# Does minibatch size affect accuracy?

**URL:** <https://discourse.pymc.io/t/does-minibatch-size-affect-accuracy/737>\
**Category:** Questions\
**Created:** [January 16, 2018, 12:22am UTC](https://discourse.pymc.io/t/does-minibatch-size-affect-accuracy/737 "2018-01-16T00:22:35Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![EtienneT](https://yyz2.discourse-cdn.com/flex036/user_avatar/discourse.pymc.io/etiennet/32/443_2.png) [@EtienneT](https://discourse.pymc.io/u/EtienneT)\
**Post date:** [January 16, 2018, 12:22am UTC](https://discourse.pymc.io/t/does-minibatch-size-affect-accuracy/737/1 "2018-01-16T00:22:35Z")

</div>

I am currently playing with ADVI and minibatches. I tried different size of minibatch size and I get very different results. I keep the n parameter in pm.fit constant to 10000 and the draws variable in sample to 10000 also…

inference = pm.ADVI()  
approx = pm.fit(method=inference, n=10000)  
approx.sample(draws=10000)

If I use minibatch with a size of 100, it finish in 37s, but if I test on unseen data (with PPC) I get an R-squared of 76% when I usually get 97-98% when using normal ADVI without minibatches. If I gradually increase batch size it seems to help or if I increase the pm.fit n parameter… Where could I get the intuition of setting those parameters to sensible values?

Thanks,

---

<div class="post-metadata">

**Author:** ![junpenglao](https://yyz2.discourse-cdn.com/flex036/user_avatar/discourse.pymc.io/junpenglao/32/8_2.png) [@junpenglao](https://discourse.pymc.io/u/junpenglao)\
**Post date:** [January 16, 2018, 5:35am UTC](https://discourse.pymc.io/t/does-minibatch-size-affect-accuracy/737/2 "2018-01-16T05:35:22Z")

</div>

In general, there is no good intuition of setting the batch size and the number of iteration in `pm.fit`. Your observation is normal that you need more iteration if your batch size is small. It is a difficult balance to keep, as small batch size is faster in terms of training, but takes more iteration. In general, I suggest plotting the `approx.hist` for different batch size and compare to make sure the model is converged to a (local) optimal.

Also, you probably dont need that many iterations in `approx.sample`…