MAINT: randomly sample chunks for partial_fit calls #276
Conversation
|
I've had to reformat |
dask_ml/_partial.py
Outdated
|
|
||
| nblocks = len(x.chunks[0]) | ||
| order = list(range(nblocks)) | ||
| random.shuffle(order) |
There was a problem hiding this comment.
For reproducibility, Incremental should take a random_state and then use that for shuffling the data.
I also think the shuffling should be optionally enabled / disabled (not sure what the default should be) by a hyper parameter.
Finally, I'm not sure what the parameter name should be. Typically scikit-learn uses shuffle=True/False. We can either re-use that, or call it shuffle_blocks, and reserve shuffle for when we actually shuffle data between blocks.
There was a problem hiding this comment.
I think shuffling should be enabled by default. That's what sklearn default too, and it's what SGD theory points too.
I've used shuffle_blocks; that describes exactly what it does. We can use shuffle for full shuffling, and shuffle_each_block to shuffle in each block.
|
I have actually tested this now, and believe it's ready for merge. |
e955a83 to
e3627f0
Compare
|
This all seems fine to me. The sklearn-dev failures are interesting. Do we know if it is due to this PR or something upstream? |
|
I don't see how it could be this PR. All the failures are with |
|
OK. Fixing it separately here: #283
…On Thu, Jul 5, 2018 at 2:13 PM, Scott Sievert ***@***.***> wrote:
I don't see how it could be this PR. All the failures are with
RobustScaler, which doesn't call fit or Incremental.
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#276 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/AASszAuYbtB4cWg8BEY6QNhEEQOVjPDPks5uDldIgaJpZM4VBaAM>
.
|
|
There was a PR on |
|
That would help explain things :)
…On Thu, Jul 5, 2018 at 2:21 PM, Scott Sievert ***@***.***> wrote:
There was a PR on RobustScalar that was merged earlier this morning:
scikit-learn/scikit-learn#11308
<scikit-learn/scikit-learn#11308>.
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#276 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/AASszDBJ0_aXlMEz2Jdlu6neAFXhv3VNks5uDllAgaJpZM4VBaAM>
.
|
mrocklin
left a comment
There was a problem hiding this comment.
Small comment. Otherwise this looks fine to me.
My guess is that @TomAugspurger is out of contact for a few days. I'll wait until this afternoon before merging in case he wants to jump in with something.
tests/test_incremental.py
Outdated
| from dask.array.utils import assert_eq | ||
| from sklearn.base import clone | ||
| from sklearn.linear_model import SGDClassifier | ||
| import numpy.linalg as LA |
There was a problem hiding this comment.
The LA pronoun isn't very common. I'd prefer the fully qualified name if possible.
import numpy as np
np.linalg.svd(...)
dask_ml/wrappers.py
Outdated
| a single NumPy array, which may exhaust the memory of your worker. | ||
| You probably want to always specify `scoring`. | ||
| random_state : int or numpy.random.RandomState |
dask_ml/wrappers.py
Outdated
| random_state : int or numpy.random.RandomState | ||
| Random object that determines how to shuffle blocks. | ||
| shuffle_blocks : bool |
dask_ml/wrappers.py
Outdated
| Random object that determines how to shuffle blocks. | ||
| shuffle_blocks : bool | ||
| Whether to randomly shuffle the blocks or now |
There was a problem hiding this comment.
typo: now -> not.
Could you elaborate on what this means? Specifically, clarify shuffling the block order from shuffling within blocks (from shuffling between blocks).
|
Thanks for the review @mrocklin and @TomAugspurger. I've addressed those comments. |
|
Thanks! |
Closes #274.
A screenshot of different
partial_fitcalls:I only show the before for some comparison, though it's not exact (and was in the debug process). It is "almost" because it uses a fixed seed every time, not a random seed. It does not run through the blocks in sequential order as mentioned in the previous docs.