Skip to content

Commit

Permalink
Merge branch 'master' into update_gtest_cmake
Browse files Browse the repository at this point in the history
  • Loading branch information
@daquexian
daquexian authored Jul 1, 2019
2 parents 02c694f + c12ff51 commit ecf04a3
Show file tree
Hide file tree
Showing 3 changed files with 60 additions and 31 deletions.
24 changes: 12 additions & 12 deletions README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -8,7 +8,7 @@

*Enjoy binary neural networks on mobile!*

[English](README.md) [中文](README_CN.md)
[[English](README.md)] [[Chinese/中文](README_CN.md)]

Join chat at [Gitter (English)](https://gitter.im/dabnn/dabnn) or QQ Group (Chinese, 1021964010, answer: nndab)

Expand All @@ -23,21 +23,21 @@ To our best knowledge, dabnn is the first highly-optimized binary neural network
Benchmark result on Google Pixel 1 (single thread):

```
2019-05-02 18:00:29
2019-05-06 10:36:48
Running data/local/tmp/dabnn_benchmark
Run on (4 X 1593.6 MHz CPU s)
***WARNING*** CPU scaling is enabled, the benchmark real time measurements may be noisy and will incur extra overhead.
--------------------------------------------------------------------
Benchmark Time CPU Iterations
--------------------------------------------------------------------
dabnn_5x5_256 3658193 ns 3636875 ns 192 <--- input: 14*14*256, kernel: 256*5*5*256, output: 14*14*256, padding: 2
dabnn_3x3_64 1285949 ns 1261826 ns 552 <--- input: 56*56*64, kernel: 64*3*3*64, output: 56*56*64, padding: 1
dabnn_3x3_128 988757 ns 981547 ns 721 <--- input: 28*28*128, kernel: 128*3*3*128, output: 28*28*128, padding: 1
dabnn_3x3_256 1018918 ns 1008007 ns 689 <--- input: 14*14*256, kernel: 256*3*3*256, output: 14*14*256, padding: 1
dabnn_3x3_256_s2 269234 ns 268085 ns 2613 <--- input: 14*14*256, kernel: 256*3*3*256, output: 7*7*256, padding: 1, stride: 2
dabnn_3x3_512 1226245 ns 1203749 ns 579 <--- input: 7* 7*512, kernel: 512*3*3*512, output: 7* 7*512, padding: 1
dabnn_bireal18_imagenet 61809506 ns 61056865 ns 10 <--- Bi-Real Net 18, 56.4% top-1 on ImageNet
dabnn_bireal18_imagenet_stem 43279353 ns 41533009 ns 14 <--- Bi-Real Net 18 with stem module (The network structure will be described in detail in the coming paper), 56.4% top-1 on ImageNet
dabnn_5x5_256 3661928 ns 3638192 ns 191 <--- input: 14*14*256, kernel: 256*5*5*256, output: 14*14*256, padding: 2
dabnn_3x3_64 1306391 ns 1281553 ns 546 <--- input: 56*56*64, kernel: 64*3*3*64, output: 56*56*64, padding: 1
dabnn_3x3_128 958388 ns 954754 ns 735 <--- input: 28*28*128, kernel: 128*3*3*128, output: 28*28*128, padding: 1
dabnn_3x3_256 975123 ns 969810 ns 691 <--- input: 14*14*256, kernel: 256*3*3*256, output: 14*14*256, padding: 1
dabnn_3x3_256_s2 268310 ns 267712 ns 2618 <--- input: 14*14*256, kernel: 256*3*3*256, output: 7*7*256, padding: 1, stride: 2
dabnn_3x3_512 1281832 ns 1253921 ns 588 <--- input: 7* 7*512, kernel: 512*3*3*512, output: 7* 7*512, padding: 1
dabnn_bireal18_imagenet 61920154 ns 61339185 ns 10 <--- Bi-Real Net 18, 56.4% top-1 on ImageNet
dabnn_bireal18_imagenet_stem 43294019 ns 41401923 ns 14 <--- Bi-Real Net 18 with stem module (The network structure will be described in detail in the coming paper), 56.4% top-1 on ImageNet
```

The following is the comparison between our dabnn and [Caffe](http://caffe.berkeleyvision.org) (full precision), [TensorFlow Lite](https://www.tensorflow.org/lite) (full precision) and [BMXNet](https://github.com/hpi-xnor/BMXNet) (binary). We surprisingly observe that BMXNet is even slower than the full precision TensorFlow Lite. It suggests that the potential of binary neural networks is far from exploited until our dabnn is published.
Expand All @@ -62,9 +62,9 @@ After conversion, the generated dabnn model can be deployed on ARM devices (e.g.

We publish two pretrained binary neural network models based on [Bi-Real Net](https://arxiv.org/abs/1808.00278) on ImageNet. More pretrained models will be published in the future.

* [Bi-Real Net 18](https://drive.google.com/uc?export=download&id=1Oau5CtFR9nWXmlBBU47Jg5ypMiIEMtvo), 56.4% top-1 on ImageNet, 61.8ms/image on Google Pixel 1 (single thread).
* Bi-Real Net 18, 56.4% top-1 on ImageNet, 61.3ms/image on Google Pixel 1 (single thread). [[dabnn](https://drive.google.com/uc?export=download&id=1Oau5CtFR9nWXmlBBU47Jg5ypMiIEMtvo)] [[ONNX](https://drive.google.com/uc?export=download&id=1Xp3HB51H6Nhl6e555ieJubVutQake5sR)]

* [Bi-Real Net 18 with Stem Module](https://drive.google.com/uc?export=download&id=1ArsirMdbtJ9lvHSjc1hkQ7dIXDKh-D1t), 56.4% top-1 on ImageNet, 43.2ms/image on Google Pixel 1 (single thread). The detailed network structure will be described in the coming paper.
* Bi-Real Net 18 with Stem Module, 56.4% top-1 on ImageNet, 43.2ms/image on Google Pixel 1 (single thread). The detailed network structure will be described in the coming paper. [[dabnn](https://drive.google.com/uc?export=download&id=1ArsirMdbtJ9lvHSjc1hkQ7dIXDKh-D1t)] [[ONNX](https://drive.google.com/uc?export=download&id=1zu48CFptAGZ91IDCBPJSPM0bxDuPm9HS)]

## Implementation Details

Expand Down
24 changes: 12 additions & 12 deletions README_CN.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -6,7 +6,7 @@
[![Gitter Chat](https://img.shields.io/gitter/room/dabnn/dabnn.svg)](https://gitter.im/dabnn/dabnn)
[![PRs Welcome](https://img.shields.io/badge/PRs-welcome-brightgreen.svg)](https://github.com/JDAI-CV/dabnn/pulls)

[English](README.md) [中文](README_CN.md)
[[English](README.md)] [[Chinese/中文](README_CN.md)]

Gitter (English): [dabnn/dabnn](https://gitter.im/dabnn/dabnn)

Expand All @@ -25,21 +25,21 @@ QQ 群：1021964010, 入群答案: nndab
Google Pixel 1 上的 benchmark (单线程):

```
2019-05-02 18:00:29
2019-05-06 10:36:48
Running data/local/tmp/dabnn_benchmark
Run on (4 X 1593.6 MHz CPU s)
***WARNING*** CPU scaling is enabled, the benchmark real time measurements may be noisy and will incur extra overhead.
--------------------------------------------------------------------
Benchmark Time CPU Iterations
--------------------------------------------------------------------
dabnn_5x5_256 3658193 ns 3636875 ns 192 <--- input: 14*14*256, kernel: 256*5*5*256, output: 14*14*256, padding: 2
dabnn_3x3_64 1285949 ns 1261826 ns 552 <--- input: 56*56*64, kernel: 64*3*3*64, output: 56*56*64, padding: 1
dabnn_3x3_128 988757 ns 981547 ns 721 <--- input: 28*28*128, kernel: 128*3*3*128, output: 28*28*128, padding: 1
dabnn_3x3_256 1018918 ns 1008007 ns 689 <--- input: 14*14*256, kernel: 256*3*3*256, output: 14*14*256, padding: 1
dabnn_3x3_256_s2 269234 ns 268085 ns 2613 <--- input: 14*14*256, kernel: 256*3*3*256, output: 7*7*256, padding: 1, stride: 2
dabnn_3x3_512 1226245 ns 1203749 ns 579 <--- input: 7* 7*512, kernel: 512*3*3*512, output: 7* 7*512, padding: 1
dabnn_bireal18_imagenet 61809506 ns 61056865 ns 10 <--- Bi-Real Net 18, ImageNet top-1 为 56.4%
dabnn_bireal18_imagenet_stem 43279353 ns 41533009 ns 14 <--- 带有 stem 模块的 Bi-Real Net 18 (将在论文中描述), ImageNet top-1 为 56.4%
dabnn_5x5_256 3661928 ns 3638192 ns 191 <--- input: 14*14*256, kernel: 256*5*5*256, output: 14*14*256, padding: 2
dabnn_3x3_64 1306391 ns 1281553 ns 546 <--- input: 56*56*64, kernel: 64*3*3*64, output: 56*56*64, padding: 1
dabnn_3x3_128 958388 ns 954754 ns 735 <--- input: 28*28*128, kernel: 128*3*3*128, output: 28*28*128, padding: 1
dabnn_3x3_256 975123 ns 969810 ns 691 <--- input: 14*14*256, kernel: 256*3*3*256, output: 14*14*256, padding: 1
dabnn_3x3_256_s2 268310 ns 267712 ns 2618 <--- input: 14*14*256, kernel: 256*3*3*256, output: 7*7*256, padding: 1, stride: 2
dabnn_3x3_512 1281832 ns 1253921 ns 588 <--- input: 7* 7*512, kernel: 512*3*3*512, output: 7* 7*512, padding: 1
dabnn_bireal18_imagenet 61920154 ns 61339185 ns 10 <--- Bi-Real Net 18, ImageNet top-1 为 56.4%
dabnn_bireal18_imagenet_stem 43294019 ns 41401923 ns 14 <--- 带有 stem 模块的 Bi-Real Net 18 (将在论文中描述), ImageNet top-1 为 56.4%
```

在 Google Pixel 1 上与 [Caffe](http://caffe.berkeleyvision.org)（全精度）, [TensorFlow Lite](https://www.tensorflow.org/lite)（全精度）和 [BMXNet](https://github.com/hpi-xnor/BMXNet)（二值）的对比如下。我们很惊讶的发现现有的二值 inference 框架 BMXNet 甚至比全精度的 TensorFlow Lite 还要慢，这表明，直到 dabnn 推出之前，二值网络的潜力都远远没有被挖掘出来。
Expand All @@ -64,9 +64,9 @@ dabnn_bireal18_imagenet_stem 43279353 ns 41533009 ns 14 <---

我们提供两个在 ImageNet 上训练的、基于 [Bi-Real Net](https://arxiv.org/abs/1808.00278) 的二值网络模型，将来还会有其它的模型发布。

* [Bi-Real Net 18](https://drive.google.com/uc?export=download&id=1Oau5CtFR9nWXmlBBU47Jg5ypMiIEMtvo), ImageNet top-1 为 56.4%, 在 Google Pixel 1 上的速度为 61.8ms/image （单线程）。
* Bi-Real Net 18, ImageNet top-1 为 56.4%, 在 Google Pixel 1 上的速度为 61.8ms/image （单线程）。[[dabnn](https://drive.google.com/uc?export=download&id=1Oau5CtFR9nWXmlBBU47Jg5ypMiIEMtvo)] [[ONNX](https://drive.google.com/uc?export=download&id=1Xp3HB51H6Nhl6e555ieJubVutQake5sR)]

* [Bi-Real Net 18 with Stem Module](https://drive.google.com/uc?export=download&id=1ArsirMdbtJ9lvHSjc1hkQ7dIXDKh-D1t), ImageNet top-1 为 56.4%, 在 Google Pixel 1 上的速度为 43.2ms/image （单线程）。详细的网络结构将在论文中描述。
* Bi-Real Net 18 with Stem Module, ImageNet top-1 为 56.4%, 在 Google Pixel 1 上的速度为 43.2ms/image （单线程）。详细的网络结构将在论文中描述。 [[dabnn](https://drive.google.com/uc?export=download&id=1ArsirMdbtJ9lvHSjc1hkQ7dIXDKh-D1t)] [[ONNX](https://drive.google.com/uc?export=download&id=1zu48CFptAGZ91IDCBPJSPM0bxDuPm9HS)]

## 技术细节

Expand Down
43 changes: 36 additions & 7 deletions tools/onnx2bnn/onnx2bnn.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -6,33 +6,62 @@
#include <string>

#include <common/StrKeyMap.h>
#include <common/argh.h>
#include <glog/logging.h>
#include "NodeAttrHelper.h"
#include "OnnxConverter.h"
#include <common/argh.h>
#include "common/log_helper.h"

using std::string;
using std::vector;

void usage(const std::string &filename) {
std::cout << "Usage: " << filename << " onnx_model output_filename" << std::endl;
std::cout
<< "Usage: " << filename
<< " onnx_model output_filename [--optimize strict|moderate|aggressive]"
<< std::endl;
std::cout << "Example: " << filename
<< " model.onnx model.dab (The optimization leval will be "
"\"aggressive\")"
<< std::endl;
std::cout << "Example: " << filename
<< " model.onnx model.dab --optimize strict (The optimization "
"level will be \"strict\")"
<< std::endl;
}

int main(int argc, char **argv) {
argh::parser cmdl(argc, argv);
argh::parser cmdl;
cmdl.add_param("optimize");
cmdl.parse(argc, argv);
google::InitGoogleLogging(cmdl[0].c_str());
FLAGS_alsologtostderr = true;
if (!cmdl(2)) {
usage(cmdl[0]);
return -1;
}
bnn::OnnxConverter::Level opt_level = bnn::OnnxConverter::Level::kModerate;
if (cmdl["strict"]) {
opt_level = bnn::OnnxConverter::Level::kStrict;
// flags like 'onnx2bnn --strict' is not supported now
for (const auto flag : cmdl.flags()) {
std::cout << "Invalid flag: " << flag << std::endl;
usage(cmdl[0]);
return -2;
}
if (cmdl["aggressive"]) {

const std::string opt_level_str =
cmdl("optimize").str().empty() ? "aggressive" : cmdl("optimize").str();

bnn::OnnxConverter::Level opt_level;
if (opt_level_str == "strict") {
opt_level = bnn::OnnxConverter::Level::kStrict;
} else if (opt_level_str == "moderate") {
opt_level = bnn::OnnxConverter::Level::kModerate;
} else if (opt_level_str == "aggressive") {
opt_level = bnn::OnnxConverter::Level::kAggressive;
} else {
std::cout << "Invalid optimization level: " << opt_level_str
<< std::endl;
usage(cmdl[0]);
return -3;
}

ONNX_NAMESPACE::ModelProto model_proto;
Expand Down

0 comments on commit ecf04a3

Please sign in to comment.