The Sequentiallayer was introducing more problems
than the ones it solved. Unfortunately, the old checkpoints
will not be compatible but they can be easily converted using:
```python
old_model = tf.keras.models.load_model()
new_model = bob.learn.tensorflow.models.incepetion_resenet_v2.IncepetionResnetV2()
new_model.set_weights(old_model.get_weights)
new_model.save(".../new_checkpoint")
```