Quick overview on Keras with John (3/5/18)

[Chang Lee]

Last week, I have been exploring Keras for autoencoders on time-series data, so John showed me his gist (hey John, do you want to share it?) on a simple Keras model for a simulated time series data with a sequential model API. 

In my first attempt to use Keras with Tensorflow backend, the most confusing part was to get data into the right shape. So seeing how he converts the data into a the right tensor shape that Keras uses was helpful. I got my first Keras pipeline to work later that day!

He also told me he read that LSTMs are good for picking up time series patterns. I'm using convolutional networks right now, and LSTM seems to be a natural next step.

Here are two links from John -

https://keras.io/getting-started/faq/#how-can-i-use-stateful-rnns

https://machinelearningmastery.com/use-different-batch-sizes-training-predicting-python-keras/

Bonus: I have a talk coming up in a few days, so I asked John what was his process in preparing a talk. He said that he'd come up with everything that's related to the topic, then refine from there. I took the approach and got my slides down. We'll see how the talk goes :D

Chang

[JnBrymn]

Here is the Keras example that I was working with https://gist.github.com/JnBrymn-EB/1eacd73e1d73f329c6e3f84b931577e2. I am learning Keras myself right now and like Chang I thought that getting the tensor shapes correct was really difficult. I finally figured things out by creating a toy data set and figuring out how to make it learn the pattern in the data. Now I'm currently applying what I learned to the Kaggle Toxic Comment challenge https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge I'm attempting to use LTSMs to learn a language model based on training from these comments.

As always, nice chatting with you Chang!