Keras often gets overlooked in the research environment in favour of Tensorflow.
Despite some limitations, I still think Keras is a viable option for prototyping research ideas with lightning speed! I’m glad you enjoyed this.
My preferred way of doing this is to have an “Abstract” Model which also implements the basic stuff like
predict() etc whose internals usually don’t change between models unless you switch from example keras to pytorch or something. I then inherit those methods into RandomForestModel.
Overall nice post! Good directory structure is too often overlooked in ML projects.
Thanks for reading. The
step function is applied to each time step (this is technically a for loop.)
Regarding point #1,
_stm is of size
(batch_size, dim, timesteps) because we use
K.repeat to effectively implement equation 1. Equation 1 is calculated for every character
j in the input sequence…
I think adding unit tests should be a more common piece of advice throughout this piece. I was able to diagnose and solve quite a few problems using this approach, particularly with my data loaders. However, now the issue becomes how to design good tests!
I don’t think there is a wrong way to learn programming. In fact my history is similar to yours, planned to be a doctor but programmed to keep me entertained during the summers. I learned PHP and most of what I learned was because I wanted to emulate Facebook and Twitter and Neopets.