DOC - Update the examples in the gallery so that they load datasets from their path

[rcap107]

#1852 added dataset paths to the fetchers, which means that now it is possible to write

from skrub.datasets import fetch_employee_salaries
import pandas as pd

path = fetch_employee_salaries().path
df = pd.read_csv(path)

instead of having magical access to the dataframe from the bunch object.

#1852 has updated some of the examples so that they use the path. All examples in the gallery that rely on skrub's fetchers should be updated so that they load from the path rather than using the dataset loaded in the bunch.

At some point in the future we are also going to remove loading from the fetchers entirely, so that it becomes necessary to load from the path: this is a more realistic situation anyway, and we want the examples to reflect it.

So for this issue (or through multiple issues) we want the examples in the gallery and elsewhere in the documentation to be updated so that they all load from file rather. For some examples this is quite simple to implement, some others may require more attention/rewriting.

The example changes in #1852 (like here https://github.com/skrub-data/skrub/pull/1852/changes#diff-050e3180cbff14a3de42224ccfae3ee7e9ae1a12eea4d38210e5535963e93e1c) can be used as templates

[jeromedockes]

👍 . also, at least one of the DataOps examples (or upcoming howto sections) should show a DataOp that takes the path as an input, rather than an already-loaded dataset (the read_csv is done inside the pipeline, not outside at the start of the example)

[MuditAtrey]

Hi, I'd like to work on this! I'll use the changes in #1852 as a template and start updating the gallery examples.

[MuditAtrey]

my first PR (#1940) updated 4 gallery examples 0040_fuzzy_joining.py, 0060_multiple_key_join.py, 0070_join_aggregation.py, 0080_interpolation_join.py to load datasets from their file paths using pd.read_csv().

Looking at the remaining examples like examples/data_ops/1110_data_ops_intro.py still accesses the dataframe directly from the bunch:

training_data = fetch_employee_salaries(split="train").employee_salaries

This one also uses split="train", which makes it more complex than the straightforward cases. Is this example in scope for this issue, and if so, what's the preferred approach for handling the train/test split?

[rcap107]

Hi @MuditAtrey, apologies for the slow response. The fetch_employee_salaries function with split="train" will return the path to the training split of the dataset in employee_salaries_path, and the same for the testing split. That means that the example does not change, so long as you specify the split (either train or test): you can load the file from the given path and it will work. You can look at the code of fetch_employee_salaries to find more detail on that.

In other words, you can make the same changes you already did in the other PR for this example. Should I assign the issue to you?

[MuditAtrey]

@rcap107 thanks for the response, yes- assign it to me, I'll start working on it.