Keras: How to use fit_generator with multiple images input and one output?

Question

Is it possible make that fit_generator?

I'm creating a U-net network and I want to use as input a picture high 500, weight 500, and 5 channels, and on the output high 500, weight 500, and 1 channel

if it is not possible - i can create 500x500x5 np.arrays by myself and then I'll need a generator for load numpy objects from HDD

my code now (just for rgb picture)

train_generator=datagen.flow_from_directory('/content/data/',
                                                  target_size=(500,500),
                                                  color_mode='rgb',
                                                  batch_size=32,
                                                  class_mode='categorical', shuffle=False)

Answer 1

You should create your own generator. It is essential to get a while True : structure and yield your data. The code looks like this

batch_size=16
step_ep=data_size//batch_size

def generator():
  while True:
    for i in range(step_ep):
       process your images
       X=images
       Y=labels
       yield X,Y

With X of shape (batch_size,height,width,channel) and Yof shape (batch_size,height,width,output_channel)

You should use model.fit() instead of model.fit_generator because it will be soon deprecated.

You can also create a generator for your validation data.

Your model.fit() looks like :

model.fit(generator(),epochs,steps_per_epoch=step_ep, 
          validation_data=val_generator,validation_steps=val_steps)