An introduction to the batch reading of Dataset images in tensorflow and the operation of dimensions in detail

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

import tensorflow as tf
 
import glob
import os
 
'''
  Dataset Batch Read Images
'''
 
def _parse_function(filename):
  # print(filename)
  image_string = tf.read_file(filename)
  image_decoded = .decode_image(image_string) # (375, 500, 3)
 
  image_decoded = tf.expand_dims(image_decoded, axis=0)
 
  image_resized = .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 = .resize_image_with_crop_or_pad( tf.truncated_normal((1,220,300,3))*10, 200, 200) This four-dimensional form is fine
 
with () as sess:
 
  print( ( img ).shape  ) # Direct initialization can be converted to four-dimensional reported an error, I do not know why, if anyone wants to understand, please leave a message reported an error!
  #InvalidArgumentError (see above for traceback): Input shape axis 0 must equal 4, got shape [5]

import tensorflow as tf
 
import glob
import os
 
'''
  Dataset Batch read images:

    Reason:
      1. first define the list of image names, stored in Dataset from_tensor_slices ()
      2. Mapping function, in the function, the list of images to read, and resize, details
        tf.read_file(filename) return is three-dimensional, because this each time out of a picture, put into the queue, do not need to be converted to four-dimensional
        and then resize the picture, and then each batch to access the function, so get_next() returns [batch, w, h, c].
      3. Set up shuffle and batch repeat.

      4. iterator = dataset.make_one_shot_iterator() set iterator

      5. iterator.get_next() get the image for each batch
'''
 
def _parse_function(filename):
  # print(filename)
  image_string = tf.read_file(filename)
  image_decoded = .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 = .resize_images(label, [200, 200])
  ''' images three dimensional，Any of the 4D's will do.
     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 = .resize_image_with_crop_or_pad(image_decoded, 200, 200)
 
  # return (mage_resized,axis=0)
  return image_resized
 
filenames = ( ('../pascal/VOCdevkit/VOC2012/JPEGImages', "*." + 'jpg') )
 
 
dataset = .from_tensor_slices((filenames))
 
dataset = (_parse_function)
 
dataset = (10).batch(2).repeat(10)
iterator = dataset.make_one_shot_iterator()
 
img = iterator.get_next()
 
with () as sess:
  # print( (img).shape ) #(4, 200, 200, 3)
  for _ in range (10):
    print( (img).shape )