Make automatic centering in PCA methods optional #808

hristog · 2021-03-20T12:43:53Z

Closes: Flag to skip centering in PCA #734.
Tests/black/pyflakes passing: Yes.
Tests added to the following files:**
- tests/test_pca.py
- tests/test_incremental_pca.py
tl;dr of design choices made:
- IncrementalPCA doesn't get affected by the new parameter and IncrementalPCA(..., center=False ,...) is unsupported.
- whiten=True does not imply, nor enforce center=True. Instead it warns about possible unexpected behavior, if the underlying data hasn't been centered already.

hristog · 2021-03-21T19:39:02Z

Once the maintainers confirm approval of the proposed changes, I'll be happy to update the .rst docs accordingly, to document the updated API, in a separate commit (as part of this PR).

dask_ml/decomposition/incremental_pca.py

TomAugspurger · 2021-03-23T22:29:38Z

dask_ml/decomposition/pca.py

+            warnings.warn(
+                "Whitening requires centering. Please, ensure that your data "
+                "is already centered, in order to avoid unexpected behavior.",
+                RuntimeWarning,


Users should have some way to silence this warning. I'm inclined to just document this requirement, and trust that the user has done it (no warning).

Sure, I'll remove the warning (the new param has been already added to the doc-string). Please, let me know if any other documentation-related changes are necessary. I've checked api.rst and compose.rst but it doesn't seem that we need any updates to either.

dask_ml/decomposition/pca.py

dask_ml/decomposition/incremental_pca.py

hristog · 2021-03-24T18:55:22Z

From the latest CI runs (e.g., this job).

black, version 19.10b0
would reformat /home/vsts/work/1/s/tests/test_pca.py
Oh no! 💥 💔 💥
1 file would be reformatted, 99 files would be left unchanged.

Oops - pretty sure that I did this check locally. Will re-run black to confirm.

Update: This is strange. I'll check if my local branch has actually propagated, and is the same as what the Azure pipeline has picked up.

Result of ./ci/code_checks.sh run locally from the root dir of `dask-ml`:

Checking flake8...
Checking flake8... DONE
Checking black...
black, version 20.8b1
would reformat /git/dask-ml/dask_ml/model_selection/utils_test.py
would reformat /git/dask-ml/tests/ensemble/test_blockwise.py
would reformat /git/dask-ml/tests/model_selection/test_keras.py
would reformat /git/dask-ml/tests/model_selection/test_hyperband.py
would reformat /git/dask-ml/tests/model_selection/dask_searchcv/test_model_selection.py
would reformat /git/dask-ml/tests/model_selection/test_incremental.py
Oh no! 💥 💔 💥
6 files would be reformatted, 94 files would be left unchanged.
Checking black... DONE
Checking isort...
5.6.4
ERROR: /git/dask-ml/tests/test_spectral_clustering.py Imports are incorrectly sorted and/or formatted.
ERROR: /git/dask-ml/tests/test_incremental_pca.py Imports are incorrectly sorted and/or formatted.
ERROR: /git/dask-ml/tests/linear_model/test_glm.py Imports are incorrectly sorted and/or formatted.
ERROR: /git/dask-ml/tests/model_selection/dask_searchcv/test_model_selection_sklearn.py Imports are incorrectly sorted and/or formatted.
ERROR: /git/dask-ml/docs/dimensions.py Imports are incorrectly sorted and/or formatted.
ERROR: /git/dask-ml/dask_ml/_compat.py Imports are incorrectly sorted and/or formatted.
...

Update (2): Seems that it's the same code version, both locally and remotely.

Locally:

$ git show
commit 01dd9d4f71ee380e5b3e1a969f3bd46bebda6cd9 (HEAD -> 734-pca-skip-centering, origin/734-pca-skip-centering)
...
    Add param validation to PCA classes (#808)

Remotely: The most recent commit (and only one since Tom's review): 01dd9d4.

I'll try reproducing black's results via the same version as per the CI pipeline (black, version 19.10b0).

Update (3): Victory! It was indeed due to the different local version of black==black-20.8b1 vs remote (black==19.10b0, as per .pre-commit-config.yaml and ci/*.yaml).

hristog · 2021-04-08T17:11:28Z

Hi @TomAugspurger - I'm writing to check if you've had a chance to look into the further changes that I pushed since your review, and whether you'd like any additional refinements to be done on top of that.
Thanks!

TomAugspurger · 2021-04-12T11:41:03Z

It'll be a week or so before I can take a look. Thanks for your patience!

In the meantime, perhaps @jameslamb has a chance to glance through this (no worries if you don't though)?

hristog · 2021-04-12T12:25:20Z

It'll be a week or so before I can take a look. Thanks for your patience!

It's all good - just wanted to check if this is still on your radar. I'm happy to wait as long as necessary, and incorporate potential further feedback.
Thanks for your confirmation, @TomAugspurger!

jameslamb

I had some time to go through and review today. Overall, I'm in favor of the change and I can understand how it could be a nice optimization for experienced users.

Please consider some of the review comments I left.

I also think it would improve understanding for people who find this pull request from search or see it in a changelog if you change the PR title to something like "Make automatic centering in PCA methods optional". By itself, the fact that there is now a parameter called center doesn't tell you much about the functional benefit of the changes in this pull request.

jameslamb · 2021-04-12T20:56:32Z

dask_ml/decomposition/incremental_pca.py

@@ -127,6 +127,7 @@ def __init__(
        self,
        n_components=None,
        whiten=False,
+        center=True,


Introducing new arguments in the middle of a signature is a breaking change for anyone using these classes with positional arguments. Can you please put this new argument at the end of the signature instead of in the middle?

@hristog you marked this as resolved, but I don't think it is yet. Could you put copy back between whiten and batch_size and make center the last keyword argument?

jameslamb · 2021-04-12T21:05:16Z

dask_ml/decomposition/pca.py

+    center : bool, optional (default True)
+        When False (True by default), the underlying data gets centered at zero
+        by subtracting the mean of the data from the data itself.
+
+        PCA is performed on centered data due to its being a regression model,
+        without an intercept. As such, its pricipal components originate at the
+        origin of the transformed space.
+
+        `center` set to False may be employed when performing PCA on already
+        centered data.
+
+        Since centering is a required step as part of whitening, `center` set
+        to False and `whiten` set to True is a combination which may result in
+        unexpected behavior, if performed on not previously centered data.
+


Suggested change

center : bool, optional (default True)

When False (True by default), the underlying data gets centered at zero

by subtracting the mean of the data from the data itself.

PCA is performed on centered data due to its being a regression model,

without an intercept. As such, its pricipal components originate at the

origin of the transformed space.

`center` set to False may be employed when performing PCA on already

centered data.

Since centering is a required step as part of whitening, `center` set

to False and `whiten` set to True is a combination which may result in

unexpected behavior, if performed on not previously centered data.

center : bool, optional (default True)

When True (the default), the underlying data gets centered at zero

by subtracting the mean of the data from the data itself.

PCA is performed on centered data due to its being a regression model,

without an intercept. As such, its principal components originate at the

origin of the transformed space.

``center=False`` may be employed when performing PCA on already

centered data.

Since centering is a required step as part of whitening, ``center`` set

to False and ``whiten`` set to True is a combination which may result in

unexpected behavior, if performed on not previously centered data.

I believe there are a few mistakes in this documentation as it's currently written.

two backticks should be used for inline code formatting, since these docstrings are interpreted as reStructuredText by Sphinx

the docstring should say that the default behavior (center=True) is to automatically center the data (it currently says the opposite)

the docstring should say that the default behavior (center=True) is to automatically center the data (it currently says the opposite)

Well spotted, thanks!
Updating as per the double backtick suggestion. I'll update the rest of the affected docstrings, to conform to the same guideline (unless we'd like to do that in a separate, dedicated PR?).

If I was a maintainer here, I'd prefer that a docstring-only change be made in a separate pull request, and that this comment only apply to the lines you are already adding / changing as of this pull request

jameslamb · 2021-04-12T21:05:51Z

dask_ml/decomposition/pca.py

@@ -149,21 +164,28 @@ class PCA(sklearn.decomposition.PCA):
    >>> dX = da.from_array(X, chunks=X.shape)
    >>> pca = PCA(n_components=2)
    >>> pca.fit(dX)
-    PCA(copy=True, iterated_power='auto', n_components=2, random_state=None,
-      svd_solver='auto', tol=0.0, whiten=False)
+    PCA(n_components=2)


why were these changes necessary? They seem unrelated to the current pull request. Could you please revert them or explain why they're necessary?

It really depends whether those are considered to be demonstrative examples (in doctest style) or actual doctests which are meant to pass.

If the case is the latter, then the proposed changes should probably still hold.

For example, in their original state, these doctests fail when run with:

py.test --verbose --doctest-modules dask_ml/

If, on the other hand, they're just meant to serve for demonstration only (section 15 of the numpydoc standard), then I'll be definitely happy to revert back these changes.

This PR is adding one new parameter with a default value set to the current behavior. I wouldn't expect any tests to fail, and if they are I'd like to know more about what failed (a link to logs, if you have it).

I don't think the doctests get run as part of the CI steps, and I imagine the doctests haven't been updated for quite some time (haven't checked with git blame) or the intention has never been that they get run as actual first-class-citizen test (but, instead, serve for demonstration).

E.g., when running against the latest main:

$ py.test --verbose --doctest-modules dask_ml/ ... dask-ml/dask_ml/wrappers.py:92: DocTestFailure ___________________ [doctest] dask_ml.decomposition.pca.PCA ____________________ 142 143 Examples 144 -------- 145 >>> import numpy as np 146 >>> import dask.array as da 147 >>> from dask_ml.decomposition import PCA 148 >>> X = np.array([[-1, -1], [-2, -1], [-3, -2], [1, 1], [2, 1], [3, 2]]) 149 >>> dX = da.from_array(X, chunks=X.shape) 150 >>> pca = PCA(n_components=2) 151 >>> pca.fit(dX) Expected: PCA(copy=True, iterated_power='auto', n_components=2, random_state=None, svd_solver='auto', tol=0.0, whiten=False) Got: PCA(n_components=2) ...

huh, ok I see. I'm not too familiar with that type of test, apologies. But it does seem unrelated to the change in this pull request, so I think it would be better to revert those changes here and open a separate issue / PR about the general topic of "doctest tests are failing / not run"

I don't recall if we run doctests here or not. I wouldn't worry about that failing one. Scikit-Learn changed the repr of their objects to only show the parameters that are not the default, so the docstring changed.

dask_ml/decomposition/pca.py

jameslamb · 2021-04-12T21:12:42Z

dask_ml/decomposition/incremental_pca.py

        self.copy = copy
        self.batch_size = batch_size
        self.svd_solver = svd_solver
        self.iterated_power = iterated_power
        self.random_state = random_state

+    def _check_params(self):
+        super()._check_params()


What is the purpose of defining an empty _check_params() on the parent class and then calling it here? In my opinion, that is adding a bit of additional indirection for no functional benefit. Could you please remove this line and the parent class's empty _check_params()?

The purpose is to future-proof this by ensuring any checks that get eventually introduced in the superclass, get considered at this level as well.
Happy to remove the perceived as redundant check, at the current stage. Just my opinion is that it introduces a bit more exposure to bugs, if the class hierarchy doesn't get utilized in cases like this. But it may as well be done at a further stage.

hristog · 2021-04-12T21:27:56Z

I had some time to go through and review today. Overall, I'm in favor of the change and I can understand how it could be a nice optimization for experienced users.

Please consider some of the review comments I left.

I also think it would improve understanding for people who find this pull request from search or see it in a changelog if you change the PR title to something like "Make automatic centering in PCA methods optional". By itself, the fact that there is now a parameter called center doesn't tell you much about the functional benefit of the changes in this pull request.

Thanks for your review, @jameslamb! I'll try to address your comments and make changes, wherever applicable, ASAP.

…#808)

TomAugspurger

I think that https://github.com/dask/dask-ml/pull/808/files#r615299186 is the only outstanding issue. Good to go once that's fixed.

…k#808

TomAugspurger reviewed Mar 23, 2021

View reviewed changes

hristog added a commit to hristog/dask-ml that referenced this pull request Mar 24, 2021

Add param validation to PCA classes (dask#808)

01dd9d4

hristog commented Mar 24, 2021

View reviewed changes

dask_ml/decomposition/incremental_pca.py Show resolved Hide resolved

hristog force-pushed the 734-pca-skip-centering branch 2 times, most recently from 266ae3f to 4c82add Compare March 31, 2021 09:23

hristog requested a review from jameslamb April 12, 2021 21:15

jameslamb reviewed Apr 12, 2021

View reviewed changes

hristog changed the title ~~Introduce a center param to PCA (#734)~~ Make automatic centering in PCA methods optional Apr 12, 2021

hristog added a commit to hristog/dask-ml that referenced this pull request Apr 15, 2021

Improve docstrings and remove redundant checks (fixes dask#734) (dask…

b2689d3

…#808)

TomAugspurger approved these changes Apr 17, 2021

View reviewed changes

Make automatic centering in PCA methods optional (fixes dask#734) das…

d241040

…k#808

hristog force-pushed the 734-pca-skip-centering branch from b2689d3 to d241040 Compare April 20, 2021 15:21

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Make automatic centering in PCA methods optional #808

Make automatic centering in PCA methods optional #808

hristog commented Mar 20, 2021 •

edited

hristog commented Mar 21, 2021

TomAugspurger Mar 23, 2021

hristog Mar 24, 2021

hristog commented Mar 24, 2021 •

edited

hristog commented Apr 8, 2021

TomAugspurger commented Apr 12, 2021

hristog commented Apr 12, 2021

jameslamb left a comment

jameslamb Apr 12, 2021

TomAugspurger Apr 17, 2021

jameslamb Apr 12, 2021

hristog Apr 15, 2021

jameslamb Apr 15, 2021

jameslamb Apr 12, 2021

hristog Apr 15, 2021

jameslamb Apr 15, 2021

hristog Apr 15, 2021

jameslamb Apr 15, 2021

TomAugspurger Apr 17, 2021

jameslamb Apr 12, 2021

hristog Apr 15, 2021 •

edited

hristog commented Apr 12, 2021 •

edited

TomAugspurger left a comment

Make automatic centering in PCA methods optional #808

Are you sure you want to change the base?

Make automatic centering in PCA methods optional #808

Conversation

hristog commented Mar 20, 2021 • edited

hristog commented Mar 21, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

hristog commented Mar 24, 2021 • edited

hristog commented Apr 8, 2021

TomAugspurger commented Apr 12, 2021

hristog commented Apr 12, 2021

jameslamb left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

hristog Apr 15, 2021 • edited

Choose a reason for hiding this comment

hristog commented Apr 12, 2021 • edited

TomAugspurger left a comment

Choose a reason for hiding this comment

hristog commented Mar 20, 2021 •

edited

hristog commented Mar 24, 2021 •

edited

hristog Apr 15, 2021 •

edited

hristog commented Apr 12, 2021 •

edited