fit_generator does not expose a sample_weight parameter like fit

[loretoparisi]

While the signature of fit is

fit(x=None, y=None, batch_size=None, epochs=1, verbose=1, callbacks=None, validation_split=0.0, validation_data=None, shuffle=True, class_weight=None, sample_weight=None, initial_epoch=0, steps_per_epoch=None, validation_steps=None)

the fit_generator is

fit_generator(generator, steps_per_epoch=None, epochs=1, verbose=1, callbacks=None, validation_data=None, validation_steps=None, class_weight=None, max_queue_size=10, workers=1, use_multiprocessing=False, shuffle=True, initial_epoch=0)

so class_weight is defined for both, the validation_data can have optionally sample weights in the form of tuple (x_val, y_val, val_sample_weights), that will work on validation set only, but the fit_generator has no sample_weight option, i.e. the array of weights for the training samples.

Why training the model for a given number of epochs (i.e. fit) support sample weighting while training the model on data generated batch-by-batch (i.e. fit_generator) does not?

At the end of the day fit_generator will use the model.train_on_batch api:

def fit_generator(model,
                  generator,
                  steps_per_epoch=None,
                  epochs=1,
                  verbose=1,
                  callbacks=None,
                  validation_data=None,
                  validation_steps=None,
                  class_weight=None,
                  max_queue_size=10,
                  workers=1,
                  use_multiprocessing=False,
                  shuffle=True,
                  initial_epoch=0):

keras/keras/engine/training_generator.py

Line 215 in 351e7a9

outs = model.train_on_batch(x, y,

Now the method train_on_batch support both class_weight than sample_weight parameters:

keras/keras/engine/training.py

Line 1172 in 351e7a9

def train_on_batch(self, x, y,

def train_on_batch(self, x, y,
                       sample_weight=None,
                       class_weight=None):

Now what happens internally here

keras/keras/engine/training_generator.py

Line 334 in 351e7a9

if len(generator_output) == 2:

is like

if len(generator_output) == 2:
                x, y = generator_output
                sample_weight = None
            elif len(generator_output) == 3:
                x, y, sample_weight = generator_output
            else:
                raise ValueError('Output of generator should be a tuple '
                                 '(x, y, sample_weight) '
                                 'or (x, y). Found: ' +
                                 str(generator_output))

so a sample_weight is mapped from the generator output when your return the tuple (x_val, y_val, val_sample_weights), so apparently it will be used on the training batch, depending on your generator.
That said, why the fit_generator is not exposing the sample_weight argument then?

[gabrieldemarmiesse]

I see you opened two issue related to the same problem. Would it be possible for you to close one and follow everything on one single thread? thank you.

[loretoparisi]

@gabrieldemarmiesse correct, l have realized later that those issues could be related to the same problem. Which one do you suggest to close? Thank you.

[brge17]

Read the docs on fit_generator.

The output of the generator must be either

a tuple (inputs, targets)
a tuple (inputs, targets, sample_weights).

The generator for train or validation simply needs to

 yield X, y, w

[loretoparisi]

@brge17 thanks, as for #11805 (comment) the fit_generator and fit will be soon merged into one taking in account the api signature of both methods.