[SPARK-20214][ML] Make sure converted csc matrix has sorted indices #17532
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
viirya
force-pushed
the
make-sure-sorted-indices
branch
from
f88eb6b
to
10be94a
Compare
|
Test build #75528 has finished for PR 17532 at commit
|
|
Test build #75530 has finished for PR 17532 at commit
|
|
I don't think is directly related, but, I tried to make a similar change to the Spark Scala implementation of this type of conversion and couldn't make it work, just because it changed a bunch of little behaviors. (I could look up the JIRA.) This could be fine, being more narrow. |
jkbradley
reviewed
Apr 5, 2017
python/pyspark/ml/linalg/__init__.py
Outdated
| @@ -72,7 +72,9 @@ def _convert_to_vector(l): | |||
| return DenseVector(l) | |||
| elif _have_scipy and scipy.sparse.issparse(l): | |||
| assert l.shape[1] == 1, "Expected column vector" | |||
| csc = l.tocsc() | |||
| # Make sure the converted csc_matrix has sorted indices. | |||
| csc = l.tocsr().tocsc() | |||
There was a problem hiding this comment.
I'd just call tocsc. You can then check the attribute has_sorted_indices and call sort_indices() if needed.
|
Are you able to write a unit test which passes data through _convert_to_vector and fails before this fix? |
|
Btw, I'd really like to get this into 2.2, which will be cut soon. Let me know if you'd like me to take it over. Thanks! |
|
@jkbradley Thanks for comment. I will add the unit test now. |
|
@jkbradley An unit test is added. Please check this again. Thanks. |
viirya
commented
Apr 5, 2017
| @@ -853,6 +853,17 @@ def serialize(l): | |||
| self.assertEqual(sv, serialize(lil.tocsr())) | |||
| self.assertEqual(sv, serialize(lil.todok())) | |||
|
|
|||
| def test_convert_to_vector(self): | |||
There was a problem hiding this comment.
We have no Scipy-related tests in ml submodule, so I don't add it. If we really need one, please let me know. I can add it quickly.
There was a problem hiding this comment.
No it's Ok; I made a separate JIRA to port the tests there.
|
Test build #75560 has finished for PR 17532 at commit
|
|
LGTM |
asfgit
pushed a commit
that referenced
this pull request
Apr 6, 2017
## What changes were proposed in this pull request?
`_convert_to_vector` converts a scipy sparse matrix to csc matrix for initializing `SparseVector`. However, it doesn't guarantee the converted csc matrix has sorted indices and so a failure happens when you do something like that:
from scipy.sparse import lil_matrix
lil = lil_matrix((4, 1))
lil[1, 0] = 1
lil[3, 0] = 2
_convert_to_vector(lil.todok())
File "/home/jenkins/workspace/python/pyspark/mllib/linalg/__init__.py", line 78, in _convert_to_vector
return SparseVector(l.shape[0], csc.indices, csc.data)
File "/home/jenkins/workspace/python/pyspark/mllib/linalg/__init__.py", line 556, in __init__
% (self.indices[i], self.indices[i + 1]))
TypeError: Indices 3 and 1 are not strictly increasing
A simple test can confirm that `dok_matrix.tocsc()` won't guarantee sorted indices:
>>> from scipy.sparse import lil_matrix
>>> lil = lil_matrix((4, 1))
>>> lil[1, 0] = 1
>>> lil[3, 0] = 2
>>> dok = lil.todok()
>>> csc = dok.tocsc()
>>> csc.has_sorted_indices
0
>>> csc.indices
array([3, 1], dtype=int32)
I checked the source codes of scipy. The only way to guarantee it is `csc_matrix.tocsr()` and `csr_matrix.tocsc()`.
## How was this patch tested?
Existing tests.
Please review http://spark.apache.org/contributing.html before opening a pull request.
Author: Liang-Chi Hsieh <viirya@gmail.com>
Closes #17532 from viirya/make-sure-sorted-indices.
(cherry picked from commit 1220605)
Signed-off-by: Joseph K. Bradley <joseph@databricks.com>
asfgit
pushed a commit
that referenced
this pull request
Apr 6, 2017
## What changes were proposed in this pull request?
`_convert_to_vector` converts a scipy sparse matrix to csc matrix for initializing `SparseVector`. However, it doesn't guarantee the converted csc matrix has sorted indices and so a failure happens when you do something like that:
from scipy.sparse import lil_matrix
lil = lil_matrix((4, 1))
lil[1, 0] = 1
lil[3, 0] = 2
_convert_to_vector(lil.todok())
File "/home/jenkins/workspace/python/pyspark/mllib/linalg/__init__.py", line 78, in _convert_to_vector
return SparseVector(l.shape[0], csc.indices, csc.data)
File "/home/jenkins/workspace/python/pyspark/mllib/linalg/__init__.py", line 556, in __init__
% (self.indices[i], self.indices[i + 1]))
TypeError: Indices 3 and 1 are not strictly increasing
A simple test can confirm that `dok_matrix.tocsc()` won't guarantee sorted indices:
>>> from scipy.sparse import lil_matrix
>>> lil = lil_matrix((4, 1))
>>> lil[1, 0] = 1
>>> lil[3, 0] = 2
>>> dok = lil.todok()
>>> csc = dok.tocsc()
>>> csc.has_sorted_indices
0
>>> csc.indices
array([3, 1], dtype=int32)
I checked the source codes of scipy. The only way to guarantee it is `csc_matrix.tocsr()` and `csr_matrix.tocsc()`.
## How was this patch tested?
Existing tests.
Please review http://spark.apache.org/contributing.html before opening a pull request.
Author: Liang-Chi Hsieh <viirya@gmail.com>
Closes #17532 from viirya/make-sure-sorted-indices.
(cherry picked from commit 1220605)
Signed-off-by: Joseph K. Bradley <joseph@databricks.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What changes were proposed in this pull request?
_convert_to_vectorconverts a scipy sparse matrix to csc matrix for initializingSparseVector. However, it doesn't guarantee the converted csc matrix has sorted indices and so a failure happens when you do something like that:A simple test can confirm that
dok_matrix.tocsc()won't guarantee sorted indices:I checked the source codes of scipy. The only way to guarantee it is
csc_matrix.tocsr()andcsr_matrix.tocsc().How was this patch tested?
Existing tests.
Please review http://spark.apache.org/contributing.html before opening a pull request.