Frequently Asked Questions
. Can I train a deep learning model using RSMTool? [#]
rsmtool
supports a large number of learners but not deep
learning algorithms. However, you can still use rsmeval
to obtain a comprehensive evaluation report from any model.
. How can I customize rsmtool
and rsmeval
reports? [#]
The reports generated by rsmtool
and rsmeval
are designed to be fully
customizable. You can:
Omit sections using the general_sections configuration option;
Change the order of the sections using the section_order option;
Create your own “custom” sections.
. I get the following error: No responses remaining after filtering out
non-numeric feature values. No further analysis can be run.
What
happened? [#]
rsmtool
is designed to work only with numeric features. Non-numeric values
including missing values are filtered out. Some of the common reasons for this
error are:
The human score column or one of the features contains only non-numeric values. You can either exclude this feature or convert it to one-hot encoding
You have features with missing values. The solution is to replace missing values with zeros. Note that this applies even if you use the
.jsonlines
format. We have an open issue to provide an option to automatically convert missing values to zeros.You have a lot of missing feature values and none of the responses has numeric features for every single feature. Inspect the *_excluded_responses files in the output directory to see what responses have been excluded.
. Can I pass a set of learners to the model
option in the
configuration file or do I need a separate run for each leaner I want to try?
[#]
rsmtool
cannot train multiple models via a configuration file. If you wish
to use multiple learners/models, you should use the RSMTool API
instead of the command line.
. Can I compute predictions on new data using a model trained
with rsmtool
? [#]
Yes! We have built rsmpredict to do exactly this!
. Why did rsmtool
change the sign of some features in
feature.csv
? I thought rsmtool
assumes a positive sign for all
features? [#]
rsmtool
indeed assumes a default positive sign for all raw features.
However, if you set select_transformations
to true
in your experiment
configuration file, RSMTool will automatically apply
transformations to some of the features.
Some transformations such as inv
(inverse transform) change the polarity of
the feature. In such cases, RSMTool takes this into account and changes the
sign accordingly. See note here.
. Why does rsmpredict
exclude a different number of
responses compared to rsmtool
? [#]
This mismatch may happen due to the following:
First,
rsmtool
is run with one of the built-in feature selection models and excludes responses where any of the candidate features has missing/non-numeric values in either the training or the evaluation set. Next, it performs the feature selection, trains a model with those features using the filtered training set and generates predictions on the filtered test set.Next,
rsmpredict
is run to generate predictions on the evaluation set using the trained model and excludes only those responses where any of the selected features (a subset of the original features) has missing/non-numeric values.
As a result, there is a mismatch in the number of predictions generated by
rsmtool
and rsmpredict
.
. The relative betas for the linear regression model in my
rsmtool
run do not sum to 1. Is that a problem? [#]
Please check if your model has any negative coefficients. Relative coefficients only make sense when all model coefficients are positive. Their sum is expected to be less than 1 if there are negative coefficients. Note that if this is the case, the relative coefficients will not be included into the report.
. Can I use Weights & Biases to log artifacts and reports? [#]
Yes! W&B logging is natively supported by the following command-line tools:
rsmtool
, rsmeval
, rsmxval
, rsmexplain
, rsmcompare
, and
rsmsummarize
. To enable it, set the use_wandb
configuration option to true
and set the wandb_entity
and wandb_project options appropriately.
. It’s so difficult to remember all the configuration options. [#]
We agree that it can be quite overwhelming to remember all the configuration options. This is why all command-line tools support automatic generation of configuration files!