# Some troubleshooting for setting up Numerai Compute

**URL:** <https://forum.numer.ai/t/some-troubleshooting-for-setting-up-numerai-compute/3623>\
**Category:** Tournament\
**Created:** [June 19, 2021, 7:04pm UTC](https://forum.numer.ai/t/some-troubleshooting-for-setting-up-numerai-compute/3623 "2021-06-19T19:04:14Z")\
**Posts on this page:** 1\
**Showing post:** 7

<div class="post-metadata">

**Author:** ![ml\_is\_lyf](https://avatars.discourse-cdn.com/v4/letter/m/e56c9b/32.png) [@ml\_is\_lyf](https://forum.numer.ai/u/ml_is_lyf)\
**Post date:** [June 20, 2021, 8:43pm UTC](https://forum.numer.ai/t/some-troubleshooting-for-setting-up-numerai-compute/3623/7 "2021-06-20T20:43:06Z")

</div>

I managed to get it working, looks like I just needed to be a bit smarter on how to load the dataset using pandas. Previously I was casting to np.float16 AFTER loading in the dataset. But I saw JRB’s post where he sets it to np.float16 during reading via the dtype arg:

> [@Saving memory with uint8 features](http://forum.numer.ai/t/saving-memory-with-uint8-features/254):
>
> I’ve noticed that there have been quite a few complaints in the chat rooms about memory usage on compute since the test set got bigger. I’d originally contributed the float16 patch to example\_model.py to help alleviate this, a couple of months ago. I just realized that there’s a simple way improve on this and to nearly halve the memory usage in your inference pipeline if you’re using a tree based classifier (Xgboost, Catboost, Lightgbm, etc). We just need to rescale the features and cast them t…

So I changed my code to set the dtype as np.float16 for the feature column and target like so:

```auto
tournament_data = pd.read_csv(tournament_data_file_path, dtype={COLUMN_NAME: np.float16 for COLUMN_NAME in FEATURE_COLUMNS+["target"]})

```

And now my prediction script runs successfully on the gen-md node.

---

_[View the full topic](https://forum.numer.ai/t/some-troubleshooting-for-setting-up-numerai-compute/3623)._