浅谈tensorflow中Dataset图片的批量读取及维度的操作详解

三维的读取图片（w, h, c）:

import tensorflow as tf
 
import glob
import os
 
 
def _parse_function(filename):
  # print(filename)
  image_string = tf.read_file(filename)
  image_decoded = tf.image.decode_image(image_string) # (375, 500, 3)
 
  image_resized = tf.image.resize_image_with_crop_or_pad(image_decoded, 200, 200)
  return image_resized
 
 
 
 
with tf.Session() as sess:
 
  print( sess.run( img ).shape  )

读取批量图片的读取图片（b, w, h, c）:

import tensorflow as tf
 
import glob
import os
 
'''
  Dataset 批量读取图片
'''
 
def _parse_function(filename):
  # print(filename)
  image_string = tf.read_file(filename)
  image_decoded = tf.image.decode_image(image_string) # (375, 500, 3)
 
  image_decoded = tf.expand_dims(image_decoded, axis=0)
 
  image_resized = tf.image.resize_image_with_crop_or_pad(image_decoded, 200, 200)
  return image_resized
 
 
 
img = _parse_function('../pascal/VOCdevkit/VOC2012/JPEGImages/2007_000068.jpg')
 
# image_resized = tf.image.resize_image_with_crop_or_pad( tf.truncated_normal((1,220,300,3))*10, 200, 200) 这种四维 形式是可以的
 
with tf.Session() as sess:
 
  print( sess.run( img ).shape  ) #直接初始化就可以 ，转换成四维报错误，不知道为什么，若谁想明白，请留言 报错误
  #InvalidArgumentError (see above for traceback): Input shape axis 0 must equal 4, got shape [5]

Databae的操作：

import tensorflow as tf
 
import glob
import os
 
'''
  Dataset 批量读取图片：
  
    原因：
      1. 先定义图片名的list,存放在Dataset中 from_tensor_slices（）
      2. 映射函数， 在函数中，对list中的图片进行读取，和resize,细节
        tf.read_file(filename) 返回的是三维的，因为这个每次取出一张图片，放进队列中的，不需要转化为四维
        然后对图片进行resize, 然后每个batch进行访问这个函数 ，所以get_next() 返回的是 [batch, w, h, c ]
      3. 进行shuffle , batch repeat的设置
      
      4. iterator = dataset.make_one_shot_iterator() 设置迭代器
      
      5. iterator.get_next() 获取每个batch的图片
'''
 
def _parse_function(filename):
  # print(filename)
  image_string = tf.read_file(filename)
  image_decoded = tf.image.decode_image(image_string) #(375, 500, 3)
  '''
    Tensor` with type `uint8` with shape `[height, width, num_channels]` for
     BMP, JPEG, and PNG images and shape `[num_frames, height, width, 3]` for
     GIF images.
  '''
 
  # image_resized = tf.image.resize_images(label, [200, 200])
  ''' images 三维，四维的都可以
     images: 4-D Tensor of shape `[batch, height, width, channels]` or
      3-D Tensor of shape `[height, width, channels]`.
    size: A 1-D int32 Tensor of 2 elements: `new_height, new_width`. The
       new size for the images.
  
  '''
  image_resized = tf.image.resize_image_with_crop_or_pad(image_decoded, 200, 200)
 
  # return tf.squeeze(mage_resized,axis=0)
  return image_resized
 
filenames = glob.glob( os.path.join('../pascal/VOCdevkit/VOC2012/JPEGImages', "*." + 'jpg') )
 
 
dataset = tf.data.Dataset.from_tensor_slices((filenames))
 
dataset = dataset.map(_parse_function)
 
dataset = dataset.shuffle(10).batch(2).repeat(10)
iterator = dataset.make_one_shot_iterator()
 
img = iterator.get_next()
 
with tf.Session() as sess:
  # print( sess.run(img).shape ) #(4, 200, 200, 3)
  for _ in range (10):
    print( sess.run(img).shape )