The training procedure is very similar to those previously described in the book, and we invite the readers to refer to the notebook attached to this chapter. Using the model described previously, we reach a precision level of 72% on the validation set.
This result can be compared to the state-of-the-art precision level of 94%, which is obtained when using more advanced techniques. Our simple model could be enhanced by improving frame sampling, using data augmentation, using a different sequence length, or by optimizing the size of the layers.
