"args_0:0" when printing Tensor

Question

I have an augmentation function that is being mapped into a generator; however, for some reason, the tfa.image.rotate function is causing an error.

def customGenerator(input_file_paths, dims, data_type):
    for i, file_path in enumerate(input_file_paths):
        if data_type.decode("utf-8") in ["png" or "tif"]:
            img = plt.imread((file_path.decode("utf-8")))
        elif data_type.decode("utf-8") == "npy":
            img = np.load(file_path.decode("utf-8"))
        x = resize(img[:,:,:3], dims)           
        yield x, x

def augment(image,label) :
    print('image', image)
    print('shape', image.shape)
    print('type', type(image))

    #angle = random.uniform(0, tf.constant(np.pi))
    image = tfa.image.rotate(image, tf.constant(np.pi)


train_dataset = tf.data.Dataset.from_generator(generator=customGenerator, 
                                                 output_types=(np.float32, np.float32), 
                                                 output_shapes=(dims, dims), 
                                                 args=[X_train_paths, dims, "png"])

train_dataset = train_dataset.map(augment, num_parallel_calls=AUTOTUNE)

I looked at the implementation of tfa.image.rotate other people had used, and theirs was working fine. I tried printing the image variable in the augment function. And this resulted:

print('image', image) # these lines is in the augment function, result below
print('type', type(image))

image Tensor("args_0:0", shape=(256, 256, 3), dtype=float32)
type <class 'tensorflow.python.framework.ops.Tensor'>

In contrast, when I go to other user's implementations and print their image, which is not mapped into a dataset. Their print(image) and print(type(image)) prints out this:

image tf.Tensor(
[[[1. 1. 1.]
  [1. 1. 1.]
  [1. 1. 1.]
  ...
  [1. 1. 1.]
  [1. 1. 1.]
  [1. 1. 1.]]

 ...

 [[1. 1. 1.]
  [1. 1. 1.]
  [1. 1. 1.]
  ...
  [1. 1. 1.]
  [1. 1. 1.]
  [1. 1. 1.]]], shape=(256, 256, 3), dtype=float32)

type <class 'tensorflow.python.framework.ops.EagerTensor'>

I expected this to be printing when I printed image in the augment function. So I am unsure what is going one. So a couple of questions. Also, tf.executing_eagerly() results in True

What exactly does "args_0:0" mean?

Should the image in the augment function be of type <class 'tensorflow.python.framework.ops.EagerTensor'> instead of a normal tensor?

Is there some way I can make "args_0:0" into the format I expected to have where it prints the array of numbers? Cause I believe that this will fix the rotation function

Finally, if not, is there a better way to augment the image with a random rotation?

Thank you for your time and help.

Can you share the result of `tf.executing_eagerly()`? – Aniket Bote Aug 29 '20 at 02:27 — Aniket Bote, Aug 29 '20 at 02:27
The result of tf.executing_eagerly() is true – dinodeep Aug 29 '20 at 02:30 — dinodeep, Aug 29 '20 at 02:30
What is dims arguments and the resize function used? – Aniket Bote Aug 29 '20 at 03:00 — Aniket Bote, Aug 29 '20 at 03:00

score 2 · Accepted Answer · answered Aug 29 '20 at 04:41

args_0:0 is tensor. See here

I made some changes to your code in order to make it work.

Code:

import tensorflow_addons as tfa
import os
import matplotlib.pyplot as plt
import tensorflow as tf
import numpy as np


def customGenerator(input_file_paths, dims, data_type):
    for i, file_path in enumerate(input_file_paths):
        image = tf.io.read_file(file_path)
        image = tf.image.decode_png(image, channels = 3)
        image = tf.image.convert_image_dtype(image, tf.float32)
        image = tf.image.resize(image, [dims[0],dims[1]])
        yield image, image

def augment(image,label) :
    img = tfa.image.rotate(image, tf.constant(np.pi/8))   
    return (img, label)

X_train_paths = [os.path.join('data','img',name) for name in os.listdir('data/img')]
dims = (256,256,3)

train_dataset = tf.data.Dataset.from_generator(generator=customGenerator, 
                                                 output_types=(tf.float32, tf.float32), 
                                                 output_shapes=(dims, dims), 
                                                 args=[X_train_paths, dims, "png"])

train_dataset = train_dataset.map(augment)

Iterating over the dataset:

for images in train_dataset:
    rotatedimg, normalimg= images[0],images[1]
    break

Output:

plt.imshow(rotatedimg)

plt.imshow(normalimg)

Things to remember:

The map function is not executed eagerly.
Always use TensorFlow function inside your generator function and map function because TensorFlow executes these functions as a part of the graph in order to speed up its execution. See here
If you use other functions TensorFlow might not be able to convert these operations into graphs which will result in an error.

Can you please accept and upvote the answer if it worked. – Aniket Bote Aug 30 '20 at 08:08 — Aniket Bote, Aug 30 '20 at 08:08

"args_0:0" when printing Tensor

1 Answers1