Flexible pipelines with multiple tables, previews and easy hyperparameter tuning #1233
Conversation
|
Very much WIP, but already a cool example to look at https://output.circle-artifacts.com/output/job/34c076e7-2a4d-4616-946d-bfb5c27d2167/artifacts/0/doc/auto_examples/10_expressions.html#sphx-glr-auto-examples-10-expressions-py |
examples/10_expressions.py
Outdated
| # To help us visualize intermediate results and make development more | ||
| # interactive, we can provide a placeholder value for each variable. Skrub will | ||
| # use it to compute previews of the results. This helps inspect what the | ||
| # results will look like, catch errors early, and provide better tab-completion | ||
| # on attribute and item names. | ||
| # | ||
| # Instead of writing ``products = skrub.var("products")`` as before, let us | ||
| # start over, this time passing a placeholder value to make our expression a | ||
| # little bit more helpful: | ||
|
|
||
| # %% | ||
| products = skrub.var("products", placeholder=products_df) |
There was a problem hiding this comment.
I don't think this part is explained well. This section is called "Eager previews", but until now it's not clear that expressions are evaluated lazily. I think it should be made more explicit that with_total.skb.eval({"products": products_df}) does not print anything because it hasn't been executed yet.
You could say something like "if we try to print with_total at this stage, we do not get useful information. In order to obtain a preview of the current step, we need to add a placeholder value, which Skrub uses to compute a preview of the results".
Two more things:
- A preview of the results means that it contains the results up until this point, or in this particular operation, or a subset of the overall results (a sampling for example)?
- Is the
placeholdervariable used for something else? If it's only used for previews, maybe it would be better to be more explicit about it. Someone might get the idea that "you could put any placeholder df here to activate the preview". I'm not sure what the best variable name would be here, but I fearplaceholdermay lead to misunderstandings 🤔
There was a problem hiding this comment.
Thanks for reading through this @rcap107 . I'll try to improve this part based on your comment.
with_total.skb.eval({"products": products_df}) does not print anything
do you mean with_total does not print anything? because eval runs the computation, so the command above does produce a result
A preview of the results means that it contains the results up until this point, or in this particular operation, or a subset of the overall results (a sampling for example)?
the result of evaluating the expression with the bindings provided by the placeholder values
>>> a = skrub.var('a', 1)
>>> e = a + a + a
>>> e.skb.preview
3
but I fear placeholder may lead to misunderstandings
yes I agree it's not a good name. these values are used
- for the previews
- to provide data for cross-validation, fitting the estimator, or the hyperparam search.
In all cases they can be replaced by passing a dictionary so
pred.skb.cross_validate() # uses the "placeholder" data
pred.skb.cross_validate({'customers': cust, 'orders': orders, 'labels': labels}) # uses the provided values instead
maybe we should call it something more generic like value or data? or default_value?
There was a problem hiding this comment.
@rcap107 WDYT of default_value instead of placeholder? it conveys better that this value can be replaced by something else, whereas a placeholder must be replaced?
very much WIP, I'm opening the PR so the first example renders and we can discuss it.
The added examples are the last ones in the gallery
then, TODO: