CLR failing with KeyError: 'batch_size' during model.fit about keras-one-cycle

[iarganda]

A possible solution is to use the set_params method inherited from Callback

from clr import OneCycleLR

lr_manager = OneCycleLR(max_lr=10, maximum_momentum=0.9, verbose=True)
lr_manager.set_params({"batch_size":  64})

As explained here, that does not work because 'batch_size' is not included in params when it is built by fit_generator(). The solution I found is to modify the code and add the batch size and number of samples as parameters to OneCycleLR.

[prateekgulati]

keras-one-cycle/clr.py

Line 149 in 799e704

self.batch_size = self.params['batch_size']

[GCBallesteros]

The problem is that params is inherited from keras.callbacks.Callback and it is initialised based on the parameters passed to fit/fit_generator. You are not setting batch_size in your call to fit_generator (because you can't) and therefore can't be found.

I assume the example in the repo manages to find the batch_size argument because it extracts it from datagen.flow?

[GCBallesteros]

A possible solution is to use the set_params method inherited from Callback

from clr import OneCycleLR

lr_manager = OneCycleLR(max_lr=10, maximum_momentum=0.9, verbose=True)
lr_manager.set_params({"batch_size":  64})