如何配置Keras模型来预测图像？

Question

The main task is to predict a mask for the input image. So I have the following data for training:

• lot's of 768x768 original pics like this:

• and output mask pics(also 768x768) like this:

Also I have validation original pics.

I prepare some kind of neural model that should predict the output mask. I prepared keras model configuaration that should have a topology which looks like below:

The code I prepared for training is there.

import keras
epochs=100

image_datagen = keras.preprocessing.image.ImageDataGenerator()
mask_datagen = keras.preprocessing.image.ImageDataGenerator()
seed = 1
image_generator = image_datagen.flow_from_directory(
    'H:/LABS/ship_detection/test_train/',
    color_mode='rgb',batch_size=32,target_size=(768,768),
    seed=seed)

mask_generator = mask_datagen.flow_from_directory(
    'H:/LABS/ship_detection/test_mask/',
    class_mode="categorical",batch_size=32,target_size=(768,768),
    seed=seed)

train_generator = zip(image_generator, mask_generator)

model.fit_generator(generator=train_generator,
                    epochs=epochs,
                    callbacks=callbacks,steps_per_epoch=1)

But when I try to fit generator for prediction I have an issue:

c:\users\harwister\appdata\local\programs\python\python36\lib\site-packages\keras\engine\training_generator.py in fit_generator(model, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)
    208                     batch_size = list(x.values())[0].shape[0]
    209                 else:
--> 210                     batch_size = x.shape[0]
    211                 batch_logs['batch'] = batch_index
    212                 batch_logs['size'] = batch_size

AttributeError: 'tuple' object has no attribute 'shape'

I do something wrong for sure, but I can't understand anything from these kind of errors. The simple question I can't find a response in Google is: How can I push into Keras two images (input and output images) for training and after training get an output image providing an input image?

Answer 1

Since you have separate generators for the images and the labels (ie masks), you need to set the class_mode argument to None to prevent the generators from producing any labels arrays:

image_generator = image_datagen.flow_from_directory(class_mode=None, ...)
mask_generator = mask_datagen.flow_from_directory(class_mode=None, ...)

This way, image_generator would only generate the input images and the mask_generator would only generate the mask (ie true label) images.