From 78400d17da78d8a0949077187ef7e65542bf50ff Mon Sep 17 00:00:00 2001
From: nihui <shuizhuyuanluo@126.com>
Date: Thu, 5 Oct 2017 17:16:27 +0800
Subject: [PATCH] implement ssd Normalize Permute PriorBox, introduce ConcatV2
 for interleave

---
 src/CMakeLists.txt      |   4 +
 src/layer/concatv2.cpp  | 298 +++++++++++++++++++++++++++++++++++++
 src/layer/concatv2.h    |  43 ++++++
 src/layer/normalize.cpp | 245 ++++++++++++++++++++++++++++++
 src/layer/normalize.h   |  51 +++++++
 src/layer/permute.cpp   | 189 +++++++++++++++++++++++
 src/layer/permute.h     |  43 ++++++
 src/layer/priorbox.cpp  | 321 ++++++++++++++++++++++++++++++++++++++++
 src/layer/priorbox.h    |  58 ++++++++
 tools/caffe2ncnn.cpp    | 171 ++++++++++++++++++++-
 10 files changed, 1422 insertions(+), 1 deletion(-)
 create mode 100644 src/layer/concatv2.cpp
 create mode 100644 src/layer/concatv2.h
 create mode 100644 src/layer/normalize.cpp
 create mode 100644 src/layer/normalize.h
 create mode 100644 src/layer/permute.cpp
 create mode 100644 src/layer/permute.h
 create mode 100644 src/layer/priorbox.cpp
 create mode 100644 src/layer/priorbox.h

diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index d39d9e521..8c77b24e2 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -131,6 +131,10 @@ ncnn_add_layer(ConvolutionDepthWise)
 ncnn_add_layer(Padding)
 ncnn_add_layer(Squeeze)
 ncnn_add_layer(ExpandDims)
+ncnn_add_layer(Normalize)
+ncnn_add_layer(Permute)
+ncnn_add_layer(PriorBox)
+ncnn_add_layer(ConcatV2)
 
 add_library(ncnn STATIC ${ncnn_SRCS})
 
diff --git a/src/layer/concatv2.cpp b/src/layer/concatv2.cpp
new file mode 100644
index 000000000..0b2a6f729
--- /dev/null
+++ b/src/layer/concatv2.cpp
@@ -0,0 +1,298 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#include "concatv2.h"
+
+namespace ncnn {
+
+DEFINE_LAYER_CREATOR(ConcatV2)
+
+ConcatV2::ConcatV2()
+{
+}
+
+#if NCNN_STDIO
+#if NCNN_STRING
+int ConcatV2::load_param(FILE* paramfp)
+{
+    int nscan = fscanf(paramfp, "%d", &dim);
+    if (nscan != 1)
+    {
+        fprintf(stderr, "ConcatV2 load_param failed %d\n", nscan);
+        return -1;
+    }
+
+    return 0;
+}
+#endif // NCNN_STRING
+int ConcatV2::load_param_bin(FILE* paramfp)
+{
+    fread(&dim, sizeof(int), 1, paramfp);
+
+    return 0;
+}
+#endif // NCNN_STDIO
+
+int ConcatV2::load_param(const unsigned char*& mem)
+{
+    dim = *(int*)(mem);
+    mem += 4;
+
+    return 0;
+}
+
+int ConcatV2::forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs) const
+{
+    int dims = bottom_blobs[0].dims;
+
+    if (dims == 1) // dim == 0
+    {
+        // concat vector
+        // total length
+        int top_w = 0;
+        for (size_t b=0; b<bottom_blobs.size(); b++)
+        {
+            const Mat& bottom_blob = bottom_blobs[b];
+            top_w += bottom_blob.w;
+        }
+
+        Mat& top_blob = top_blobs[0];
+        top_blob.create(top_w);
+        if (top_blob.empty())
+            return -100;
+
+        float* outptr = top_blob;
+        for (size_t b=0; b<bottom_blobs.size(); b++)
+        {
+            const Mat& bottom_blob = bottom_blobs[b];
+
+            int w = bottom_blob.w;
+
+            const float* ptr = bottom_blob;
+            for (int i=0; i<w; i++)
+            {
+                outptr[i] = ptr[i];
+            }
+
+            outptr += w;
+        }
+
+        return 0;
+    }
+
+    if (dims == 2 && dim == 0)
+    {
+        // concat image
+        int w = bottom_blobs[0].w;
+
+        // total height
+        int top_h = 0;
+        for (size_t b=0; b<bottom_blobs.size(); b++)
+        {
+            const Mat& bottom_blob = bottom_blobs[b];
+            top_h += bottom_blob.h;
+        }
+
+        Mat& top_blob = top_blobs[0];
+        top_blob.create(w, top_h);
+        if (top_blob.empty())
+            return -100;
+
+        float* outptr = top_blob;
+        for (size_t b=0; b<bottom_blobs.size(); b++)
+        {
+            const Mat& bottom_blob = bottom_blobs[b];
+
+            int size = w * bottom_blob.h;
+
+            const float* ptr = bottom_blob;
+            for (int i=0; i<size; i++)
+            {
+                outptr[i] = ptr[i];
+            }
+
+            outptr += size;
+        }
+
+        return 0;
+    }
+
+    if (dims == 2 && dim == 1)
+    {
+        // interleave image row
+        int h = bottom_blobs[0].h;
+
+        // total width
+        int top_w = 0;
+        for (size_t b=0; b<bottom_blobs.size(); b++)
+        {
+            const Mat& bottom_blob = bottom_blobs[b];
+            top_w += bottom_blob.w;
+        }
+
+        Mat& top_blob = top_blobs[0];
+        top_blob.create(top_w, h);
+        if (top_blob.empty())
+            return -100;
+
+        #pragma omp parallel for
+        for (int i=0; i<h; i++)
+        {
+            float* outptr = top_blob.row(i);
+            for (size_t b=0; b<bottom_blobs.size(); b++)
+            {
+                const Mat& bottom_blob = bottom_blobs[b];
+
+                const float* ptr = bottom_blob.row(i);
+                for (int j=0; j<bottom_blob.w; j++)
+                {
+                    outptr[j] = ptr[j];
+                }
+
+                outptr += bottom_blob.w;
+            }
+        }
+
+        return 0;
+    }
+
+    if (dims == 3 && dim == 0)
+    {
+        // concat dim
+        int w = bottom_blobs[0].w;
+        int h = bottom_blobs[0].h;
+
+        // total channels
+        int top_channels = 0;
+        for (size_t b=0; b<bottom_blobs.size(); b++)
+        {
+            const Mat& bottom_blob = bottom_blobs[b];
+            top_channels += bottom_blob.c;
+        }
+
+        Mat& top_blob = top_blobs[0];
+        top_blob.create(w, h, top_channels);
+        if (top_blob.empty())
+            return -100;
+
+        int q = 0;
+        for (size_t b=0; b<bottom_blobs.size(); b++)
+        {
+            const Mat& bottom_blob = bottom_blobs[b];
+
+            int channels = bottom_blob.c;
+            int size = bottom_blob.cstep * channels;
+
+            const float* ptr = bottom_blob;
+            float* outptr = top_blob.channel(q);
+            for (int i=0; i<size; i++)
+            {
+                outptr[i] = ptr[i];
+            }
+
+            q += channels;
+        }
+
+        return 0;
+    }
+
+    if (dims == 3 && dim == 1)
+    {
+        // interleave dim height
+        int w = bottom_blobs[0].w;
+        int channels = bottom_blobs[0].c;
+
+        // total height
+        int top_h = 0;
+        for (size_t b=0; b<bottom_blobs.size(); b++)
+        {
+            const Mat& bottom_blob = bottom_blobs[b];
+            top_h += bottom_blob.h;
+        }
+
+        Mat& top_blob = top_blobs[0];
+        top_blob.create(w, top_h, channels);
+        if (top_blob.empty())
+            return -100;
+
+        #pragma omp parallel for
+        for (int q=0; q<channels; q++)
+        {
+            float* outptr = top_blob.channel(q);
+
+            for (size_t b=0; b<bottom_blobs.size(); b++)
+            {
+                const Mat& bottom_blob = bottom_blobs[b];
+
+                int size = bottom_blob.w * bottom_blob.h;
+
+                const float* ptr = bottom_blob.channel(q);
+                for (int i=0; i<size; i++)
+                {
+                    outptr[i] = ptr[i];
+                }
+            }
+        }
+
+        return 0;
+    }
+
+    if (dims == 3 && dim == 2)
+    {
+        // interleave dim width
+        int h = bottom_blobs[0].h;
+        int channels = bottom_blobs[0].c;
+
+        // total height
+        int top_w = 0;
+        for (size_t b=0; b<bottom_blobs.size(); b++)
+        {
+            const Mat& bottom_blob = bottom_blobs[b];
+            top_w += bottom_blob.w;
+        }
+
+        Mat& top_blob = top_blobs[0];
+        top_blob.create(top_w, h, channels);
+        if (top_blob.empty())
+            return -100;
+
+        #pragma omp parallel for
+        for (int q=0; q<channels; q++)
+        {
+            float* outptr = top_blob.channel(q);
+
+            for (int i=0; i<h; i++)
+            {
+                for (size_t b=0; b<bottom_blobs.size(); b++)
+                {
+                    const Mat& bottom_blob = bottom_blobs[b];
+
+                    const float* ptr = bottom_blob.channel(q).row(i);
+                    for (int j=0; j<bottom_blob.w; j++)
+                    {
+                        outptr[j] = ptr[j];
+                    }
+
+                    outptr += bottom_blob.w;
+                }
+            }
+        }
+
+        return 0;
+    }
+
+    return 0;
+}
+
+} // namespace ncnn
diff --git a/src/layer/concatv2.h b/src/layer/concatv2.h
new file mode 100644
index 000000000..7353f5618
--- /dev/null
+++ b/src/layer/concatv2.h
@@ -0,0 +1,43 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#ifndef LAYER_CONCATV2_H
+#define LAYER_CONCATV2_H
+
+#include "layer.h"
+
+namespace ncnn {
+
+class ConcatV2 : public Layer
+{
+public:
+    ConcatV2();
+
+#if NCNN_STDIO
+#if NCNN_STRING
+    virtual int load_param(FILE* paramfp);
+#endif // NCNN_STRING
+    virtual int load_param_bin(FILE* paramfp);
+#endif // NCNN_STDIO
+    virtual int load_param(const unsigned char*& mem);
+
+    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs) const;
+
+public:
+    int dim;
+};
+
+} // namespace ncnn
+
+#endif // LAYER_CONCATV2_H
diff --git a/src/layer/normalize.cpp b/src/layer/normalize.cpp
new file mode 100644
index 000000000..bc5375ca0
--- /dev/null
+++ b/src/layer/normalize.cpp
@@ -0,0 +1,245 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#include "normalize.h"
+#include <math.h>
+
+namespace ncnn {
+
+DEFINE_LAYER_CREATOR(Normalize)
+
+Normalize::Normalize()
+{
+    one_blob_only = true;
+    support_inplace = false;
+}
+
+#if NCNN_STDIO
+#if NCNN_STRING
+int Normalize::load_param(FILE* paramfp)
+{
+    int nscan = fscanf(paramfp, "%d %d %f %d",
+                       &across_spatial, &channel_shared, &eps, &scale_data_size);
+    if (nscan != 4)
+    {
+        fprintf(stderr, "Normalize load_param failed %d\n", nscan);
+        return -1;
+    }
+
+    return 0;
+}
+#endif // NCNN_STRING
+int Normalize::load_param_bin(FILE* paramfp)
+{
+    fread(&across_spatial, sizeof(int), 1, paramfp);
+
+    fread(&channel_shared, sizeof(int), 1, paramfp);
+
+    fread(&eps, sizeof(float), 1, paramfp);
+
+    fread(&scale_data_size, sizeof(int), 1, paramfp);
+
+    return 0;
+}
+
+int Normalize::load_model(FILE* binfp)
+{
+    int nread;
+
+    scale_data.create(1, scale_data_size);
+    nread = fread(scale_data, scale_data_size * sizeof(float), 1, binfp);
+    if (nread != 1)
+    {
+        fprintf(stderr, "Normalize read scale_data failed %d\n", nread);
+        return -1;
+    }
+
+    return 0;
+}
+#endif // NCNN_STDIO
+
+int Normalize::load_param(const unsigned char*& mem)
+{
+    across_spatial = *(int*)(mem);
+    mem += 4;
+
+    channel_shared = *(int*)(mem);
+    mem += 4;
+
+    eps = *(float*)(mem);
+    mem += 4;
+
+    scale_data_size = *(float*)(mem);
+    mem += 4;
+
+    return 0;
+}
+
+int Normalize::load_model(const unsigned char*& mem)
+{
+    scale_data = Mat(1, scale_data_size, (float*)mem);
+    mem += scale_data_size * sizeof(float);
+
+    return 0;
+}
+
+int Normalize::forward(const Mat& bottom_blob, Mat& top_blob) const
+{
+    int w = bottom_blob.w;
+    int h = bottom_blob.h;
+    int channels = bottom_blob.c;
+    int size = w * h;
+
+    top_blob.create(w, h, channels);
+    if (top_blob.empty())
+        return -100;
+
+    if (across_spatial)
+    {
+        // square
+        Mat square_sum_blob;
+        square_sum_blob.create(channels);
+        if (square_sum_blob.empty())
+            return -100;
+
+        float* square_sum_ptr = square_sum_blob;
+        #pragma omp parallel for
+        for (int q=0; q<channels; q++)
+        {
+            const float* ptr = bottom_blob.channel(q);
+
+            float ssum = 0.f;
+            for (int i=0; i<size; i++)
+            {
+                ssum += ptr[i] * ptr[i];
+            }
+
+            square_sum_ptr[q] = ssum;
+        }
+
+        // sum + eps
+        float ssum = eps;
+        for (int q=0; q<channels; q++)
+        {
+            ssum += square_sum_ptr[q];
+        }
+
+        // 1 / sqrt(ssum)
+        float a = 1.f / sqrt(ssum);
+
+        if (channel_shared)
+        {
+            float scale = a * scale_data.data[0];
+
+            #pragma omp parallel for
+            for (int q=0; q<channels; q++)
+            {
+                const float* ptr = bottom_blob.channel(q);
+                float* outptr = top_blob.channel(q);
+
+                for (int i=0; i<size; i++)
+                {
+                    outptr[i] = ptr[i] * scale;
+                }
+            }
+        }
+        else
+        {
+            #pragma omp parallel for
+            for (int q=0; q<channels; q++)
+            {
+                const float* ptr = bottom_blob.channel(q);
+                float* outptr = top_blob.channel(q);
+                float scale = a * scale_data.data[q];
+
+                for (int i=0; i<size; i++)
+                {
+                    outptr[i] = ptr[i] * scale;
+                }
+            }
+        }
+    }
+    else
+    {
+        // square sum, 1 / sqrt(ssum)
+        Mat square_sum_blob;
+        square_sum_blob.create(w, h);
+        if (square_sum_blob.empty())
+            return -100;
+
+        float* ssptr = square_sum_blob;
+
+        if (channel_shared)
+        {
+            float scale = scale_data.data[0];
+
+            #pragma omp parallel for
+            for (int i=0; i<size; i++)
+            {
+                float ssum = eps;
+                for (int q=0; q<channels; q++)
+                {
+                    const float* ptr = bottom_blob.channel(q);
+                    ssum += ptr[i] * ptr[i];
+                }
+
+                ssptr[i] = 1.f / sqrt(ssum) * scale;
+            }
+
+            #pragma omp parallel for
+            for (int q=0; q<channels; q++)
+            {
+                const float* ptr = bottom_blob.channel(q);
+                float* outptr = top_blob.channel(q);
+
+                for (int i=0; i<size; i++)
+                {
+                    outptr[i] = ptr[i] * ssptr[i];
+                }
+            }
+        }
+        else
+        {
+            #pragma omp parallel for
+            for (int i=0; i<size; i++)
+            {
+                float ssum = eps;
+                for (int q=0; q<channels; q++)
+                {
+                    const float* ptr = bottom_blob.channel(q);
+                    ssum += ptr[i] * ptr[i];
+                }
+
+                ssptr[i] = 1.f / sqrt(ssum);
+            }
+
+            #pragma omp parallel for
+            for (int q=0; q<channels; q++)
+            {
+                const float* ptr = bottom_blob.channel(q);
+                float* outptr = top_blob.channel(q);
+                float scale = scale_data.data[q];
+
+                for (int i=0; i<size; i++)
+                {
+                    outptr[i] = ptr[i] * ssptr[i] * scale;
+                }
+            }
+        }
+    }
+
+    return 0;
+}
+
+} // namespace ncnn
diff --git a/src/layer/normalize.h b/src/layer/normalize.h
new file mode 100644
index 000000000..e1ac7f015
--- /dev/null
+++ b/src/layer/normalize.h
@@ -0,0 +1,51 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#ifndef LAYER_NORMALIZE_H
+#define LAYER_NORMALIZE_H
+
+#include "layer.h"
+
+namespace ncnn {
+
+class Normalize : public Layer
+{
+public:
+    Normalize();
+
+#if NCNN_STDIO
+#if NCNN_STRING
+    virtual int load_param(FILE* paramfp);
+#endif // NCNN_STRING
+    virtual int load_param_bin(FILE* paramfp);
+    virtual int load_model(FILE* binfp);
+#endif // NCNN_STDIO
+    virtual int load_param(const unsigned char*& mem);
+    virtual int load_model(const unsigned char*& mem);
+
+    virtual int forward(const Mat& bottom_blob, Mat& top_blob) const;
+
+public:
+    // param
+    int across_spatial;
+    int channel_shared;
+    float eps;
+    int scale_data_size;
+
+    Mat scale_data;
+};
+
+} // namespace ncnn
+
+#endif // LAYER_NORMALIZE_H
diff --git a/src/layer/permute.cpp b/src/layer/permute.cpp
new file mode 100644
index 000000000..9502ca3a1
--- /dev/null
+++ b/src/layer/permute.cpp
@@ -0,0 +1,189 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#include "permute.h"
+
+namespace ncnn {
+
+DEFINE_LAYER_CREATOR(Permute)
+
+Permute::Permute()
+{
+    one_blob_only = true;
+    support_inplace = false;
+}
+
+#if NCNN_STDIO
+#if NCNN_STRING
+int Permute::load_param(FILE* paramfp)
+{
+    int nscan = fscanf(paramfp, "%d", &order_type);
+    if (nscan != 1)
+    {
+        fprintf(stderr, "Permute load_param failed %d\n", nscan);
+        return -1;
+    }
+
+    return 0;
+}
+#endif // NCNN_STRING
+int Permute::load_param_bin(FILE* paramfp)
+{
+    fread(&order_type, sizeof(int), 1, paramfp);
+
+    return 0;
+}
+#endif // NCNN_STDIO
+
+int Permute::load_param(const unsigned char*& mem)
+{
+    order_type = *(int*)(mem);
+    mem += 4;
+
+    return 0;
+}
+
+int Permute::forward(const Mat& bottom_blob, Mat& top_blob) const
+{
+    int w = bottom_blob.w;
+    int h = bottom_blob.h;
+    int channels = bottom_blob.c;
+    int size = w * h;
+
+    // order_type
+    // 0 = w h c
+    // 1 = h w c
+    // 2 = w c h
+    // 3 = c w h
+    // 4 = h c w
+    // 5 = c h w
+
+    if (order_type == 0)
+    {
+        top_blob = bottom_blob;
+    }
+    else if (order_type == 1)
+    {
+        top_blob.create(h, w, channels);
+        if (top_blob.empty())
+            return -100;
+
+        #pragma omp parallel for
+        for (int q=0; q<channels; q++)
+        {
+            const float* ptr = bottom_blob.channel(q);
+            float* outptr = top_blob.channel(q);
+
+            for (int i = 0; i < w; i++)
+            {
+                for (int j = 0; j < h; j++)
+                {
+                    outptr[i*h + j] = ptr[j*w + i];
+                }
+            }
+        }
+    }
+    else if (order_type == 2)
+    {
+        top_blob.create(w, channels, h);
+        if (top_blob.empty())
+            return -100;
+
+        #pragma omp parallel for
+        for (int q=0; q<h; q++)
+        {
+            float* outptr = top_blob.channel(q);
+
+            for (int i = 0; i < channels; i++)
+            {
+                const float* ptr = bottom_blob.channel(i).row(q);
+
+                for (int j = 0; j < w; j++)
+                {
+                    outptr[i*w + j] = ptr[j];
+                }
+            }
+        }
+    }
+    else if (order_type == 3)
+    {
+        top_blob.create(channels, w, h);
+        if (top_blob.empty())
+            return -100;
+
+        #pragma omp parallel for
+        for (int q=0; q<h; q++)
+        {
+            float* outptr = top_blob.channel(q);
+
+            for (int i = 0; i < w; i++)
+            {
+                for (int j = 0; j < channels; j++)
+                {
+                    const float* ptr = bottom_blob.channel(j).row(q);
+
+                    outptr[i*channels + j] = ptr[i];
+                }
+            }
+        }
+    }
+    else if (order_type == 4)
+    {
+        top_blob.create(h, channels, w);
+        if (top_blob.empty())
+            return -100;
+
+        #pragma omp parallel for
+        for (int q=0; q<w; q++)
+        {
+            float* outptr = top_blob.channel(q);
+
+            for (int i = 0; i < channels; i++)
+            {
+                const float* ptr = bottom_blob.channel(i);
+
+                for (int j = 0; j < h; j++)
+                {
+                    outptr[i*channels + j] = ptr[j*w + q];
+                }
+            }
+        }
+    }
+    else if (order_type == 5)
+    {
+        top_blob.create(channels, h, w);
+        if (top_blob.empty())
+            return -100;
+
+        #pragma omp parallel for
+        for (int q=0; q<w; q++)
+        {
+            float* outptr = top_blob.channel(q);
+
+            for (int i = 0; i < h; i++)
+            {
+                for (int j = 0; j < channels; j++)
+                {
+                    const float* ptr = bottom_blob.channel(j);
+
+                    outptr[i*channels + j] = ptr[i*w + q];
+                }
+            }
+        }
+    }
+
+    return 0;
+}
+
+} // namespace ncnn
diff --git a/src/layer/permute.h b/src/layer/permute.h
new file mode 100644
index 000000000..dce4a7d16
--- /dev/null
+++ b/src/layer/permute.h
@@ -0,0 +1,43 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#ifndef LAYER_PERMUTE_H
+#define LAYER_PERMUTE_H
+
+#include "layer.h"
+
+namespace ncnn {
+
+class Permute : public Layer
+{
+public:
+    Permute();
+
+#if NCNN_STDIO
+#if NCNN_STRING
+    virtual int load_param(FILE* paramfp);
+#endif // NCNN_STRING
+    virtual int load_param_bin(FILE* paramfp);
+#endif // NCNN_STDIO
+    virtual int load_param(const unsigned char*& mem);
+
+    virtual int forward(const Mat& bottom_blob, Mat& top_blob) const;
+
+public:
+    int order_type;
+};
+
+} // namespace ncnn
+
+#endif // LAYER_PERMUTE_H
diff --git a/src/layer/priorbox.cpp b/src/layer/priorbox.cpp
new file mode 100644
index 000000000..d9c1b0c62
--- /dev/null
+++ b/src/layer/priorbox.cpp
@@ -0,0 +1,321 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#include "priorbox.h"
+#include <math.h>
+
+namespace ncnn {
+
+DEFINE_LAYER_CREATOR(PriorBox)
+
+PriorBox::PriorBox()
+{
+    one_blob_only = false;
+    support_inplace = false;
+}
+
+#if NCNN_STDIO
+#if NCNN_STRING
+int PriorBox::load_param(FILE* paramfp)
+{
+    int nscan = fscanf(paramfp, "%d %d %d %f %f %f %f %d %d %d %d %f %f %f",
+                       &num_min_size, &num_max_size, &num_aspect_ratio,
+                       &variances[0], &variances[1], &variances[2], &variances[3],
+                       &flip, &clip, &image_width, &image_height,
+                       &step_width, &step_height, &offset);
+    if (nscan != 14)
+    {
+        fprintf(stderr, "PriorBox load_param failed %d\n", nscan);
+        return -1;
+    }
+
+    min_sizes.create(num_min_size);
+    if (min_sizes.empty())
+        return -100;
+    float* min_sizes_ptr = min_sizes;
+    for (int i=0; i<num_min_size; i++)
+    {
+        int nscan = fscanf(paramfp, "%f", &min_sizes_ptr[i]);
+        if (nscan != 1)
+        {
+            fprintf(stderr, "PriorBox load_param failed %d\n", nscan);
+            return -1;
+        }
+    }
+
+    max_sizes.create(num_max_size);
+    if (max_sizes.empty())
+        return -100;
+    float* max_sizes_ptr = max_sizes;
+    for (int i=0; i<num_max_size; i++)
+    {
+        int nscan = fscanf(paramfp, "%f", &max_sizes_ptr[i]);
+        if (nscan != 1)
+        {
+            fprintf(stderr, "PriorBox load_param failed %d\n", nscan);
+            return -1;
+        }
+    }
+
+    aspect_ratios.create(num_aspect_ratio);
+    if (aspect_ratios.empty())
+        return -100;
+    float* aspect_ratios_ptr = aspect_ratios;
+    for (int i=0; i<num_aspect_ratio; i++)
+    {
+        int nscan = fscanf(paramfp, "%f", &aspect_ratios_ptr[i]);
+        if (nscan != 1)
+        {
+            fprintf(stderr, "PriorBox load_param failed %d\n", nscan);
+            return -1;
+        }
+    }
+
+    return 0;
+}
+#endif // NCNN_STRING
+int PriorBox::load_param_bin(FILE* paramfp)
+{
+    fread(&num_min_size, sizeof(int), 1, paramfp);
+
+    fread(&num_max_size, sizeof(int), 1, paramfp);
+
+    fread(&num_aspect_ratio, sizeof(int), 1, paramfp);
+
+    fread(&variances[0], sizeof(float), 1, paramfp);
+
+    fread(&variances[1], sizeof(float), 1, paramfp);
+
+    fread(&variances[2], sizeof(float), 1, paramfp);
+
+    fread(&variances[3], sizeof(float), 1, paramfp);
+
+    fread(&flip, sizeof(int), 1, paramfp);
+
+    fread(&clip, sizeof(int), 1, paramfp);
+
+    fread(&image_width, sizeof(int), 1, paramfp);
+
+    fread(&image_height, sizeof(int), 1, paramfp);
+
+    fread(&step_width, sizeof(float), 1, paramfp);
+
+    fread(&step_height, sizeof(float), 1, paramfp);
+
+    fread(&offset, sizeof(float), 1, paramfp);
+
+    min_sizes.create(num_min_size);
+    if (min_sizes.empty())
+        return -100;
+    float* min_sizes_ptr = min_sizes;
+    fread(min_sizes_ptr, sizeof(float), num_min_size, paramfp);
+
+    max_sizes.create(num_max_size);
+    if (max_sizes.empty())
+        return -100;
+    float* max_sizes_ptr = max_sizes;
+    fread(max_sizes_ptr, sizeof(float), num_max_size, paramfp);
+
+    aspect_ratios.create(num_aspect_ratio);
+    if (aspect_ratios.empty())
+        return -100;
+    float* aspect_ratios_ptr = aspect_ratios;
+    fread(aspect_ratios_ptr, sizeof(float), num_aspect_ratio, paramfp);
+
+    return 0;
+}
+#endif // NCNN_STDIO
+
+int PriorBox::load_param(const unsigned char*& mem)
+{
+    num_min_size = *(int*)(mem);
+    mem += 4;
+
+    num_max_size = *(int*)(mem);
+    mem += 4;
+
+    num_aspect_ratio = *(int*)(mem);
+    mem += 4;
+
+    variances[0] = *(float*)(mem);
+    mem += 4;
+
+    variances[1] = *(float*)(mem);
+    mem += 4;
+
+    variances[2] = *(float*)(mem);
+    mem += 4;
+
+    variances[3] = *(float*)(mem);
+    mem += 4;
+
+    flip = *(int*)(mem);
+    mem += 4;
+
+    clip = *(int*)(mem);
+    mem += 4;
+
+    image_width = *(int*)(mem);
+    mem += 4;
+
+    image_height = *(int*)(mem);
+    mem += 4;
+
+    step_width = *(float*)(mem);
+    mem += 4;
+
+    step_height = *(float*)(mem);
+    mem += 4;
+
+    offset = *(float*)(mem);
+    mem += 4;
+
+    min_sizes = Mat(num_min_size, (float*)mem);
+    mem += num_min_size * sizeof(float);
+
+    max_sizes = Mat(num_max_size, (float*)mem);
+    mem += num_max_size * sizeof(float);
+
+    aspect_ratios = Mat(num_aspect_ratio, (float*)mem);
+    mem += num_aspect_ratio * sizeof(float);
+
+    return 0;
+}
+#include <stdio.h>
+int PriorBox::forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs) const
+{
+    int w = bottom_blobs[0].w;
+    int h = bottom_blobs[0].h;
+
+    int image_w = image_width;
+    int image_h = image_height;
+    if (image_w == -233)
+        image_w = bottom_blobs[1].w;
+    if (image_h == -233)
+        image_h = bottom_blobs[1].h;
+
+    float step_w = step_width;
+    float step_h = step_height;
+    if (step_w == -233)
+        step_w = (float)image_w / w;
+    if (step_h == -233)
+        step_h = (float)image_h / h;
+
+    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);
+
+    #pragma omp parallel for
+    for (int i = 0; i < h; i++)
+    {
+        float* box = top_blob.data + 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++)
+        {
+            float box_w;
+            float box_h;
+
+            for (int k = 0; k < num_min_size; k++)
+            {
+                float min_size = min_sizes.data[k];
+
+                // min size box
+                box_w = box_h = min_size;
+
+                box[0] = (center_x - box_w * 0.5f) / image_w;
+                box[1] = (center_y - box_h * 0.5f) / image_h;
+                box[2] = (center_x + box_w * 0.5f) / image_w;
+                box[3] = (center_y + box_h * 0.5f) / image_h;
+
+                box += 4;
+
+                if (num_max_size > 0)
+                {
+                    float max_size = max_sizes.data[k];
+
+                    // max size box
+                    box_w = box_h = sqrt(min_size * max_size);
+
+                    box[0] = (center_x - box_w * 0.5f) / image_w;
+                    box[1] = (center_y - box_h * 0.5f) / image_h;
+                    box[2] = (center_x + box_w * 0.5f) / image_w;
+                    box[3] = (center_y + box_h * 0.5f) / image_h;
+
+                    box += 4;
+                }
+
+                // all aspect_ratios
+                for (int p = 0; p < num_aspect_ratio; p++)
+                {
+                    float ar = aspect_ratios[p];
+
+                    box_w = min_size * sqrt(ar);
+                    box_h = min_size / sqrt(ar);
+
+                    box[0] = (center_x - box_w * 0.5f) / image_w;
+                    box[1] = (center_y - box_h * 0.5f) / image_h;
+                    box[2] = (center_x + box_w * 0.5f) / image_w;
+                    box[3] = (center_y + box_h * 0.5f) / image_h;
+
+                    box += 4;
+
+                    if (flip)
+                    {
+                        box[0] = (center_x - box_h * 0.5f) / image_h;
+                        box[1] = (center_y - box_w * 0.5f) / image_w;
+                        box[2] = (center_x + box_h * 0.5f) / image_h;
+                        box[3] = (center_y + box_w * 0.5f) / image_w;
+
+                        box += 4;
+                    }
+                }
+            }
+
+            center_x += step_w;
+        }
+
+        center_y += step_h;
+    }
+
+    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);
+        }
+    }
+
+    // set variance
+    float* var = top_blob.row(1);
+    for (int i = 0; i < top_blob.w / 4; i++)
+    {
+        var[0] = variances[0];
+        var[1] = variances[1];
+        var[2] = variances[2];
+        var[3] = variances[3];
+
+        var += 4;
+    }
+
+    return 0;
+}
+
+} // namespace ncnn
diff --git a/src/layer/priorbox.h b/src/layer/priorbox.h
new file mode 100644
index 000000000..17e640cde
--- /dev/null
+++ b/src/layer/priorbox.h
@@ -0,0 +1,58 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#ifndef LAYER_PRIORBOX_H
+#define LAYER_PRIORBOX_H
+
+#include "layer.h"
+
+namespace ncnn {
+
+class PriorBox : public Layer
+{
+public:
+    PriorBox();
+
+#if NCNN_STDIO
+#if NCNN_STRING
+    virtual int load_param(FILE* paramfp);
+#endif // NCNN_STRING
+    virtual int load_param_bin(FILE* paramfp);
+#endif // NCNN_STDIO
+    virtual int load_param(const unsigned char*& mem);
+
+    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs) const;
+
+public:
+    int num_min_size;
+    int num_max_size;
+    int num_aspect_ratio;
+    float variances[4];
+
+    int flip;
+    int clip;
+    int image_width;
+    int image_height;
+    float step_width;
+    float step_height;
+    float offset;
+
+    Mat min_sizes;
+    Mat max_sizes;
+    Mat aspect_ratios;
+};
+
+} // namespace ncnn
+
+#endif // LAYER_PRIORBOX_H
diff --git a/tools/caffe2ncnn.cpp b/tools/caffe2ncnn.cpp
index 9d9e5ccd5..d2aeae2d3 100644
--- a/tools/caffe2ncnn.cpp
+++ b/tools/caffe2ncnn.cpp
@@ -14,6 +14,7 @@
 
 #include <stdio.h>
 #include <limits.h>
+#include <math.h>
 
 #include <fstream>
 #include <set>
@@ -319,7 +320,15 @@ int main(int argc, char** argv)
 
         // layer definition line, repeated
         // [type] [name] [bottom blob count] [top blob count] [bottom blobs] [top blobs] [layer specific params]
-        if (layer.type() == "Convolution")
+        if (layer.type() == "Concat")
+        {
+            const caffe::ConcatParameter& concat_param = layer.concat_param();
+            if (concat_param.axis() != 1)
+                fprintf(pp, "%-16s", "ConcatV2");
+            else
+                fprintf(pp, "%-16s", "Concat");
+        }
+        else if (layer.type() == "Convolution")
         {
             const caffe::ConvolutionParameter& convolution_param = layer.convolution_param();
             if (convolution_param.group() != 1)
@@ -426,6 +435,15 @@ int main(int argc, char** argv)
             std::vector<float> zeros(mean_blob.data_size(), 0.f);
             fwrite(zeros.data(), sizeof(float), zeros.size(), bp);// bias
         }
+        else if (layer.type() == "Concat")
+        {
+            const caffe::ConcatParameter& concat_param = layer.concat_param();
+            if (concat_param.axis() != 1)
+            {
+                int dim = concat_param.axis() >= 1 ? concat_param.axis() - 1 : 0;
+                fprintf(pp, " %d", dim);
+            }
+        }
         else if (layer.type() == "Convolution")
         {
             const caffe::LayerParameter& binlayer = net.layer(netidx);
@@ -641,6 +659,77 @@ int main(int argc, char** argv)
             const caffe::MemoryDataParameter& memory_data_param = layer.memory_data_param();
             fprintf(pp, " %d %d %d", memory_data_param.channels(), memory_data_param.width(), memory_data_param.height());
         }
+        else if (layer.type() == "Normalize")
+        {
+            const caffe::LayerParameter& binlayer = net.layer(netidx);
+            const caffe::BlobProto& scale_blob = binlayer.blobs(0);
+            const caffe::NormalizeParameter& norm_param = layer.norm_param();
+            fprintf(pp, " %d %d %f %d", norm_param.across_spatial(), norm_param.channel_shared(), norm_param.eps(), scale_blob.data_size());
+
+            fwrite(scale_blob.data().data(), sizeof(float), scale_blob.data_size(), bp);
+        }
+        else if (layer.type() == "Permute")
+        {
+            const caffe::PermuteParameter& permute_param = layer.permute_param();
+            int order_size = permute_param.order_size();
+            int order_type = 0;
+            if (order_size == 0)
+                order_type = 0;
+            if (order_size == 1)
+            {
+                int order0 = permute_param.order(0);
+                if (order0 == 0)
+                    order_type = 0;
+                // permute with N not supported
+            }
+            if (order_size == 2)
+            {
+                int order0 = permute_param.order(0);
+                int order1 = permute_param.order(1);
+                if (order0 == 0)
+                {
+                    if (order1 == 1) // 0 1 2 3
+                        order_type = 0;
+                    else if (order1 == 2) // 0 2 1 3
+                        order_type = 2;
+                    else if (order1 == 3) // 0 3 1 2
+                        order_type = 4;
+                }
+                // permute with N not supported
+            }
+            if (order_size == 3 || order_size == 4)
+            {
+                int order0 = permute_param.order(0);
+                int order1 = permute_param.order(1);
+                int order2 = permute_param.order(2);
+                if (order0 == 0)
+                {
+                    if (order1 == 1)
+                    {
+                        if (order2 == 2) // 0 1 2 3
+                            order_type = 0;
+                        if (order2 == 3) // 0 1 3 2
+                            order_type = 1;
+                    }
+                    else if (order1 == 2)
+                    {
+                        if (order2 == 1) // 0 2 1 3
+                            order_type = 2;
+                        if (order2 == 3) // 0 2 3 1
+                            order_type = 3;
+                    }
+                    else if (order1 == 3)
+                    {
+                        if (order2 == 1) // 0 3 1 2
+                            order_type = 4;
+                        if (order2 == 2) // 0 3 2 1
+                            order_type = 5;
+                    }
+                }
+                // permute with N not supported
+            }
+            fprintf(pp, " %d", order_type);
+        }
         else if (layer.type() == "Pooling")
         {
             const caffe::PoolingParameter& pooling_param = layer.pooling_param();
@@ -659,6 +748,86 @@ int main(int argc, char** argv)
             fprintf(pp, " %d", slope_blob.data_size());
             fwrite(slope_blob.data().data(), sizeof(float), slope_blob.data_size(), bp);
         }
+        else if (layer.type() == "PriorBox")
+        {
+            const caffe::PriorBoxParameter& prior_box_param = layer.prior_box_param();
+
+            int num_aspect_ratio = prior_box_param.aspect_ratio_size();
+            for (int j=0; j<prior_box_param.aspect_ratio_size(); j++)
+            {
+                float ar = prior_box_param.aspect_ratio(j);
+                if (fabs(ar - 1.) < 1e-6) {
+                    num_aspect_ratio--;
+                }
+            }
+
+            float variances[4] = {0.1f, 0.1f, 0.1f, 0.1f};
+            if (prior_box_param.variance_size() == 4)
+            {
+                variances[0] = prior_box_param.variance(0);
+                variances[1] = prior_box_param.variance(1);
+                variances[2] = prior_box_param.variance(2);
+                variances[3] = prior_box_param.variance(3);
+            }
+            else if (prior_box_param.variance_size() == 1)
+            {
+                variances[0] = prior_box_param.variance(0);
+                variances[1] = prior_box_param.variance(0);
+                variances[2] = prior_box_param.variance(0);
+                variances[3] = prior_box_param.variance(0);
+            }
+
+            int flip = prior_box_param.has_flip() ? prior_box_param.flip() : 1;
+            int clip = prior_box_param.has_clip() ? prior_box_param.clip() : 0;
+            int image_width = -233;
+            int image_height = -233;
+            if (prior_box_param.has_img_size())
+            {
+                image_width = prior_box_param.img_size();
+                image_height = prior_box_param.img_size();
+            }
+            else if (prior_box_param.has_img_w() && prior_box_param.has_img_h())
+            {
+                image_width = prior_box_param.img_w();
+                image_height = prior_box_param.img_h();
+            }
+
+            float step_width = -233;
+            float step_height = -233;
+            if (prior_box_param.has_step())
+            {
+                step_width = prior_box_param.step();
+                step_height = prior_box_param.step();
+            }
+            else if (prior_box_param.has_step_w() && prior_box_param.has_step_h())
+            {
+                step_width = prior_box_param.step_w();
+                step_height = prior_box_param.step_h();
+            }
+
+            fprintf(pp, " %d %d %d %f %f %f %f %d %d %d %d %f %f %f", prior_box_param.min_size_size(),
+                    prior_box_param.max_size_size(), num_aspect_ratio,
+                    variances[0], variances[1], variances[2], variances[3],
+                    flip, clip, image_width, image_height,
+                    step_width, step_height, prior_box_param.offset());
+
+            for (int j=0; j<prior_box_param.min_size_size(); j++)
+            {
+                fprintf(pp, " %f", prior_box_param.min_size(j));
+            }
+            for (int j=0; j<prior_box_param.max_size_size(); j++)
+            {
+                fprintf(pp, " %f", prior_box_param.max_size(j));
+            }
+            for (int j=0; j<prior_box_param.aspect_ratio_size(); j++)
+            {
+                float ar = prior_box_param.aspect_ratio(j);
+                if (fabs(ar - 1.) < 1e-6) {
+                    continue;
+                }
+                fprintf(pp, " %f", ar);
+            }
+        }
         else if (layer.type() == "Proposal")
         {
             const caffe::PythonParameter& python_param = layer.python_param();