fit_generator and Multi-Input Models

[mhubri...@gmail.com]

Hello,

using the Merge layer to create multi-input models, is it still possible to train via fit_generator?

Let's say I have two input branches. This code

model.fit_generator([generator1, generator2])

gives me the following error message:
Exception: Error when checking model input: data should be a Numpy array, or list/dict of Numpy arrays. Found: <my.project.MyGenerator object at 0x2ae8364d868>

I'm aware that a list of Numpy arrays would solve the problem (and using model.fit() at the same time). Generators are just more convenient.

[Klemen Grm]

You can still use a generator with multiple input models, but it can only be a single generator. Your generator should return two lists, containing respectively a batch of each of your model's inputs and outputs.

[mhubri...@gmail.com]

Of course! It's always simpler than you think it is ;)

Thank you.

[jin...@eng.ucsd.edu]

Hi, but what if I want the fit_generator to generate two output with different batch_size? The batch_per_epoch can also take in two values?