Augmentation - Problems in training

Until a few days ago I worked with the version 0.13.4 of Rasa Core and had no problems in training. The configuration of my Policy was the following:

policies:
- name: KerasPolicy
	epochs: 90
	batch_size: 2048
	rnn_size: 100
	featurizer:
	- name: MaxHistoryTrackerFeaturizer
	max_history: 6
	state_featurizer:
		- name: BinarySingleStateFeaturizer
- name: MemoizationPolicy
	max_history: 6
- name: FallbackPolicy
	nlu_threshold: 0.6
	core_threshold: 0.3

At the time of training, this is my log:

(rasa1) PS C:\Byte\rasa1\Byte\bytebot-fcr-mig\bytebot-fcr-core> rasa train core --config .\policy.yml -d .\domain.yml --debug
2019-06-25 12:33:41 DEBUG    rasa.skill  - Selected skills:
Training Core model...
2019-06-25 12:33:43 DEBUG    rasa.core.training.generator  - Generated trackers will be deduplicated based on their unique last 6 states.
2019-06-25 12:33:43 DEBUG    rasa.core.training.generator  - Number of augmentation rounds is 3
2019-06-25 12:33:43 DEBUG    rasa.core.training.generator  - Starting data generation round 0 ... (with 1 trackers)
Processed Story Blocks: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 360/360 [00:00<00:00, 683.37it/s, # trackers=1]
2019-06-25 12:33:43 DEBUG    rasa.core.training.generator  - Finished phase (353 training samples found).
2019-06-25 12:33:43 DEBUG    rasa.core.training.generator  - Data generation rounds finished.
2019-06-25 12:33:43 DEBUG    rasa.core.training.generator  - Found 0 unused checkpoints
2019-06-25 12:33:43 DEBUG    rasa.core.training.generator  - Starting augmentation round 0 ... (with 50 trackers)
Processed Story Blocks: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 360/360 [00:15<00:00, 21.07it/s, # trackers=27]
2019-06-25 12:33:58 DEBUG    rasa.core.training.generator  - Finished phase (11601 training samples found).
2019-06-25 12:33:59 DEBUG    rasa.core.training.generator  - Starting augmentation round 1 ... (with 50 trackers)
Processed Story Blocks: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 360/360 [00:19<00:00, 18.73it/s, # trackers=44]
2019-06-25 12:34:18 DEBUG    rasa.core.training.generator  - Finished phase (24777 training samples found).
2019-06-25 12:34:18 DEBUG    rasa.core.training.generator  - Starting augmentation round 2 ... (with 50 trackers)
Processed Story Blocks: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 360/360 [00:19<00:00, 18.49it/s, # trackers=42]
2019-06-25 12:34:37 DEBUG    rasa.core.training.generator  - Finished phase (37790 training samples found).
2019-06-25 12:34:37 DEBUG    rasa.core.training.generator  - Found 37790 training trackers.
2019-06-25 12:34:37 DEBUG    rasa.core.training.generator  - Subsampled to 500 augmented training trackers.
2019-06-25 12:34:37 DEBUG    rasa.core.training.generator  - There are 353 original trackers.
2019-06-25 12:34:38 DEBUG    rasa.core.agent  - Agent trainer got kwargs: {'dump_stories': False}
2019-06-25 12:34:38 DEBUG    rasa.core.featurizers  - Creating states and action examples from collected trackers (by MaxHistoryTrackerFeaturizer(BinarySingleStateFeaturizer))...
Processed trackers: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 853/853 [00:03<00:00, 233.28it/s, # actions=6650]
2019-06-25 12:34:42 DEBUG    rasa.core.featurizers  - Created 6650 action examples.
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
masking (Masking)            (None, 6, 451)            0
_________________________________________________________________
lstm (LSTM)                  (None, 100)               220800
_________________________________________________________________
dense (Dense)                (None, 186)               18786
_________________________________________________________________
activation (Activation)      (None, 186)               0
=================================================================
Total params: 239,586
Trainable params: 239,586
Non-trainable params: 0
_________________________________________________________________
2019-06-25 12:34:42 INFO     rasa.core.policies.keras_policy  - Fitting model with 6650 total samples and a validation split of 0.1
2019-06-25 12:34:42 DEBUG    rasa.core.policies.policy  - Parameters ignored by `model.fit(...)`: {}
Epoch 1/90
- 2s - loss: 5.1224 - acc: 0.1150
Epoch 2/90
- 1s - loss: 4.4750 - acc: 0.3254
Epoch 3/90
- 1s - loss: 3.6577 - acc: 0.3280
Epoch 4/90
- 1s - loss: 3.4156 - acc: 0.3280
Epoch 5/90
- 1s - loss: 3.3060 - acc: 0.3280
Epoch 6/90
- 1s - loss: 3.2438 - acc: 0.3280
Epoch 7/90
- 1s - loss: 3.2002 - acc: 0.3280
Epoch 8/90
- 1s - loss: 3.1624 - acc: 0.3280
Epoch 9/90
- 1s - loss: 3.1265 - acc: 0.3280
Epoch 10/90
Traceback (most recent call last):
...

Why do you subsample my training trackers to 500?

Going deeper into the source rasa_core/generator.py at master · RasaHQ/rasa_core · GitHub, I realized that commenting from line 372 to line 386 made the training went correctly, I did not do the sub-sample and the predictions were correct.

    if self.config.augmentation_factor > 0:
        augmented_trackers, original_trackers = [], []
        for t in finished_trackers:
            if t.is_augmented:
                augmented_trackers.append(t)
            else:
                original_trackers.append(t)
        augmented_trackers = self._subsample_trackers(
            augmented_trackers, self.config.max_number_of_augmented_trackers
        )
        logger.debug(
            "Subsampled to {} augmented training trackers."
            "".format(len(augmented_trackers))
        )
        logger.debug(
            "There are {} original trackers.".format(len(original_trackers))
        )
        finished_trackers = original_trackers + augmented_trackers

Could it be considered a bug? Are there any parameters that you may be missing? Please help.

subsampling were added because with more stories the number of augmented trackers increased exponentially.

if you run rasa train core ... --augmentation 100, augmented trackers will be subsampled to a 1000 (so 10x) I recommend to try different augmentation factors (default 50) instead of creating uncontrollably huge number of trackers