replication of model prediction - deviation of ~2% random seeds missing

[partizanos]

Hello,

I try to replicate the training and evaluation using LSTM network for the phenotype benchmark.
I would like to ask if there are specific random seeds set in your system for the model training. (for the folders I saw the code provided them but for training no)
ie
from tensorflow import set_random_seed set_random_seed(2)
I get very close results (+/- 0.02 deviation) and I suspect this might be the issue. But let me know if there was some change

screenshot refers to this version of the paper:https://arxiv.org/pdf/1703.07771.pdf

Some more potential reasons for non replicability

• I believe my data folders have the correct size ie: train/test size: 35622/6282
• Regarding the environment I have the following dependencies using conda .
  I faced issues with newer versions of python (I use python=3.7.12=hf930737_100_cpython)
  complete set of dependencies: deps.txt

Thank you again for your work and active maintenance of this repository.