Skip to content

Commit

Permalink
mxnet-ssd wip ...
Browse files Browse the repository at this point in the history
  • Loading branch information
@nihui
nihui committed Nov 29, 2018
1 parent d9d7d6b commit c77383f
Show file tree
Hide file tree
Showing 5 changed files with 157 additions and 49 deletions.
8 changes: 6 additions & 2 deletions src/layer/detectionoutput.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -33,6 +33,10 @@ int DetectionOutput::load_param(const ParamDict& pd)
nms_top_k = pd.get(2, 300);
keep_top_k = pd.get(3, 100);
confidence_threshold = pd.get(4, 0.5f);
variances[0] = pd.get(5, 0.1f);
variances[1] = pd.get(6, 0.1f);
variances[2] = pd.get(7, 0.2f);
variances[3] = pd.get(8, 0.2f);

return 0;
}
Expand Down Expand Up @@ -161,14 +165,14 @@ int DetectionOutput::forward(const std::vector<Mat>& bottom_blobs, std::vector<M

const float* location_ptr = location;
const float* priorbox_ptr = priorbox.row(0);
const float* variance_ptr = priorbox.row(1);
const float* variance_ptr = priorbox.h == 2 ? priorbox.row(1) : 0;

#pragma omp parallel for num_threads(opt.num_threads)
for (int i = 0; i < num_prior; i++)
{
const float* loc = location_ptr + i * 4;
const float* pb = priorbox_ptr + i * 4;
const float* var = variance_ptr + i * 4;
const float* var = variance_ptr ? variance_ptr + i * 4 : variances;

float* bbox = bboxes.row(i);

Expand Down
1 change: 1 addition & 0 deletions src/layer/detectionoutput.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -34,6 +34,7 @@ class DetectionOutput : public Layer
int nms_top_k;
int keep_top_k;
float confidence_threshold;
float variances[4];
};

} // namespace ncnn
Expand Down
33 changes: 33 additions & 0 deletions src/layer/permute.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -38,6 +38,39 @@ int Permute::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) c
int channels = bottom_blob.c;
size_t elemsize = bottom_blob.elemsize;

int dims = bottom_blob.dims;

if (dims == 2)
{
// order_type
// 0 = w h
// 1 = h w

if (order_type == 0)
{
top_blob = bottom_blob;
}
else if (order_type == 1)
{
top_blob.create(h, w, elemsize, opt.blob_allocator);
if (top_blob.empty())
return -100;

const float* ptr = bottom_blob;
float* outptr = top_blob;

for (int i = 0; i < w; i++)
{
for (int j = 0; j < h; j++)
{
outptr[i*h + j] = ptr[j*w + i];
}
}
}

return 0;
}

// order_type
// 0 = w h c
// 1 = h w c
Expand Down
107 changes: 85 additions & 22 deletions src/layer/priorbox.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -31,10 +31,10 @@ int PriorBox::load_param(const ParamDict& pd)
min_sizes = pd.get(0, Mat());
max_sizes = pd.get(1, Mat());
aspect_ratios = pd.get(2, Mat());
variances[0] = pd.get(3, 0.f);
variances[1] = pd.get(4, 0.f);
variances[2] = pd.get(5, 0.f);
variances[3] = pd.get(6, 0.f);
variances[0] = pd.get(3, 0.1f);
variances[1] = pd.get(4, 0.1f);
variances[2] = pd.get(5, 0.2f);
variances[3] = pd.get(6, 0.2f);
flip = pd.get(7, 1);
clip = pd.get(8, 0);
image_width = pd.get(9, 0);
Expand All @@ -51,6 +51,83 @@ int PriorBox::forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& to
int w = bottom_blobs[0].w;
int h = bottom_blobs[0].h;

if (bottom_blobs.size() == 1 && image_width == -233 && image_height == -233 && max_sizes.empty())
{
// mxnet style _contrib_MultiBoxPrior
fprintf(stderr, "mxnet style _contrib_MultiBoxPrior\n");

float step_w = step_width;
float step_h = step_height;
if (step_w == -233)
step_w = 1.f / (float)w;
if (step_h == -233)
step_h = 1.f / (float)h;

int num_sizes = min_sizes.w;
int num_ratios = aspect_ratios.w;

int num_prior = num_sizes - 1 + num_ratios;

Mat& top_blob = top_blobs[0];
top_blob.create(4 * w * h * num_prior, 4u, opt.blob_allocator);
if (top_blob.empty())
return -100;

#pragma omp parallel for num_threads(opt.num_threads)
for (int i = 0; i < h; i++)
{
float* box = (float*)top_blob + i * w * num_prior * 4;

float center_x = offset * step_w;
float center_y = offset * step_h + i * step_h;

for (int j = 0; j < w; j++)
{
// ratio = 1, various sizes
for (int k = 0; k < num_sizes; k++)
{
float size = min_sizes[k];
float cw = size * h / w / 2;
float ch = size / 2;

box[0] = center_x - cw;
box[1] = center_y - ch;
box[2] = center_x + cw;
box[3] = center_y + ch;
box += 4;
}

// various ratios, size = min_size = size[0]
float size = min_sizes[0];
for (int p = 1; p < num_ratios; p++)
{
float ratio = sqrt(aspect_ratios[p]);
float cw = size * h / w * ratio / 2;
float ch = size / ratio / 2;

box[0] = center_x - cw;
box[1] = center_y - ch;
box[2] = center_x + cw;
box[3] = center_y + ch;
box += 4;
}

center_x += step_w;
}
}

if (clip)
{
float* box = top_blob;
for (int i = 0; i < top_blob.w; i++)
{
box[i] = std::min(std::max(box[i], 0.f), 1.f);
}
}

return 0;
}

int image_w = image_width;
int image_h = image_height;
if (image_w == -233)
Expand All @@ -69,26 +146,14 @@ int PriorBox::forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& to
int num_max_size = max_sizes.w;
int num_aspect_ratio = aspect_ratios.w;

Mat min_sizes_copy = min_sizes.clone();
Mat max_sizes_copy = max_sizes.clone();

for (int k = 0; k < num_min_size; k++)
{
if (min_sizes_copy[k] < 0.f)
min_sizes_copy[k] = -min_sizes_copy[k] * image_w;
}
for (int k = 0; k < num_max_size; k++)
{
if (max_sizes_copy[k] < 0.f)
max_sizes_copy[k] = -max_sizes_copy[k] * image_w;
}

int num_prior = num_min_size * num_aspect_ratio + num_min_size + num_max_size;
if (flip)
num_prior += num_min_size * num_aspect_ratio;

Mat& top_blob = top_blobs[0];
top_blob.create(4 * w * h * num_prior, 2, 4u, opt.blob_allocator);
if (top_blob.empty())
return -100;

#pragma omp parallel for num_threads(opt.num_threads)
for (int i = 0; i < h; i++)
Expand All @@ -105,7 +170,7 @@ int PriorBox::forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& to

for (int k = 0; k < num_min_size; k++)
{
float min_size = min_sizes_copy[k];
float min_size = min_sizes[k];

// min size box
box_w = box_h = min_size;
Expand All @@ -119,7 +184,7 @@ int PriorBox::forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& to

if (num_max_size > 0)
{
float max_size = max_sizes_copy[k];
float max_size = max_sizes[k];

// max size box
box_w = box_h = sqrt(min_size * max_size);
Expand Down Expand Up @@ -161,8 +226,6 @@ int PriorBox::forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& to

center_x += step_w;
}

center_y += step_h;
}

if (clip)
Expand Down
57 changes: 32 additions & 25 deletions tools/mxnet/mxnet2ncnn.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -703,6 +703,10 @@ int main(int argc, char** argv)
}
continue;
}
else if (n.op == "_contrib_MultiBoxTarget")
{
n.output_size = 3;
}
else if (n.op == "SliceChannel")
{
n.output_size = n.attr("num_outputs");
Expand Down Expand Up @@ -775,6 +779,11 @@ int main(int argc, char** argv)
}
}

// for (std::map<int, int>::iterator it = node_reference.begin(); it != node_reference.end(); it++)
// {
// fprintf(stderr, "ref %d %d\n", it->first, it->second);
// }

// op chain fusion
int reduced_node_count = 0;
for (int i=0; i<node_count; i++)
Expand Down Expand Up @@ -1277,42 +1286,40 @@ int main(int argc, char** argv)
int keep_top_k = 100;
fprintf(pp, " 3=%d", keep_top_k);
fprintf(pp, " 4=%f", threshold);

std::vector<float> variances = n.attr("variances");
if (variances.empty())
{
fprintf(pp, " 5=0.1");
fprintf(pp, " 6=0.1");
fprintf(pp, " 7=0.2");
fprintf(pp, " 8=0.2");
}
else
{
fprintf(pp, " 5=%f", variances[0]);
fprintf(pp, " 6=%f", variances[1]);
fprintf(pp, " 7=%f", variances[2]);
fprintf(pp, " 8=%f", variances[3]);
}
}
else if (n.op == "_contrib_MultiBoxPrior")
{
// mxnet-ssd encode size as scale factor, fill min_size
std::vector<float> sizes = n.attr("sizes");
float min_size = sizes[0];
float max_size = sizes[1];

// mxnet-ssd encode size as scale factor
fprintf(pp, " -23300=%d", 1);
fprintf(pp, ",%f", -min_size);

fprintf(pp, " -23301=%d", 1);
fprintf(pp, ",%f", -max_size);

// drop 1.0 ratio
std::vector<float> ratios = n.attr("ratios");
std::vector<float> aspect_ratios;
for (int j=0; j<ratios.size(); j++)
fprintf(pp, " -23300=%d", (int)sizes.size());
for (int j=0; j<(int)sizes.size(); j++)
{
if (ratios[j] == 1.f)
continue;
aspect_ratios.push_back(ratios[j]);
fprintf(pp, ",%f", sizes[j]);
}

std::vector<float> aspect_ratios = n.attr("ratios");
fprintf(pp, " -23302=%d", (int)aspect_ratios.size());
for (int j=0; j<(int)aspect_ratios.size(); j++)
{
fprintf(pp, ",%f", aspect_ratios[j]);
}

float variances[4] = {0.1f, 0.1f, 0.2f, 0.2f};
fprintf(pp, " 3=%f", variances[0]);
fprintf(pp, " 4=%f", variances[1]);
fprintf(pp, " 5=%f", variances[2]);
fprintf(pp, " 6=%f", variances[3]);

int flip = 0;
fprintf(pp, " 7=%d", flip);

Expand All @@ -1327,8 +1334,8 @@ int main(int argc, char** argv)
if (steps.empty() || (steps[0] == -1.f && steps[1] == -1.f))
{
// auto step
fprintf(pp, " 11=-233");
fprintf(pp, " 12=-233");
fprintf(pp, " 11=-233.0");
fprintf(pp, " 12=-233.0");
}
else
{
Expand Down

0 comments on commit c77383f

Please sign in to comment.