Skip to content

/ brainiak

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Distributed srm #220

Merged
merged 14 commits into from May 11, 2017
Merged

Distributed srm #220

merged 14 commits into from May 11, 2017

Conversation

@mjanderson09
Copy link
Contributor

@mjanderson09 mjanderson09 commented Apr 28, 2017

No description provided.

@mjanderson09
switched to distributed srm
caae62a
@mjanderson09
passes test for -n 3
65fe066
@mjanderson09
passes pr-check
8ff2ef7
@mjanderson09
Added distributed example
Loading status checks…
f8d05bd
@mjanderson09
pr changes
Loading status checks…
cf8d818
@mihaic
mihaic requested changes Apr 28, 2017
View changes
brainiak/funcalign/srm.py Outdated
@@ -132,6 +144,8 @@ class SRM(BaseEstimator, TransformerMixin):
rho2_ : array, shape=[subjects]
The estimated noise variance :math:`\\rho_i^2` for each subject
comm_ : mpi4py.MPI.Intracomm
The MPI communicator containing the data

This comment has been minimized.

Sign in to view
@mihaic

mihaic Apr 28, 2017
Contributor

comm should be used instead of comm_ everywhere, because it is given as a parameter, not computed from data:
http://scikit-learn.org/stable/developers/contributing.html#instantiation
http://scikit-learn.org/stable/developers/contributing.html#estimated-attributes
brainiak/funcalign/srm.py Outdated
for subject in range(subjects):
mu.append(np.mean(data[subject], 1))
np.random.seed(subject)

This comment has been minimized.

Sign in to view
@mihaic

mihaic Apr 28, 2017
Contributor

How does this interact with self.rand_seed?

This comment has been minimized.

Sign in to view
@mihaic

mihaic Apr 28, 2017
Contributor

Not addressed.

This comment has been minimized.

Sign in to view
@TuKo

TuKo Apr 28, 2017
Contributor

Check issue #114
Is it right to initialize the distributed SRM randomly?

This comment has been minimized.

Sign in to view
@mjanderson09

mjanderson09 May 1, 2017
Author Contributor

I removed this random seed per subject code. Instead, each SRM class instantiation can set a random seed (as before). This means for multiple ranks, each rank can potentially set a different seed. I don't think we can guarantee exactly deterministic results for the distributed implementation due to differences in parallel scheduling per run.
tests/funcalign/test_srm.py Outdated
@@ -191,3 +223,4 @@ def test_det_srm():
print("Test: different number of samples per subject")


test_can_instantiate()

This comment has been minimized.

Sign in to view
@mihaic

mihaic Apr 28, 2017
Contributor

Wrap this call in if __name__ == "__main__", please.
@mihaic mihaic requested a review from TuKo Apr 28, 2017
mjanderson09 added 2 commits May 1, 2017
@mjanderson09
one random seed per rank
Loading status checks…
cb4e6dd
@mjanderson09
fixed over indent
Loading status checks…
6a950a3
@mihaic
Copy link
Contributor

@mihaic mihaic commented May 1, 2017

Jenkins, retest this please.
1 similar comment
@mihaic
Copy link
Contributor

@mihaic mihaic commented May 2, 2017

Jenkins, retest this please.
@mihaic
mihaic approved these changes May 2, 2017
View changes
@TuKo
Copy link
Contributor

@TuKo TuKo commented May 2, 2017

@mjanderson09 great job!
Please, consider adding the TS-SVD and TS-QR code as well. We should allow for reproducibility of the IEEE BigData algorithm. Also, all SRM flavors will benefit from it, and some other methods as well.

Don't need to do it now. You can do it in a different PR.

Thank you!
mjanderson09 added 2 commits May 3, 2017
@mjanderson09
Added old srm test
Loading status checks…
2fea955
@mjanderson09
Gets same answer as v0_4
Loading status checks…
120fc09
@mihaic
mihaic requested changes May 8, 2017
View changes
tests/funcalign/test_srm.py Outdated
@@ -62,6 +63,9 @@ def test_can_instantiate():

# Check that runs with 2 subject
s.fit(X)
sr_v0_4 = np.load('sr_v0_4.npy')
diff_v0_4 = np.linalg.norm(sr_v0_4 - s.s_)
assert(diff_v0_4 < 1e-5)

This comment has been minimized.

Sign in to view
@mihaic

mihaic May 8, 2017
Contributor

Shouldn't the results be element-wise equal, i.e., np.allclose(sr_v0_4, s.s_), maybe with an atol value?
@mihaic
mihaic requested changes May 8, 2017
View changes
tests/funcalign/test_srm.py Outdated
@@ -62,6 +63,9 @@ def test_can_instantiate():

# Check that runs with 2 subject
s.fit(X)
sr_v0_4 = np.load('sr_v0_4.npy')

This comment has been minimized.

Sign in to view
@mihaic

mihaic May 8, 2017
Contributor

Add this to imports:
from pathlib import Path

Then:
Path(__file__).parent / "sr_v0_4.npy"
@mjanderson09
np.allclose and fixed file path
Loading status checks…
af5a756
@mihaic
mihaic approved these changes May 10, 2017
View changes
@mihaic
Copy link
Contributor

@mihaic mihaic commented May 10, 2017

@TuKo, what say you?
@TuKo
Copy link
Contributor

@TuKo TuKo commented May 10, 2017

@TuKo says NIPS, sorry...
@mjanderson09
merging random number PR
Loading status checks…
59d5503
@mjanderson09
new npy file
Loading status checks…
742d1b2
@mjanderson09
changed random number generation
Loading status checks…
fc595fb
@mjanderson09
switched to npz
Loading status checks…
1ee3e5b
@mihaic
mihaic requested changes May 11, 2017
View changes
tests/funcalign/test_srm.py Outdated
@@ -190,4 +194,4 @@ def test_det_srm():
model.fit(X)
print("Test: different number of samples per subject")


test_can_instantiate()

This comment has been minimized.

Sign in to view
@mihaic

mihaic May 11, 2017
Contributor

Please guard this call with if __name__ == "__main__".
@mihaic
mihaic approved these changes May 11, 2017
View changes
@TuKo
Copy link
Contributor

@TuKo TuKo commented May 11, 2017 •
edited

👍 Great work!
@TuKo
TuKo approved these changes May 11, 2017
View changes
@mihaic mihaic merged commit 3e039bd into brainiak:master May 11, 2017
3 checks passed
3 checks passed
continuous-integration/travis-ci/pr The Travis CI build passed
Details
linux Build finished.
Details
macos Build finished.
Details
@mihaic
Copy link
Contributor

@mihaic mihaic commented May 11, 2017

For the record, here is the effect of the changes on the accuracy in the image prediction example:

v0.4:
SRM: The average accuracy among all subjects is 0.657143 +/- 0.058576
Det. SRM: The average accuracy among all subjects is 0.657143 +/- 0.070891

master (i.e., per-subject random number generators):
SRM: The average accuracy among all subjects is 0.664286 +/- 0.076097
Det. SRM: The average accuracy among all subjects is 0.673214 +/- 0.068721

distributed_srm:
SRM: The average accuracy among all subjects is 0.664286 +/- 0.076097
Det. SRM: The average accuracy among all subjects is 0.673214 +/- 0.068721
danielsuo pushed a commit that referenced this pull request Nov 16, 2017
@mehrdadn @pcmoritz
int and long should be treated similarly (#220)
199b4ef
danielsuo pushed a commit that referenced this pull request Nov 16, 2017
@Wapaul1 @pcmoritz
Added functionality for retrieving variables from control dependencies (
db72978 
#220)

* Added test for retriving variables from an optimizer

* Added comments to test

* Addressed comments

* Fixed travis bug

* Added fix to circular controls

* Added set for explored operations and duplicate prefix stripping

* Removed embeded ipython

* Removed prefix, use seperate graph for each network

* Removed redundant imports

* Addressed comments and added separate graph to initializer

* fix typos

* get rid of prefix in documentation
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@mihaic mihaic

@TuKo TuKo

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Linked issues

Successfully merging this pull request may close these issues.

None yet

3 participants
@mjanderson09 @mihaic @TuKo