Skip to content
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.

Sign up for GitHub

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

Add MultiNodeBatchNormalization #106

Merged
merged 20 commits into from
Aug 24, 2017
Merged

Add MultiNodeBatchNormalization #106

Changes from 1 commit
Commits
Show all changes
20 commits
Select commit Hold shift + click to select a range
1b40b9e
Add MultiNodeBatchNormalizationFunction
iwiwi Aug 10, 2017
c6d7983
Add MultiNodeBatchNormalization
iwiwi Aug 10, 2017
71afd30
Add experimental flag
iwiwi Aug 10, 2017
24713de
Add __init__.py
iwiwi Aug 10, 2017
79b8b8f
Support CPU execution
iwiwi Aug 10, 2017
56621c2
Add tests for MultiNodeBatchNormalization
iwiwi Aug 10, 2017
12d8475
Fix PEP8
iwiwi Aug 10, 2017
9df2831
Fix PEP8
iwiwi Aug 15, 2017
74a3701
Add tests to see that this test is valid
iwiwi Aug 15, 2017
3a19dfa
Fix the test class name
iwiwi Aug 15, 2017
568c117
Add MultiNodeBatchNormalization to document
iwiwi Aug 15, 2017
3db296a
Raise RuntimeError when Chainer is version 1
iwiwi Aug 15, 2017
947563a
Add version checks to tests
iwiwi Aug 15, 2017
71dd347
Add checks to see that no exception raising
iwiwi Aug 16, 2017
bc0c2dc
Add a note about chainer version
iwiwi Aug 16, 2017
71b25cb
Add explanation of tests
iwiwi Aug 16, 2017
73efe61
Fix import error for Chainer v1
iwiwi Aug 17, 2017
629643e
Add comments to highlight the differences from original Chainer imple…
iwiwi Aug 17, 2017
680ce24
Fix PEP8
iwiwi Aug 17, 2017
37a5ab8
Fix the test by ignoring configurations with a single node
iwiwi Aug 23, 2017
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
Add comments to highlight the differences from original Chainer imple… 
…mentation
  • Loading branch information
@iwiwi
iwiwi committed Aug 17, 2017
commit 629643e7bdb7824acbf36ff7beeb22c641025ea6
9 changes: 8 additions & 1 deletion chainermn/functions/batch_normalization.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,6 @@
# This file is heavily based on Chainer's batch normalization implementation.
# See: chainer/functions/normalization/batch_normalization.py (dbb650)

import chainer
from chainer import cuda
from chainer import function
Expand Down Expand Up @@ -115,6 +118,7 @@ def forward(self, inputs):
if chainer.configuration.config.train:
axis = (0,) + tuple(range(head_ndim, x.ndim))

# ChainerMN diff (1/2) begins
mpi_comm = self.comm.mpi_comm
tmp = xp.empty(gamma.size * 2, dtype=x.dtype)
x.mean(axis=axis, out=tmp[:gamma.size])
Expand All @@ -129,6 +133,7 @@ def forward(self, inputs):
mean = tmp[:gamma.size]
sqmean = tmp[gamma.size:]
var = sqmean - xp.square(mean)
# ChainerMN diff (1/2) ends

var += self.eps
else:
Expand Down Expand Up @@ -194,7 +199,8 @@ def backward(self, inputs, grad_outputs):
# Note: If length of inputs is not 5, we must be in train mode.
assert chainer.configuration.config.train

# It is wrong to multiply m by mpi_comm.size
# ChainerMN diff (2/2) begins
# Note: It is wrong to multiply m by mpi_comm.size
# (instead of multiplying 1/size to gbeta, ggamma)
mpi_comm = self.comm.mpi_comm
tmp = xp.empty(gamma.size * 2, dtype=x.dtype)
Expand All @@ -208,6 +214,7 @@ def backward(self, inputs, grad_outputs):
tmp *= 1.0 / mpi_comm.size
gbeta = tmp[:gamma.size]
ggamma = tmp[gamma.size:]
# ChainerMN diff (2/2) ends

if xp is numpy:
gx = (gamma / self.std)[expander] * (
Expand Down