Bypass sample weights

model.py

        # Implement REINFORCE rule by using crossentropy
        # + (reward - baseline) as sample_weight.
        self.model.compile(loss='categorical_crossentropy', optimizer=Adam(lr=self.lr))

        # Adapted from model.compile: redo the computation of loss.
        loss_functions = [objectives.get('categorical_crossentropy')]

        total_loss = None

        for i in range(len(self.model.outputs)):
            y_true = self.model.targets[i]
            y_pred = self.model.outputs[i]
            sample_weight = self.model.sample_weights[i]
            output_loss = loss_functions[i](y_true, y_pred) * sample_weight
            output_loss = K.mean(output_loss)
            if total_loss is None:
                total_loss = output_loss
            else:
                total_loss += output_loss

        for r in self.model.regularizers:
            total_loss = r(total_loss)

        self.model.total_loss = total_loss
        # For callbacks; normally in .fit
        self.model.validation_data = None