自定义图像生成器 function 解释为生成器 object？

Question

I recently learned that Python interprets functions with return and functions with yield separately. I made this custom generator that uses Keras ImageDataGenerator to generates batches of images from a directory into a GAN Deep Learning network:

def loadRealImages(batch):
for gen in pixGen.flow_from_directory(picturesPath, target_size = (256, 256),
                                    batch_size = batch, class_mode = "binary"):
    realImgs = [image for image in gen]
    yield realImgs[0]

My program expect what this generator yields (tensor of batch images) to run an analysis through a network and output a prediction:

real = loadRealImages(1) # get one real image
prediction = discriminator([real, fake]) # check similarity

My problem is that because of the call to the custom generator, real is considered a generator object loadRealImages when the program is expecting an image in tensor form and I get this error:

AttributeError: 'generator' object has no attribute '_keras_mask_'

From what I found looking into the problem, it seems generators don't get executed during interpretation. Then how should I bypass this check during interpretation?

PS: in this particular line, I only need one image so I could load it in a different way but I have others of these generator calls throughout the code and I can't really delete them.

Answer 1

I don't know about the keras library, but generators are a bit different from functions in that calling them returns a generator, not the first yield object. You have to call next on the returned generator to get the image

def img_generator(num_imgs):
  for ii in range(num_imgs):
    yield np.random.random((256, 256)) > 0

one_img_generator = img_generator(1)
img = next(one_img_generator)
next(one_img_generator) # will raise StopIteration

Does that make sense?

Your generator could also be made more 'natural', ie the pythonic way of writing it would be:

def loadRealImages(batch):
  for gen in pixGen.flow_from_directory(picturesPath, target_size = (256, 256),
                                    batch_size = batch, class_mode = "binary"):
    yield next(gen)

This way, the whole list isn't placed into memory at the same time (if gen is a generator). Instead, only the requested element is in memory.