[DOC] TabICL tabular foundation model#705
Conversation
| Housing regression dataset. | ||
| """ | ||
|
|
||
| import pandas as pd |
There was a problem hiding this comment.
please perform the import wherever they're needed. It's indeed not standard, but makes the example more readable.
|
|
||
| model = make_pipeline(StandardScaler(), TabICLRegressor(device="cpu")) | ||
| model.fit(X_train, y_train) | ||
|
|
There was a problem hiding this comment.
maybe report model accuracy on the test data ?
There was a problem hiding this comment.
Model accuracy is of 0.743
| # running TabICL on the CPU is slow. We recommend to use it on larger datasets, | ||
| # be it in terms of samples and/or features and to run it on GPU for faster | ||
| # inferences. Due to technical limitations, this is something we cannot showcase | ||
| # here. |
There was a problem hiding this comment.
you can skip the last sentence imho.
The example LGTM.
* [ENH] initial implementation of LOCI. * Changed importance measure, and fixed predict_one_features_group according to failing test. * added multilabel case handling * raised fdr threshold for less agressive power testing. * Implemented multiclass test, and fixed LOCI multiclass code. * Baseline mean of y is now taken on the train set. Reduced sample and feature size for tests. --------- Co-authored-by: jpaillard <joseph.paillard@inria.fr>
|
I made the changes following your reviews. It's not the most interesting example, but it shows that we can interface HiDimStat with SOTA models. |
|
However, I can't seem to understand why the CI doesn't want to build the documentation even with the [doc] tag that forces the doc build. |
|
After a look at the CI, since this PR doesn't target the main branch, the CI will never be run ... So to see the result, this would need to be merged into the target branch, which this time targets main ... (same issue as with the PyTorch/Skorch example) |
2 participants
I added a basic example of TabICL usage with HiDimStat on the California Housing dataset. Unfortunately, Github doesn't give free access to GPUs, so I limited the example to 1000 samples which still amounts to a runtime of 1min30s for the entire script.
Also, TabICL currently runs only with CFI and PFI. TabICL requires an integer seeding, while D0CRT directly works with a seeding generator object, making them conflict.
The example is basic, I did not manage to find something interesting to do with it that did not explode computation time since everything has to run on CPU.