keras fit_generator: 'zip' object has no attribute 'shape'

Question

as in keras documentation in the (image, mask) data generator, I created data generator to have (input, output) tuple images.
When running locally on my laptop (with tensorflow 1.13.1), it runs with no issues, but when running on a server (with tensorflow-gpu 1.13.1), I get the error:

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

so as in the example, after creating two generators with flow_from_directory:

from tensorflow import keras

...

input_gen = input_datagen.flow_from_directory(
        directory=os.path.join(path_in, dirname),
        class_mode=None,
        color_mode=color_mode,
        batch_size=batch_size,
        target_size=(img_size, img_size),
        seed=seed_in)

I combined them as in the keras documentation:

train_generator = zip(input_gen, output_gen)

and feed them to the fit_generator(...)

currently solved it using a function to generate the new combined generator as suggested here:

def combine_generator(gen1, gen2):
    while True:
        yield(gen1.next(), gen2.next())   

though I would still like to understand why the tensorflow cpu version doesn't get this error, and why the zip isn't supported in the gpu version...

Answer 1

Looks like support of fit_generator in Tensorflow 1.13 has been changed. As mentioned in this issue, workaround,

Replace

train_generator = zip(input_gen, output_gen)

with

train_generator = (pair for pair in zip(input_gen, output_gen))