ImageDataGenerator for semantic segmentation

Question

I am trying to do semantic segmentation with Keras and when trying to load the images i get this error using flow_from_directory method.

Found 0 images belonging to 0 classes.
Found 0 images belonging to 0 classes.

This is my code.

from tensorflow.keras.applications.resnet50 import preprocess_input
from tensorflow.keras.preprocessing.image import ImageDataGenerator

data_generator = ImageDataGenerator()
train_generator = data_generator.flow_from_directory(
                                        directory="../input/Training_dataset/Images",
                                        target_size=(IMG_SIZE, IMG_SIZE),
                                        batch_size=16,
                                        class_mode=None,
                                        classes=None
                                        )

mask_generator = data_generator.flow_from_directory(
    directory="../input/Training_dataset/Masks/all",
    class_mode=None,
    classes=None,
    batch_size = 1,
    )

I have read this question but solution didnt work Keras for semantic segmentation, flow_from_directory() error

Hangdog

Answer 1

you need to keep your images inside one sub-folder like create a folder named "img" inside both image and mask directory.

-- image
   -- img
      -- 1.jpg
      -- 2.jpg
-- mask
   -- img
      -- 1.png
      -- 2.png

Datagenerator should be like:-

seed = 909 # (IMPORTANT) to transform image and corresponding mask with same augmentation parameter.
image_datagen = ImageDataGenerator(width_shift_range=0.1,
                 height_shift_range=0.1,
                 preprocessing_function = image_preprocessing) # custom fuction for each image you can use resnet one too.
mask_datagen = ImageDataGenerator(width_shift_range=0.1,
                 height_shift_range=0.1,
                 preprocessing_function = mask_preprocessing)  # to make mask as feedable formate (256,256,1)

image_generator =image_datagen.flow_from_directory("dataset/image/",
                                                    class_mode=None, seed=seed)

mask_generator = mask_datagen.flow_from_directory("dataset/mask/",
                                                   class_mode=None, seed=seed)

train_generator = zip(image_generator, mask_generator)

if you want to make your own custom data generator for semantic segmentation models to get better control over dataset, you can check my kaggle kernel where i have used camvid dataset to train UNET model.

https://www.kaggle.com/mukulkr/camvid-segmentation-using-unet

If you need better augmentation fuction you can check this awesome GitHub repo - https://github.com/mdbloice/Augmentor

Rind

That worked with the images, but when trying to fit the model i get error AttributeError: 'zip' object has no attribute 'shape' Thanks anyways ill check your kernel – Hangdog

I am not getting any AttributeError can tell exactly where you are getting this error. – Rind

after trying to do this train_generator = zip(image_generator, mask_generator) model.fit_generator(train_generator, steps_per_epoch=167,epochs=10) – Hangdog

I have same dir structure. But my image name is "img (1).png" while mask name is "img (1)_mask.png" ... would that work? – Soloist

Nope, it should be the same as the original image name. – Rind

after running the line 'train_generator = zip(image_generator, mask_generator) ', memory just fills up entirely, even if working with just 4 images. What is going on in that line? – Archeology

Also, shouldn't shuffle be set to False? (Default is True) Because when we zip the two generators, won't it create a mismatch between images and masks? EDIT -> No. As per documentation --> seed is "Optional random seed for shuffling and transformations." so that should take care of this. – Coff

Could you explain this solution like why you used Zip and the same folder name for both directories? – Frication