如何在python中使用张量流从图像创建形状（？，）和dtype = string的张量

Question

I have a trained model with input layer specified below:

Model input:  [<tf.Tensor 'encoded_image_string_tensor:0' shape=(None,) dtype=string>, <tf.Tensor 'key:0' shape=(None,) dtype=string>]

I have problem to create a tensor with these properties. Either i get right dtype but then i get shape() . Or i get a nonempty shape but dtype=uint8 or similar. Any tip on how to read, create and input my image to right format. The images that I want to input are grayscale , jpg , 3232x583 pixels.

Answer 1

You can do

import tensorflow as tf

a = tf.placeholder(dtype=tf.string, shape=[None, ], name="encoder_image_string_tensor")
print(a)

which prints

Tensor("encoder_image_string_tensor:0", shape=(?,), dtype=string)

For feeding value into this Tensor you can use sess.run and the feed_dict parameter inside this function.

To get the image in the right dimension you can do:

import cv2
im = cv2.imread("abc.jpg")
my_img = np.squeeze(np.reshape(im, [-1, 1]))
sess.run([], feed_dict={a: my_img})

I am sourcing my answer from here .

Answer 2

I was able to solve it with help of this anwser ( link ) I read it as a image with OpenCV. Encode it to jpg. Transform it to byte array. And make it to tensorflow with tf.constant(). Feels like the convertion from file to byte array could be made in a better way but i leave it with this for now.

code:

img = cv2.imread('IMAGEPATH')
flag, bts = cv2.imencode('.jpg', img) 
byte_arr = [bts[:,0].tobytes()]
tensor_string = tf.constant(byte_arr)

gives this which works:

Tensor("Const_14:0", shape=(1,), dtype=string)