What is the validation principle in ImageDataGenerator?

Question

Neural networks

I made a generator from the directory (flow_from_directory) with the selection of validation data (validation_split=0.25). Based on it, there are two fillings: test (train_generator1) and verification (validation_generator1).

datagen1 = ImageDataGenerator(rescale=1./255, rotation_range=60, width_shift_range=0.2, height_shift_range=0.2, zoom_range=0.5, horizontal_flip=True, vertical_flip=True, validation_split=0.25)
train_generator1 = datagen1.flow_from_directory(directory=dir1, target_size=(150,150), batch_size=39, class_mode='sparse', save_to_dir=dir2_1, save_prefix=u'XXX', subset='training')
validation_generator1 = datagen1.flow_from_directory(directory=dir1, target_size=(150,150), batch_size=39, class_mode='sparse', save_to_dir=dir2_2, save_prefix=u'XXX', subset='validation')
history = model.fit_generator(train_gen, steps_per_epoch=156, validation_data=val_gen, validation_steps=52)

After going through the generator, I looked at the test directory and found that the test images were transformed in the same way as the training ones. Tell me, what is the principle of training? Should the test images transform too or remain static?

Answer 1

the test images were transformed in the same way as the training ones

But you created them on the basis datagen1of which you specified shifts, rotations, etc. In any case, this serves to increase the data sets, there is nothing wrong with that.