zackchase/icu_rnn

possible example of the sample data

zergxyz opened this issue · 4 comments

zergxyz commented

Hello Zack:
Thanks a lot for your sharing of this great work! Just wondering if you can share a small example of the sample data because I'm having difficulties in understanding the data format used by this LSTM model.Specifically, in your comments I got this

"X should be a list of 2d arrays. First dimension idexes examples. Second dimension indexes sequences steps. Third dimension corresponds to the measurements within one sequence step."

I'm confused about the third dimension and hopefully I can get a real sample based on above comments. Thanks a again for your grate work.

U6C commented

Hi Zergxyz,

Did you mange to get the coding running with any sort of data? I think I am having errors with code, which are not related to the input data. Some seems like related to typo, with missing 's' at the end of a variable.
I was wondering whether you have the same issue, or just me.

Ahanio commented

Hi RuoxiYu,
I have the same issue with running the code. Did you manage to fix the issues?

jsbhat commented

I have the same issues.

Without touching any data, even the simplest instantiation of the class M2M_RNN doesn't work out of the box.

rnn = M2M_RNN(num_input=13, num_hidden=64, num_output=128, clip_at=0.0)

or

rnn = M2M_RNN(num_input=13, num_hidden=[64], num_output=128, clip_at=0.0)

The errors are simple to fix though...like wrong arguments order or typo but after 2-3 naive edits, they didn't go away and I do not know how deep is the rabbit hole.

Maybe the author doesn't follow this repo anymore and therefore this issue has been left open since long. I have sent him a personal email. Let's see.

zackchase commented

I hadn't followed the repo in some time. So pardon the delay. A few comments. First the inputs should just 2D numpy arrays. The first index should give the time step and the second should give the features (within a time step). I haven't run the code in about 1-2 years so it's possible some changes in Theano have caused bugs?

I'll look again soon when I can steal a few minutes. But likely the best thing to do is if citing this paper is to use the code from the followup work by my collaborator @turambar and colleagues who created an open benchmark: https://github.com/YerevaNN/mimic3-benchmarks

Among other advantages, that code is not in Theano so it should age better since Theano development is officially discontinued.