使用tfrecord建立自己的数据集
注意事项:
1.关于输入图像格式的问题
使用io.imread()的时,根据输入图像确定as_grey的参数值。 转化为字符串之后(image.tostring) ,最后输出看下image_raw的长度。因为不同的图像编码格式,存储方式不同。
我读入的灰度图jpeg格式,类型是int64,image_raw的大小是图像的大小的8倍 。 但如果是RGB图像,则统一类型是uint8。确定了类型,在之后的解码 (decode_raw)中,需要将type设置和存储方式同样的类型。
根据image_raw的长度和原图像大小,推算一下使用的类型,常用的是uint8,int32,int64.
2.转化成tfrecords的时间有点长,需要等待。
import os import tensorflow as tf import numpy as np import skimage.io as io import matplotlib.pyplot as plt import cv2 def get_data (file_path): data = [] label = [] for dirs in os.listdir(file_path): temp_path = os.path.join(file_path,dirs) i =0 for dirss in os.listdir(temp_path): data.append(os.path.join(temp_path,dirss)) num_img = len(os.listdir(temp_path)) label = np.append(label,num_img*[1]) temp = np.array([data,label]) temp = temp.transpose() np.random.shuffle(temp) image_list = list(temp[:,0]) label_list = list(temp[:,1]) label_list = [int(float(i)) for i in label_list] return image_list,label_list # 转化成字符串 def _int64_feature(value): return tf.train.Feature(int64_list=tf.train.Int64List(value=[value])) def _bytes_feature(value): return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value])) def convert_tfrecord(images,labels,save_filename): writer = tf.python_io.TFRecordWriter(save_filename) print("Transform start....") num_examples= len(labels) if np.shape(images)[0]!=num_examples: raise ValueError('Images size %d does not match label size %d.' % (images.shape[0], num_examples)) for index in np.arange(0,num_examples): try: image = io.imread(images[index],as_grey=False) #image = tf.image.decode_jpeg(images[index]) #print(image.shape) image_raw = image.tostring() #print(len(image_raw)) example = tf.train.Example(features = tf.train.Features(feature={ 'label' :_int64_feature(int(labels[index])), 'image_raw':_bytes_feature(image_raw) })) writer.write(example.SerializeToString()) except IOError as e: print('Could not read:',images[index]) print('error :%s Skip it !\n'%e) writer.close() print("success!") def read_and_decode(tfrecords_file,batch_size): reader = tf.TFRecordReader() filename_queue = tf.train.string_input_producer([tfrecords_file]) _,serialized_example = reader.read(filename_queue) features = tf.parse_single_example( serialized_example, features={ 'label': tf.FixedLenFeature([],tf.int64), 'image_raw': tf.FixedLenFeature([], tf.string) } ) #print(features['image_raw']) capacity = 1000+3*batch_size image = tf.decode_raw(features['image_raw'],tf.uint8) label = tf.cast(features['label'],tf.int32) #image = tf.image.resize_images(image,[300, 200, 1]) image = tf.reshape(image,[200,300,3]) image_batch,label_batch = tf.train.batch([image,label], batch_size=batch_size, capacity=capacity) image_batch = tf.image.resize_image_with_crop_or_pad(image_batch,100,100) image_batch = tf.cast(image_batch, tf.float32) * (1. / 255) return image_batch,label_batch def plot_images(images, labels): '''plot one batch size ''' for i in np.arange(0, 2): plt.subplot(3, 3, i + 1) plt.axis('off') # plt.title((labels[i] - 1), fontsize = 14) plt.subplots_adjust(top=1) print(labels[i]) print(images.shape) # print(images[i].shape) plt.imshow(images[i][:,:,:]) plt.show() def train(): image,label = get_data('E:\syn_data') convert_tfrecord(image,label,'1.tfrecords') x_batch, y_batch = read_and_decode('1.tfrecords', batch_size=2) with tf.Session() as sess: coord = tf.train.Coordinator() threads = tf.train.start_queue_runners(coord=coord) try: i=0 while not coord.should_stop() and i<3: # just plot one batch size image, label = sess.run([x_batch, y_batch]) plot_images(image, label) i+=1 except tf.errors.OutOfRangeError: print('done!') finally: coord.request_stop() coord.join(threads) #train()