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

openpose example #3

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
Open

openpose example #3

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
57 changes: 57 additions & 0 deletions examples/example_openpose_plus_hao28.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,57 @@
// https://github.com/tensorlayer/openpose-plus/blob/master/openpose_plus/models/models_hao28_experimental.py

// #define STDNN_OPS_HAVE_CBLAS

#include <algorithm>
#include <string>

#include "example_openpose_plus_hao28.hpp"

using std::experimental::range;

int main(int argc, char *argv[])
{
const std::string home(std::getenv("HOME"));
const auto prefix = home + "/var/models/openpose";
// const auto filename =
// home +
// "/var/data/openpose/examples/media/COCO_val2014_000000000192.jpg";

openpose_plus_hao28 openpose(prefix);

// auto paf_runner =
// create_paf_processor(32, 48, openpose.h, openpose.w, 19, 19, 13);

auto x = ttl::tensor<float, 4>(1, openpose.h, openpose.w, 3);

// TODO: input images
// auto input = ttl::tensor<uint8_t, 4>(x.shape());
// cv::Mat resized_image(cv::Size(openpose.w, openpose.h), CV_8UC(3),
// input.data());
// {
// auto img = cv::imread(filename);
// cv::resize(img, resized_image, resized_image.size(), 0, 0);
// std::transform(input.data(), data_end(input), x.data(),
// [](uint8_t p) { return p / 255.0; });
// }

int repeats = 5;
for (auto i : std::experimental::range(repeats)) {
printf("inference %d\n", i);
auto [l_conf, l_paf] = openpose(ref(x));

// TODO: run paf process
// auto conf = nn::ops::apply<ttl::tensor<float, 4>>(
// nn::ops::to_channels_first(), ref(*l_conf));
// auto paf = nn::ops::apply<ttl::tensor<float, 4>>(
// nn::ops::to_channels_first(), ref(*l_paf));

// auto human = (*paf_runner)(conf.data(), paf.data(), false);
// for (auto h : human) {
// h.print();
// draw_human(resized_image, h);
// }
// cv::imwrite("a.png", resized_image);
}
return 0;
}
178 changes: 178 additions & 0 deletions examples/example_openpose_plus_hao28.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,178 @@
#pragma once
#include <nn/layers>
#include <nn/models>
#include <nn/ops>

// https://github.com/tensorlayer/openpose-plus/blob/master/openpose_plus/models/models_hao28_experimental.py
template <typename R> struct openpose_plus_hao28_impl {
using image_order = nn::ops::nhwc;
using filter_order = nn::ops::rscd;
using relu = nn::ops::pointwise<nn::ops::relu>;
using pool = nn::layers::pool<nn::ops::pool_max, image_order>;
using concat = nn::ops::concat_channel4d<image_order>;

mutable nn::models::namescope ns;
const std::string data_dir_;

auto f(const std::string &name) const
{
const auto full = data_dir_ + "/" + ns(name) + ".idx";
return nn::ops::readfile(full);
};

auto conv(int d, int k, int p, const std::string &name = "conv") const
{
return ns.with(name, [&] {
using conv_layer =
nn::layers::conv<image_order, filter_order, false>;
const auto l = conv_layer(conv_layer::ksize(k, k), d,
conv_layer::padding(p, p));
return with_init(l, f("kernel"));
});
}

auto conv_(int d, int k, int p, const std::string &name = "conv") const
{
return ns.with(name, [&] {
using conv_layer =
nn::layers::conv<image_order, filter_order, true>;
const auto l = conv_layer(conv_layer::ksize(k, k), d,
conv_layer::padding(p, p));
return with_init(l, f("kernel"), f("bias"));
});
}

auto bn(const std::string &name = "bn") const
{
return ns.with(name, [&] {
using relu = nn::ops::pointwise<nn::ops::relu>;
using bn_layer = nn::layers::batch_norm<image_order, relu>;
return with_init(bn_layer(), f("moving_mean"), f("moving_variance"),
f("beta"), f("gamma"));
});
}

public:
openpose_plus_hao28_impl(const std::string prefix) : data_dir_(prefix) {}

const int n_joins = 19;
const int n_connections = 19;

auto operator()(const ttl::tensor_ref<R, 4> &x)
{
return ns.with("model", [&] {
const auto fm = cnn(x);
auto p1 = stage1(*fm);
auto p5 = [&] {
return ns.with("stage5", [&] {
return stage2(*fm, *p1.first, *p1.second);
});
}();
auto p6 = [&] {
return ns.with("stage6", [&] {
return stage2(*fm, *p5.first, *p5.second);
});
}();
return p6;
});
}

private:
auto cnn(const ttl::tensor_ref<R, 4> &x)
{
auto conv_layers = //
nn::models::make_sequential() //
<< conv(32, 3, 1, "conv1_1") << bn("bn1_1") //
<< conv(64, 3, 1, "conv1_2") << bn("bn1_2") //
<< pool() //
<< conv(128, 3, 1, "conv2_1") << bn("bn2_1") //
<< conv(128, 3, 1, "conv2_2") << bn("bn2_2") //
<< pool() //
<< conv(200, 3, 1, "conv3_1") << bn("bn3_1") //
<< conv(200, 3, 1, "conv3_2") << bn("bn3_2") //
<< conv(200, 3, 1, "conv3_3") << bn("bn3_3") //
<< pool() //
<< conv(384, 3, 1, "conv4_1") << bn("bn4_1") //
<< conv(384, 3, 1, "conv4_2") << bn("bn4_2") //
<< conv(256, 3, 1, "conv4_3") << bn("bn4_3") //
<< conv(128, 3, 1, "conv4_4") << bn("bn4_4") //
;
return conv_layers(x);
}

auto stage1(const ttl::tensor_ref<R, 4> &x)
{
return ns.with("stage1", [&] {
auto common = [&] {
return //
nn::models::make_sequential() //
<< conv(128, 3, 1, "c1") << bn("bn1") //
<< conv(128, 3, 1, "c2") << bn("bn2") //
<< conv(128, 3, 1, "c3") << bn("bn3") //
<< conv(128, 1, 0, "c4") << bn("bn4") //
;
};
auto left = [&] {
return ns.with("branch1", [&] {
return common() << conv_(n_joins, 1, 0, "confs");
});
}();
auto right = [&] {
return ns.with("branch2", [&] {
return common() << conv_(2 * n_connections, 1, 0, "pafs");
});
}();
return std::make_pair(left(x), right(x));
});
}

auto stage2(const ttl::tensor_ref<R, 4> &x, //
const ttl::tensor_ref<R, 4> &b1,
const ttl::tensor_ref<R, 4> &b2)
{
using T = ttl::tensor<R, 4>;
auto net =
std::unique_ptr<T>(nn::ops::new_result<T>(concat(), x, b1, b2));
auto common = [&] {
return //
nn::models::make_sequential() //
<< conv(128, 3, 1, "c1") << bn("bn1") //
<< conv(128, 3, 1, "c2") << bn("bn2") //
<< conv(128, 3, 1, "c3") << bn("bn3") //
<< conv(128, 3, 1, "c4") << bn("bn4") //
<< conv(128, 3, 1, "c5") << bn("bn5") //
<< conv(128, 1, 0, "c6") << bn("bn6") //
;
};
auto left = [&] {
return ns.with("branch1", [&] {
return common() << conv_(n_joins, 1, 0, "conf");
});
}();
auto right = [&] {
return ns.with("branch2", [&] {
return common() << conv_(2 * n_connections, 1, 0, "pafs");
});
}();
return std::make_pair(left(ref(*net)), right(ref(*net)));
}
};

struct openpose_plus_hao28 {
const std::string data_dir_;

public:
const size_t h;
const size_t w;

openpose_plus_hao28(const std::string &data_dir, int height = 256,
int width = 384)
: data_dir_(data_dir), h(height), w(width)
{
}

template <typename R> auto operator()(const ttl::tensor_ref<R, 4> &x)
{
return openpose_plus_hao28_impl<R>(data_dir_)(x);
}
};
1 change: 0 additions & 1 deletion include/nn/bits/ops/concat.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,7 +33,6 @@ template <> class concat_channel4d_impl<nhwc, 3>

const auto c1 = c_1;
const auto c2 = c_2;
const auto c3 = c_3;

for (auto l : range(n)) {
for (auto i : range(h)) {
Expand Down