Skip to content

1604WQM

Petr Baudis edited this page Apr 27, 2016 · 4 revisions

1604 WQMProp Model Tuning

Baseline based on preliminary testing: cnn model with cnnsiamese=False. On enttok, trainmodel MRR 0.55, val MRR 0.538, but this was on an older dataset version.

TODO:

  • Establish 4-way baseline performance.

  • Establish 4-way baseline performance on non-enttok.

  • Narrower filters - cdim={1:0.5,2:0.5,3:0.5}.

  • Try with attn1511.

  • Try without dropout inp_e_dropout=0 dropout=0.

  • Try with balance_class=True

  • Try with loss='binary_crossentropy'

  • Try with dot-product - ptscorer=B.dot_ptscorer.

  • Try with smaller batch (let's say 80).

Clone this wiki locally