change init method for torch pnet - Githubissues

zhanglab-aim / cancer-net

Diagnosing cancers using deep learning.

GNU General Public License v2.0

2 stars 0 forks source link

change init method for torch pnet #30

Closed zj-zhang closed 1 year ago

zj-zhang commented 1 year ago

DONE:

change kernel initializer from normal to glorot uniform to match TF; this change increased training stability
Add new scripts under reprod_report folder for generating P-net related results for reusability report.

TODO: TF components not implemented in torch that may contribute to the last bit of performance gap:

kernel regularizers for SparseLayer and FeatureLayer: L2 wregs = [0.001] * 7
kernel regularizers for intermediate outcomes: L2 wreg_outcomes = [0.01] * 6