : tf.FixedLenFeature([], tf.float32), 'image/object/bbox/xmax' : tf.FixedLenFeature([], tf.float32), 'image/object/bbox/ymin' :tf.FixedLenFeature([], tf.float32), 'image/object/bbox/ymax' :tf.FixedLenFeature([], tf.float32), 'image/object/class/text' :tf.FixedLenFeature([], tf.string), 'image/object/class/label' : tf.FixedLenFeature([], tf.int64), })   width = tf.cast(features[ 'image/width' ], tf.int32) height  = tf.cast(features[ 'image/height' ], tf.int32) filename  = tf.cast(features[ 'image/filename' ], tf.string) format = tf.cast(features[ 'image/format' ], tf.string) xmin  = tf.cast(features[ 'image/object/bbox/xmin' ], tf.float32) xmax  = tf.cast(features[ 'image/object/bbox/xmax' ], tf.float32) ymin  = tf.cast(features[ 'image/object/bbox/ymin' ], tf.float32) ymax  = tf.cast(features[ 'image/object/bbox/ymax' ], tf.float32) text  = tf.cast(features[ 'image/object/class/text' ], tf.string) label  = tf.cast(features[ 'image/object/class/label' ], tf.int64) image  = tf.image.decode_jpeg(features[ 'image/encoded' ]); image  = tf.reshape(image,tf.stack([height,width, 3 ])) with tf.Session() as sess:  init_op  = tf.initialize_all_variables() sess.run(init_op) coord = tf.train.Coordinator() threads = tf.train.start_queue_runners(coord = coord) for i  in range ( 20 ): width1,height1,filename1,format1,xmin1,xmax1,ymin1,ymax1,text1,label1,image1 = sess.run([width,height,filename, format ,xmin,xmax,ymin,ymax,text,label,image]) print (width1,height1,filename1,format1,xmin1,xmax1,ymin1,ymax1,text1,label1) x1,y1 = int (xmin1 * width1), int (ymin1 * height1) x2,y2 = int (xmax1 * width1), int (ymax1 * height1) io.imshow(cv2.rectangle(np.array(image1),(x1,y1),(x2,y2),( , 255 , ), 3 ), cmap  = 'gray' , interpolation  = 'bicubic' ) io.show() coord.request_stop() coord.join(threads)

下載下傳代碼github

最終的結果如下：