search

microsoft / rat-sql

A relation-aware semantic parsing model from English to SQL
https://arxiv.org/abs/1911.04942
MIT License
406 stars 117 forks source link

how many iterations to get good performance on GloVe models? #38

Closed mrlamm closed 2 years ago

mrlamm commented 3 years ago

hello!

I'm currently training your nl2code-glove model with default settings, except with label_smooth loss. I'm wondering how many steps you found were needed to get the numbers reported in the paper.

thanks for your help.

cindycandy commented 3 years ago

hi，I wonder if you get the answer through training. I have trained the glove model for 31000 steps, and the loss of val is increasing. I want to stop it.

berlino commented 3 years ago

Hi, sorry for the late reply.

I used 40k steps for training, you can find it in the config here https://github.com/microsoft/rat-sql/blob/727cf7055cf9f3b468eac28d68318362f8aa22e5/configs/spider/nl2code-base.libsonnet#L94

@cindycandy sometimes dev accuracy keeps improving even when val loss is increasing.