tensorflow, How get value from Tensor-CodePudding

I upload the data in BatchDataset using the image_dataset_from_directory method

seed = 64
images = tf.keras.utils.image_dataset_from_directory(
    '/content/drive/MyDrive/DATA_PYTHON/Recognize_Alphabet/Recognize_Alphabet', 
    validation_split=0.2, 
    image_size=(34, 34),
    color_mode='rgb',
    interpolation='nearest',
    subset='training',
    seed=seed)

I want to invert the colors for all the images that I uploaded. To do this, I try to write a method:

def invertColor(im, b):

  sess2 = tf2.Session()
  im = sess2.run(im)

  imI = PIL.ImageOps.invert(im)
  imIN = np.asarray(imI)
  imINC = cv2.cvtColor(imIN, cv2.COLOR_BGR2RGB)
  bI = Image.fromarray(imINC, 'RGB')
  return bI

When I call the map with this invertColor method

images2 = images.map(invertColor)

I'm getting this errors:

InvalidArgumentError: in user code:
File "<ipython-input-19-1f1e09851e25>", line 5, in invertColor  *
    sess2.run(im)

InvalidArgumentError: Graph execution error:

Detected at node 'args_0' defined at (most recent call last):
    File "/usr/lib/python3.7/runpy.py", line 193, in _run_module_as_main
      "__main__", mod_spec)

How can I get the value of im element in the invertColor method? (Or how to invert colors in the BatchDataset?)

CodePudding user response：

You can try using tf.py_function to integrate PIL operations in graph mode. Here is an example with a batch size of 1 to keep it simple (you can change the batch size afterwards):

Before

import tensorflow as tf
import matplotlib.pyplot as plt
import pathlib
import PIL
import cv2
from PIL import Image
import PIL.ImageOps    
import numpy as np

dataset_url = "https://storage.googleapis.com/download.tensorflow.org/example_images/flower_photos.tgz"
data_dir = tf.keras.utils.get_file('flower_photos', origin=dataset_url, untar=True)
data_dir = pathlib.Path(data_dir)

batch_size = 1

train_ds = tf.keras.utils.image_dataset_from_directory(
  data_dir,
  validation_split=0.2,
  subset="training",
  seed=123,
  shuffle= False,
  image_size=(180, 180),
  batch_size=batch_size)

image, _ = next(iter(train_ds.take(1)))
plt.imshow(image[0].numpy() / 255)

After

def invert_color(image):
  im = Image.fromarray(image[0].numpy().astype('uint8'), 'RGB')
  imI = PIL.ImageOps.invert(im)
  imIN = np.asarray(imI)
  imINC = cv2.cvtColor(imIN, cv2.COLOR_BGR2RGB)
  return imINC / 255

def change_data(image, label):
  return  tf.py_function(invert_color, [image], Tout=[tf.float32]), label

train_ds = train_ds.map(change_data)

image, _ = next(iter(train_ds.take(1)))
plt.imshow(image[0].numpy())