File size: 6,494 Bytes
9375c9a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
// Copyright (C) 2017  Davis E. King ([email protected])
// License: Boost Software License   See LICENSE.txt for the full license.

#include "opaque_types.h"
#include <dlib/python.h>
#include <dlib/matrix.h>
#include <dlib/dnn.h>
#include <dlib/image_transforms.h>
#include "indexing.h"
#include <pybind11/stl_bind.h>

using namespace dlib;
using namespace std;

namespace py = pybind11;


class cnn_face_detection_model_v1
{

public:

    cnn_face_detection_model_v1(const std::string& model_filename)
    {
        deserialize(model_filename) >> net;
    }

    std::vector<mmod_rect> detect (
        py::array pyimage,
        const int upsample_num_times
    )
    {
        pyramid_down<2> pyr;
        std::vector<mmod_rect> rects;

        // Copy the data into dlib based objects
        matrix<rgb_pixel> image;
        if (is_image<unsigned char>(pyimage))
            assign_image(image, numpy_image<unsigned char>(pyimage));
        else if (is_image<rgb_pixel>(pyimage))
            assign_image(image, numpy_image<rgb_pixel>(pyimage));
        else
            throw dlib::error("Unsupported image type, must be 8bit gray or RGB image.");

        // Upsampling the image will allow us to detect smaller faces but will cause the
        // program to use more RAM and run longer.
        unsigned int levels = upsample_num_times;
        while (levels > 0)
        {
            levels--;
            pyramid_up(image, pyr);
        }

        auto dets = net(image);

        // Scale the detection locations back to the original image size
        // if the image was upscaled.
        for (auto&& d : dets) {
            d.rect = pyr.rect_down(d.rect, upsample_num_times);
            rects.push_back(d);
        }

        return rects;
    }

    std::vector<std::vector<mmod_rect>> detect_mult (
        py::list imgs,
        const int upsample_num_times,
        const int batch_size = 128
    )
    {
        pyramid_down<2> pyr;
        std::vector<matrix<rgb_pixel>> dimgs;
        dimgs.reserve(len(imgs));

        for(int i = 0; i < len(imgs); i++)
        {
            // Copy the data into dlib based objects
            matrix<rgb_pixel> image;
            py::array tmp = imgs[i].cast<py::array>();
            if (is_image<unsigned char>(tmp))
                assign_image(image, numpy_image<unsigned char>(tmp));
            else if (is_image<rgb_pixel>(tmp))
                assign_image(image, numpy_image<rgb_pixel>(tmp));
            else
                throw dlib::error("Unsupported image type, must be 8bit gray or RGB image.");

            for(int i = 0; i < upsample_num_times; i++)
            {
                pyramid_up(image);
            }
            dimgs.emplace_back(std::move(image));
        }

        for(int i = 1; i < dimgs.size(); i++)
        {
            if (dimgs[i - 1].nc() != dimgs[i].nc() || dimgs[i - 1].nr() != dimgs[i].nr())
                throw dlib::error("Images in list must all have the same dimensions.");
            
        }        

        auto dets = net(dimgs, batch_size);
        std::vector<std::vector<mmod_rect> > all_rects;

        for(auto&& im_dets : dets)
        {
            std::vector<mmod_rect> rects;
            rects.reserve(im_dets.size());
            for (auto&& d : im_dets) {
                d.rect = pyr.rect_down(d.rect, upsample_num_times);
                rects.push_back(d);
            }
            all_rects.push_back(rects);
        }
        
        return all_rects;
    }

private:

    template <long num_filters, typename SUBNET> using con5d = con<num_filters,5,5,2,2,SUBNET>;
    template <long num_filters, typename SUBNET> using con5  = con<num_filters,5,5,1,1,SUBNET>;

    template <typename SUBNET> using downsampler  = relu<affine<con5d<32, relu<affine<con5d<32, relu<affine<con5d<16,SUBNET>>>>>>>>>;
    template <typename SUBNET> using rcon5  = relu<affine<con5<45,SUBNET>>>;

    using net_type = loss_mmod<con<1,9,9,1,1,rcon5<rcon5<rcon5<downsampler<input_rgb_image_pyramid<pyramid_down<6>>>>>>>>;

    net_type net;
};

// ----------------------------------------------------------------------------------------

void bind_cnn_face_detection(py::module& m)
{
    {
    py::class_<cnn_face_detection_model_v1>(m, "cnn_face_detection_model_v1", "This object detects human faces in an image.  The constructor loads the face detection model from a file. You can download a pre-trained model from http://dlib.net/files/mmod_human_face_detector.dat.bz2.")
        .def(py::init<std::string>(), py::arg("filename"))
        .def(
            "__call__", 
            &cnn_face_detection_model_v1::detect_mult, 
            py::arg("imgs"), py::arg("upsample_num_times")=0, py::arg("batch_size")=128, 
            "takes a list of images as input returning a 2d list of mmod rectangles"
            )
        .def(
            "__call__", 
            &cnn_face_detection_model_v1::detect, 
            py::arg("img"), py::arg("upsample_num_times")=0,
            "Find faces in an image using a deep learning model.\n\
          - Upsamples the image upsample_num_times before running the face \n\
            detector."
            );
    }

    m.def("set_dnn_prefer_smallest_algorithms", &set_dnn_prefer_smallest_algorithms, "Tells cuDNN to use slower algorithms that use less RAM.");

    auto cuda = m.def_submodule("cuda", "Routines for setting CUDA specific properties.");
    cuda.def("set_device", &dlib::cuda::set_device, py::arg("device_id"), 
        "Set the active CUDA device.  It is required that 0 <= device_id < get_num_devices().");
    cuda.def("get_device", &dlib::cuda::get_device, "Get the active CUDA device.");
    cuda.def("get_num_devices", &dlib::cuda::get_num_devices, "Find out how many CUDA devices are available.");

    {
    typedef mmod_rect type;
    py::class_<type>(m, "mmod_rectangle", "Wrapper around a rectangle object and a detection confidence score.")
        .def_readwrite("rect",   &type::rect)
        .def_readwrite("confidence", &type::detection_confidence);
    }
    {
    typedef std::vector<mmod_rect> type;
    py::bind_vector<type>(m, "mmod_rectangles", "An array of mmod rectangle objects.")
        .def("extend", extend_vector_with_python_list<mmod_rect>);
    }
    {
    typedef std::vector<std::vector<mmod_rect> > type;
    py::bind_vector<type>(m, "mmod_rectangless", "A 2D array of mmod rectangle objects.")
        .def("extend", extend_vector_with_python_list<std::vector<mmod_rect>>);
    } 
}