[OP] 3d conv + 3d pool #2585

antinucleon · 2016-06-30T18:58:19Z

Important changes:

Pooling shape: Need to manually change pooling pad to work with old inception models (done).
C++11 flags: Break support of CUDA 7.0, 6.5

piiswrong · 2016-06-30T23:39:49Z

breaking all previous models seems really bad...

antinucleon · 2016-06-30T23:40:58Z

Only break Inception family. I have fixed symbols and pretrained model.

piiswrong · 2016-06-30T23:41:53Z

does it affect vgg?

antinucleon · 2016-06-30T23:43:30Z

I think vgg is ok

antinucleon · 2016-07-01T00:59:43Z

@tqchen @piiswrong please check this. If it is ok we can merge this first because it fix travis.

piiswrong · 2016-07-01T18:45:42Z

example/image-classification/symbol_inception-bn-28-small.py

@@ -17,7 +17,7 @@ def DownsampleFactory(data, ch_3x3, mirror_attr):
    # conv 3x3
    conv = ConvFactory(data=data, kernel=(3, 3), stride=(2, 2), num_filter=ch_3x3, pad=(1, 1), mirror_attr=mirror_attr)
    # pool
-    pool = mx.symbol.Pooling(data=data, kernel=(3, 3), stride=(2, 2), pool_type='max', attr=mirror_attr)
+    pool = mx.symbol.Pooling(data=data, kernel=(3, 3), stride=(2, 2), pad=(1, 1), pool_type='max', attr=mirror_attr)


what about inception v3?

tqchen · 2016-07-05T17:31:05Z

what is the status on this PR?

antinucleon · 2016-07-06T17:44:35Z

Will do pairtest with Torch today

tqchen · 2016-07-06T17:45:44Z

please let me know if it is ready to merg

piiswrong · 2016-07-07T19:18:12Z

@tqchen looks like dmlc/any.h broke windows build

tqchen · 2016-07-07T19:33:56Z

will push a fix

antinucleon · 2016-07-07T20:02:34Z

I am working on CUDNN tuning. Will be finished very soon.

antinucleon · 2016-07-07T21:31:08Z

@piiswrong @tqchen It is done. There is a possible refactor, but I think we can refactor it after we fix deepmark

piiswrong · 2016-07-07T22:40:04Z

src/operator/cudnn_convolution.cc

@@ -11,6 +11,7 @@
 namespace mxnet {
 namespace op {
 #if MXNET_USE_CUDNN == 1
+// TODO(xxx): Refactor with Init CuDNN function, remove redandent code in initalization


I did this intentionally so that it would be easier to isolate this into it's own tune class

But many initialization code are exactly same, this part should be refactored.

tqchen · 2016-07-08T00:59:49Z

@antinucleon I have updated dmlc-core so that any should be compatible with windows build

tqchen · 2016-07-08T17:49:04Z

any updates?

antinucleon · 2016-07-08T17:58:19Z

Wait me 1 hour,
On Fri, Jul 8, 2016 at 10:49 Tianqi Chen [email protected] wrote:

any updates?

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#2585 (comment), or mute
the thread
https://github.com/notifications/unsubscribe/ABM13vRAd8Yx-313y7egViruPunq_tDRks5qTo2TgaJpZM4JCfHe
.

Sent from mobile phone

antinucleon · 2016-07-08T20:58:45Z

@tqchen If it is OK, merge it then I will update pretrained model

Godricly · 2016-07-15T07:27:11Z

src/operator/pooling-inl.h

+        oshape[2] = 1;
+        oshape[3] = 1;
+      } else {
+        oshape[2] = 1 + (dshape[2] + 2 * param_.pad[0] - param_.kernel[0]) / param_.stride[0];


Hi, I got some problem when testing old code with new pooling interface as the datatype used here is unsigned int. Maybe we need some more fool proof stuff here. This issue can be fixed if I claim it as global_pool.

given input shape of 6 and kernel size 7. it can exit with alert of "kernel size exceed input".

given input shape of 6 and kernel size 8, you can sneak through this check with 4294967296(max value of unsigned int).

Lorrainexun · 2016-08-11T05:59:50Z

It still cant work well with inception-BN model. COULD anyone fix this? It really cause me headache... Thanks a lot!

antinucleon · 2016-08-11T06:33:28Z

@Lorrainexun I have fixed the pretrained model symbol. It is because change of pooling shape.

WesleyZhang1991 · 2016-12-31T03:57:55Z

Hi,all. I compiled the latest mxnet with cudnn 5 and thus can use 3D conv. The vgg and alexnet in deepmark are ok. However, when I try to run c3d.py I met a problem:

mxnet.base.MXNetError: Error in operator pooling4: [11:50:48] src/operator/./pooling-inl.h:232: Check failed: param_.kernel[0] < dshape[2] + 2 * param_.pad[
0] (2 vs. 2) kernel size exceeds input

This can be seen whenever I want to collapse the last two frames into one in other C3D models.
Any suggestions? Thanks!

Bing Xu added 5 commits June 28, 2016 17:17

cudnn 3d conv

3ad28df

3d pool

48fdf7f

update

3f71d7c

Merge remote-tracking branch 'upstream/master' into 3dconv

0b5a6c3

remove ualberta

30b20db

antinucleon changed the title ~~[WIP] 3d conv + 3d pool~~ 3d conv + 3d pool Jun 30, 2016

antinucleon changed the title ~~3d conv + 3d pool~~ [OP] 3d conv + 3d pool Jun 30, 2016

fix travis

9fc84c6

update scala test

2e4f92b

piiswrong mentioned this pull request Jul 1, 2016

Problem with image classification examples #2588

Closed

piiswrong reviewed Jul 1, 2016
View reviewed changes

fix travis

d2b462e

pass pairtest

4c39be7

add tune

e90ca13

piiswrong reviewed Jul 7, 2016
View reviewed changes

update dmlc-core

43b4369

antinucleon merged commit 5532511 into apache:master Jul 8, 2016

Godricly reviewed Jul 15, 2016
View reviewed changes

Godricly mentioned this pull request Jul 16, 2016

DType dropout #2711

Closed

This was referenced Jul 17, 2016

predict-with-pretrained-model.ipynb unable to run #2718

Closed

error of "Incorrect shape" in example/cpp/image-classification-predict.cc #2709

Closed

lbin mentioned this pull request Jul 18, 2016

Incorrect shape error phunterlau/kaggle_statefarm#1

Open

antinucleon deleted the 3dconv branch July 21, 2016 18:16

erogol mentioned this pull request Jul 29, 2016

Do you plan to update pretrained networks following the latest changes? #2876

Closed

miguelgfierro mentioned this pull request Oct 2, 2016

Training of Imagenet with inception-bn in python gets good training accuracy but low validation after epoch 1 #3424

Closed

marcociccone mentioned this pull request Oct 17, 2016

Compilation Error with the latest version #3169

Closed

tqchen mentioned this pull request Jan 5, 2017

mxnet 0.9.1 seems to be slower and use much more memory #4525

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[OP] 3d conv + 3d pool #2585

[OP] 3d conv + 3d pool #2585

antinucleon commented Jun 30, 2016 •

edited

Loading

piiswrong commented Jun 30, 2016

antinucleon commented Jun 30, 2016

piiswrong commented Jun 30, 2016

antinucleon commented Jun 30, 2016

antinucleon commented Jul 1, 2016

piiswrong Jul 1, 2016

antinucleon Jul 6, 2016

tqchen commented Jul 5, 2016

antinucleon commented Jul 6, 2016

tqchen commented Jul 6, 2016

piiswrong commented Jul 7, 2016

tqchen commented Jul 7, 2016

antinucleon commented Jul 7, 2016

antinucleon commented Jul 7, 2016 •

edited

Loading

piiswrong Jul 7, 2016

antinucleon Jul 8, 2016

tqchen commented Jul 8, 2016

tqchen commented Jul 8, 2016

antinucleon commented Jul 8, 2016

antinucleon commented Jul 8, 2016

Godricly Jul 15, 2016

Lorrainexun commented Aug 11, 2016

antinucleon commented Aug 11, 2016

WesleyZhang1991 commented Dec 31, 2016

[OP] 3d conv + 3d pool #2585

[OP] 3d conv + 3d pool #2585

Conversation

antinucleon commented Jun 30, 2016 • edited Loading

piiswrong commented Jun 30, 2016

antinucleon commented Jun 30, 2016

piiswrong commented Jun 30, 2016

antinucleon commented Jun 30, 2016

antinucleon commented Jul 1, 2016

piiswrong Jul 1, 2016

Choose a reason for hiding this comment

antinucleon Jul 6, 2016

Choose a reason for hiding this comment

tqchen commented Jul 5, 2016

antinucleon commented Jul 6, 2016

tqchen commented Jul 6, 2016

piiswrong commented Jul 7, 2016

tqchen commented Jul 7, 2016

antinucleon commented Jul 7, 2016

antinucleon commented Jul 7, 2016 • edited Loading

piiswrong Jul 7, 2016

Choose a reason for hiding this comment

antinucleon Jul 8, 2016

Choose a reason for hiding this comment

tqchen commented Jul 8, 2016

tqchen commented Jul 8, 2016

antinucleon commented Jul 8, 2016

antinucleon commented Jul 8, 2016

Godricly Jul 15, 2016

Choose a reason for hiding this comment

Lorrainexun commented Aug 11, 2016

antinucleon commented Aug 11, 2016

WesleyZhang1991 commented Dec 31, 2016

antinucleon commented Jun 30, 2016 •

edited

Loading

antinucleon commented Jul 7, 2016 •

edited

Loading