Added Halide backend support for deep learning layers #1186

dkurt · 2017-05-25T12:14:34Z

Implemented some layers using Halide language backend:

Layer	Covered all configurations	Test
Concat	by channels (axis = 1)	x
Convolution	x	x
Deconvolution	no group factor	x
(attachable) ReLU, TanH, Sigmoid, Abs, BNLL, Power, PReLU, Batch normalization, Scale	x	x
Element-wise addition, multiplication, maximum	x	x
Fully-connected	axis = 1	x
LRN (across and within channels)	x	x
Maximum unpooling with indices support	x	x
Maximum pooling	x	x
Average pooling	only whole kernels, no padding	x
SoftMax	axis == 1, no spatial size more than 1x1	x

Supported CPU only, float32 computations. No ahead-of-time compilation.
Added tutorial with installation guide and sample, tests for some layers and networks, performance tests.

Generalized manual scheduling of 5 networks into the layer-wise automatic scheduling.

Intel® Core™ i7-6700K CPU @ 4.00GHz x 8

Architecture	Manual scheduling	Automatic scheduling	Ratio
AlexNet	22.38ms	25.5ms	x0.87
ResNet-50	73.91ms	84.26ms	x0.87
SqueezeNet v1.1	8.21ms	6.44ms	x1.27
Inception-5h	37.06ms	32.01ms	x1.15
ENet @ 3x512x256	41.21ms	41.07ms	x1.00

For testing purpose have taken GoogLeNet network. L1 diff is 1.26789e-09. Efficiency is 33ms using Halide (automatic scheduling) against 34.89ms using MKL (x1.05).

TODO list:

Reproduce Halide installation by tutorial (Linux) halide/Halide@abef1d9
Reproduce Halide installation by tutorial (Windows): installation reproduced but SegFault appears (Some layers works with replacing vectorization by x variable to factor 4 (from 8 or 16). All right in case of no scheduling).
Cover more layers and configurations in tests
Write scheduling tutorial
Remove local paths
Figure out about supportBackend and layer configurations (axis, etc.)
Automatic scheduling with no more than 20% efficiency loss.

alalek · 2017-05-25T14:51:48Z

modules/dnn/perf/perf_halide_net.cpp

+    { \
+        net.forward(net.getLayerId(outputLayer)); \
+    } \
+    SANITY_CHECK_NOTHING();


Please use "static void" function instead. It is very hard to debug multiline macro.

Thanks! Replaced to static function.

alalek · 2017-05-25T14:52:22Z

modules/dnn/CMakeLists.txt

+  set(DNN_LIBRARIES
+    ${DNN_LIBRARIES}
+    ${HALIDE_LIBRARIES}
+  )


list(APPEND DNN_INCLUDE_DIRS ${HALIDE_INCLUDE_DIRS})

Eliminated CMakeLists changes at all.

alalek · 2017-05-25T14:54:45Z

modules/dnn/perf/perf_halide_net.cpp

+
+inline std::string getOpenCVExtraDir()
+{
+    return "/home/dkurtaev/opencv_extra/testdata/";


#1162 + findDataFile() function

Modified modules/dnn/perf/perf_main.cpp for it.

alalek · 2017-05-25T15:12:58Z

modules/dnn/tutorials/tutorial_dnn_halide.markdown

+Measured on Intel&reg; Core&trade; i7-6700K CPU @ 4.00GHz x 8.
+
+<html>
+  <table>


Please use markdown tables. Example. Tutorial

Fixed, thanks!

alalek · 2017-05-25T15:18:12Z

modules/dnn/test/test_halide_nets.cpp

+using namespace cv;
+using namespace dnn;
+
+void loadNet(const std::string& weights, const std::string& proto,


please use "static" or "anonymous namespace" for internal functions.

BTW, Where is #include directive in this file?

It's working somehow. Should I add it for prevent fails in different build cases?

vpisarev · 2017-06-09T18:38:03Z

modules/dnn/include/opencv2/dnn/dnn.hpp

+     */
+    enum Backend
+    {
+        DEFAULT,


I would rename them to DNN_BACKEND_DEFAULT, DNN_BACKEND_HALIDE, to make it more self-explanatory

vpisarev · 2017-06-09T18:39:31Z

modules/dnn/include/opencv2/dnn/dnn.hpp

+     */
+    enum HalideTarget
+    {
+        HALIDE_CPU


again, let's use DNN_TARGET_CPU, DNN_TARGET_GPU etc. I would not use the word "Halide" here, because other backends, like OpenVX or even the default one, can be potentially customized to use CPU, GPU or other device to run the networks

vpisarev · 2017-06-09T18:41:39Z

modules/dnn/include/opencv2/dnn/dnn.hpp

+         * it helps prevent some memory management issues (if something wrong,
+         * Halide tests will be failed).
+         */
+        virtual std::vector<void*> initHalide(const std::vector<void*> &inputs);


instead of very low-level void* I would use some abstract class HalideFunc (that can encapsulate void*) with a smart pointer to it. It's comparatively low overhead, but it would be much more safe from memory management point of view. That is, what about changing std::vector<void*> to std::vector<Ptr<HalideFunc>>?

Introduced BackendNode and BackendWrapper whose derivatives store explicit Halide::Func and Halide::Buffer correspondingly. Moreover, it helps remove Halide::Func members from layers as it was before.

vpisarev · 2017-06-15T19:04:29Z

ok, let's merge it in while it's fresh! 👍

alalek · 2017-06-15T22:48:00Z

This PR is rely on opencv/opencv#8794 . Need to review/merge them too.

dkurt force-pushed the halide_support branch 2 times, most recently from 71cd4b3 to 3ee3f41 Compare May 25, 2017 14:06

alalek reviewed May 25, 2017

View reviewed changes

dkurt force-pushed the halide_support branch 6 times, most recently from 226db43 to e9b2d2f Compare May 29, 2017 15:13

bhack mentioned this pull request May 31, 2017

Framework Adoption Status dmlc/dlpack#10

Closed

dkurt force-pushed the halide_support branch 14 times, most recently from 1070e98 to 17e9f46 Compare June 8, 2017 10:33

vpisarev reviewed Jun 9, 2017

View reviewed changes

dkurt force-pushed the halide_support branch 3 times, most recently from 2c28d87 to aedce13 Compare June 14, 2017 14:00

dkurt force-pushed the halide_support branch 8 times, most recently from fc18830 to 2e94270 Compare June 15, 2017 14:27

Added Halide backend support for deep learning layers

dc93eed

dkurt force-pushed the halide_support branch from 2e94270 to dc93eed Compare June 15, 2017 14:59

vpisarev self-assigned this Jun 15, 2017

opencv-pushbot merged commit dc93eed into opencv:master Jun 15, 2017

opencv-pushbot pushed a commit that referenced this pull request Jun 15, 2017

Merge pull request #1186 from dkurt:halide_support

d95053f

dkurt deleted the halide_support branch June 21, 2017 13:47

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Added Halide backend support for deep learning layers #1186

Added Halide backend support for deep learning layers #1186

dkurt commented May 25, 2017 •

edited

Loading

alalek May 25, 2017

dkurt May 26, 2017

alalek May 25, 2017

dkurt May 26, 2017

alalek May 25, 2017

dkurt May 26, 2017

alalek May 25, 2017

dkurt May 26, 2017

alalek May 25, 2017

dkurt May 26, 2017

vpisarev Jun 9, 2017

vpisarev Jun 9, 2017

vpisarev Jun 9, 2017 •

edited

Loading

dkurt Jun 15, 2017

vpisarev commented Jun 15, 2017

alalek commented Jun 15, 2017

Added Halide backend support for deep learning layers #1186

Added Halide backend support for deep learning layers #1186

Conversation

dkurt commented May 25, 2017 • edited Loading

Implemented some layers using Halide language backend:

Generalized manual scheduling of 5 networks into the layer-wise automatic scheduling.

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

vpisarev Jun 9, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

vpisarev commented Jun 15, 2017

alalek commented Jun 15, 2017

dkurt commented May 25, 2017 •

edited

Loading

vpisarev Jun 9, 2017 •

edited

Loading