CODEWITHSUNDEEP
tensorflowkerassemantic-segmentationunet-neural-networkimagedatagenerator
Unknown_ _
Unknown_ _

Reputation: 1

Passing ImageDataGenerator to segmentation model's U-net for multiclas classification

How can I pass an ImageDataGenerator to segmentation_model's U-net,


data_generator  = ImageDataGenerator(
                    rescale = 1./255.
)

train_dataset_images = data_generator.flow_from_directory(
                            directory=image_directory,
                            target_size = (256, 256),
                            class_mode = None,
                            batch_size = 32,
                            seed=custom_seed
)

train_dataset_masks = data_generator.flow_from_directory(
                            directory=mask_directory,
                            target_size = (256, 256),
                            batch_size = 32,
                            class_mode = None,
                            color_mode = 'grayscale',
                            seed=custom_seed
)

train_generator = zip(train_dataset_images, train_dataset_masks)

When I run this i run into a valueError saying "exptected 1 input but received 2" so I tried combining them with these functions:

def combine_generator(image_generator, mask_generator):
    while True:
        image_batch = image_generator.next()
        mask_batch = mask_generator.next()
        yield (image_batch, mask_batch)

and

def combine_generator(image_gen, mask_gen):
    for img, mask in zip(image_gen, mask_gen):
        yield img, mask

These don't seem to work.

  1. How can I pass ImageDataGenerator() to the U-net?
  2. Do the images and the masks have to have the same name?
  3. Do the masks have to be one hot encoded according to the classes?

Upvotes: 0

Views: 37

Answers (1)

Sagar
Sagar

Reputation: 169

Create a subfolders images and masked images inside the train directory.Pass it to image_datagen.flow_from_directory().Masks images are generally grayscale,create a generator that yields batch of images and masks .Create a U-net model and pass the generator to fit function, masks need not be one hot encoded. i have build U-net please refer to this gist

Upvotes: 0

Related Questions