ncnn/tools/mxnet/mxnet2ncnn.cpp

// Copyright 2017 Tencent
// SPDX-License-Identifier: BSD-3-Clause

#include <limits.h>
#include <map>
#include <set>
#include <stdint.h>
#include <stdio.h>
#include <string.h>
#include <string>
#include <vector>

class MXNetParam;
class MXNetNode
{
public:
    bool has_attr(const char* key) const;
    bool is_attr_scalar(const char* key) const;

    class AttrProxy
    {
        MXNetNode const* _n;
        const char* const _key;

    public:
        AttrProxy(MXNetNode const* n, const char* key)
            : _n(n), _key(key)
        {
        }
        operator int() const
        {
            return _n->attr_i(_key);
        }
        operator float() const
        {
            return _n->attr_f(_key);
        }
        operator std::string() const
        {
            return _n->attr_s(_key);
        }
        operator std::vector<int>() const
        {
            return _n->attr_ai(_key);
        }
        operator std::vector<float>() const
        {
            return _n->attr_af(_key);
        }
    };

    AttrProxy attr(const char* key) const
    {
        return AttrProxy(this, key);
    }

    int attr_i(const char* key) const;
    float attr_f(const char* key) const;
    std::string attr_s(const char* key) const;
    std::vector<int> attr_ai(const char* key) const;
    std::vector<float> attr_af(const char* key) const;

public:
    bool is_weight() const;
    bool has_weight(int i) const;
    std::vector<float> weight(int i, int init_len = 0) const;

    std::vector<MXNetNode>* nodes;   // reference
    std::vector<MXNetParam>* params; // reference

public:
    std::string op;
    std::string name;
    int output_size;
    std::map<std::string, std::string> attrs;
    std::vector<int> inputs;
    std::vector<int> subinputs;
    std::vector<int> weights;
};

class MXNetParam
{
public:
    std::string name;
    std::vector<float> data;
    std::string init;
};

bool MXNetNode::has_attr(const char* key) const
{
    const std::map<std::string, std::string>::const_iterator it = attrs.find(key);
    return it != attrs.end();
}

bool MXNetNode::is_attr_scalar(const char* key) const
{
    const std::map<std::string, std::string>::const_iterator it = attrs.find(key);
    if (it == attrs.end())
        return false;

    if (it->second.empty())
        return false;

    return it->second[0] != '(';
}

int MXNetNode::attr_i(const char* key) const
{
    const std::map<std::string, std::string>::const_iterator it = attrs.find(key);
    if (it == attrs.end())
        return 0;

    if (it->second == "False")
        return 0;

    if (it->second == "True")
        return 1;

    int i = 0;
    int nscan = sscanf(it->second.c_str(), "%d", &i);
    if (nscan != 1)
        return 0;

    return i;
}

float MXNetNode::attr_f(const char* key) const
{
    const std::map<std::string, std::string>::const_iterator it = attrs.find(key);
    if (it == attrs.end())
        return 0.f;

    float f = 0;
    int nscan = sscanf(it->second.c_str(), "%f", &f);
    if (nscan != 1)
        return 0.f;

    return f;
}

std::string MXNetNode::attr_s(const char* key) const
{
    const std::map<std::string, std::string>::const_iterator it = attrs.find(key);
    if (it == attrs.end())
        return std::string();

    return it->second;
}

std::vector<int> MXNetNode::attr_ai(const char* key) const
{
    const std::map<std::string, std::string>::const_iterator it = attrs.find(key);
    if (it == attrs.end())
        return std::vector<int>();

    // (1,2,3)
    std::vector<int> list;

    if (is_attr_scalar(key))
    {
        list.push_back(attr_i(key));
        return list;
    }

    int i = 0;
    int c = 0;
    int nconsumed = 0;
    int nscan = sscanf(it->second.c_str() + c, "%*[\\[(,]%d%n", &i, &nconsumed);
    if (nscan != 1)
    {
        // (None
        if (strncmp(it->second.c_str() + c, "(None", 5) == 0)
        {
            i = -233;
            nconsumed = 5;
            nscan = 1;
        }
    }
    while (nscan == 1)
    {
        list.push_back(i);
        //         fprintf(stderr, "%d\n", i);

        i = 0;
        c += nconsumed;
        nscan = sscanf(it->second.c_str() + c, "%*[(,]%d%n", &i, &nconsumed);
        if (nscan != 1)
        {
            // , None
            if (strncmp(it->second.c_str() + c, ", None", 6) == 0)
            {
                i = -233;
                nconsumed = 6;
                nscan = 1;
            }
        }
    }

    return list;
}

std::vector<float> MXNetNode::attr_af(const char* key) const
{
    const std::map<std::string, std::string>::const_iterator it = attrs.find(key);
    if (it == attrs.end())
        return std::vector<float>();

    // (0.1,0.2,0.3)
    std::vector<float> list;

    if (is_attr_scalar(key))
    {
        list.push_back(attr_f(key));
        return list;
    }

    float i = 0.f;
    int c = 0;
    int nconsumed = 0;
    int nscan = sscanf(it->second.c_str() + c, "%*[(,]%f%n", &i, &nconsumed);
    while (nscan == 1)
    {
        list.push_back(i);
        //         fprintf(stderr, "%f\n", i);

        i = 0.f;
        c += nconsumed;
        nscan = sscanf(it->second.c_str() + c, "%*[(,]%f%n", &i, &nconsumed);
    }

    return list;
}

bool MXNetNode::is_weight() const
{
    for (int i = 0; i < (int)(*params).size(); i++)
    {
        const MXNetParam& p = (*params)[i];
        if (p.name == name)
            return true;
    }

    return false;
}

bool MXNetNode::has_weight(int i) const
{
    if (i < 0 || i >= (int)weights.size())
        return false;

    const std::string& node_name = (*nodes)[weights[i]].name;

    for (int j = 0; j < (int)(*params).size(); j++)
    {
        const MXNetParam& p = (*params)[j];
        if (p.name == node_name)
            return true;
    }

    return false;
}

std::vector<float> MXNetNode::weight(int i, int init_len) const
{
    if (i < 0 || i >= (int)weights.size())
        return std::vector<float>();

    const std::string& node_name = (*nodes)[weights[i]].name;

    for (int j = 0; j < (int)(*params).size(); j++)
    {
        const MXNetParam& p = (*params)[j];
        if (p.name != node_name)
            continue;

        if (!p.data.empty())
            return p.data;

        std::vector<float> data;

        if (!p.init.empty() && init_len != 0)
        {
            if (p.init == "[\\$zero\\$, {}]" || p.init == "[\\\"zero\\\", {}]" || p.init == "zeros")
            {
                data.resize(init_len, 0.f);
            }
            else if (p.init == "[\\$one\\$, {}]" || p.init == "[\\\"one\\\", {}]" || p.init == "ones")
            {
                data.resize(init_len, 1.f);
            }
        }

        return data;
    }

    return std::vector<float>();
}

static void replace_backslash_doublequote_dollar(char* s)
{
    char* a = s;
    char* b = s + 1;
    while (*a && *b)
    {
        if (*a == '\\' && *b == '\"')
        {
            *b = '$';
        }

        a++;
        b++;
    }
}

static void parse_input_list(const char* s, std::vector<int>& inputs, std::vector<int>& subinputs)
{
    inputs.clear();
    subinputs.clear();

    if (memcmp(s, "[]", 2) == 0)
        return;

    int nscan = 0;
    int nconsumed = 0;

    int id;
    int subid;

    int c = 1; // skip leading [
    nscan = sscanf(s + c, "[%d, %d%n", &id, &subid, &nconsumed);
    while (nscan == 2)
    {
        inputs.push_back(id);
        subinputs.push_back(subid);
        //         fprintf(stderr, "%d %d\n", id, subid);

        c += nconsumed;
        nscan = sscanf(s + c, "%*[^[][%d, %d%n", &id, &subid, &nconsumed);
    }
}

static bool read_mxnet_json(const char* jsonpath, std::vector<MXNetNode>& nodes)
{
    FILE* fp = fopen(jsonpath, "rb");
    if (!fp)
    {
        fprintf(stderr, "fopen %s failed\n", jsonpath);
        return false;
    }

    int internal_unknown = 0;
    int internal_underscore = 0;

    char line[1024];

    //{
    char* s = fgets(line, 1024, fp);
    if (!s)
    {
        fprintf(stderr, "fgets %s failed\n", jsonpath);
        return false;
    }

    MXNetNode n;

    bool in_nodes_list = false;
    bool in_node_block = false;
    bool in_attr_block = false;
    bool in_inputs_block = false;
    while (!feof(fp))
    {
        char* t = fgets(line, 1024, fp);
        if (!t)
            break;

        if (in_inputs_block)
        {
            //      ]
            if (memcmp(line, "      ]", 7) == 0)
            {
                in_inputs_block = false;
                continue;
            }

            //        [439, 0, 0],
            int id;
            int subid;
            int nscan = sscanf(line, "        [%d, %d", &id, &subid);
            if (nscan == 2)
            {
                n.inputs.push_back(id);
                n.subinputs.push_back(subid);
                continue;
            }
        }

        if (in_attr_block)
        {
            //      },
            if (memcmp(line, "      }", 7) == 0)
            {
                in_attr_block = false;
                continue;
            }

            // replace \" with \$
            replace_backslash_doublequote_dollar(line);

            //        "kernel": "(7,7)",
            char key[256] = {0};
            char value[256] = {0};
            int nscan = sscanf(line, "        \"%255[^\"]\": \"%255[^\"]\"", key, value);
            if (nscan == 2)
            {
                n.attrs[key] = value;
                //                 fprintf(stderr, "# %s = %s\n", key, value);
                continue;
            }
        }

        if (in_node_block)
        {
            //    },
            if (memcmp(line, "    }", 5) == 0)
            {
                // new node
                if (n.name.empty())
                {
                    // assign default unknown name
                    char unknownname[256];
                    sprintf(unknownname, "unknownncnn_%d", internal_unknown);

                    n.name = unknownname;

                    internal_unknown++;
                }
                if (n.name[0] == '_')
                {
                    // workaround for potential duplicated _plus0
                    char underscorename[256];
                    sprintf(underscorename, "underscorencnn_%d%s", internal_underscore, n.name.c_str());

                    n.name = underscorename;

                    internal_underscore++;
                }
                nodes.push_back(n);

                in_node_block = false;
                continue;
            }

            int nscan;

            //      "op": "Convolution",
            char op[256] = {0};
            nscan = sscanf(line, "      \"op\": \"%255[^\"]\",", op);
            if (nscan == 1)
            {
                n.op = op;
                //                 fprintf(stderr, "op = %s\n", op);
                continue;
            }

            //      "name": "conv0",
            char name[256] = {0};
            nscan = sscanf(line, "      \"name\": \"%255[^\"]\",", name);
            if (nscan == 1)
            {
                n.name = name;
                //                 fprintf(stderr, "name = %s\n", name);
                continue;
            }

            //      "inputs": [
            if (memcmp(line, "      \"inputs\": [\n", 18) == 0)
            {
                in_inputs_block = true;
                continue;
            }

            //      "inputs": []
            char inputs[256] = {0};
            nscan = sscanf(line, "      \"inputs\": %255[^\n]", inputs);
            if (nscan == 1)
            {
                parse_input_list(inputs, n.inputs, n.subinputs);
                //                 fprintf(stderr, "inputs = %s\n", inputs);
                continue;
            }

            //      "param": {},
            if (memcmp(line, "      \"param\": {}", 17) == 0)
            {
                continue;
            }

            // replace \" with \$
            replace_backslash_doublequote_dollar(line);

            //      "attr": {"__init__": "[\"zero\", {}]"},
            char key[256] = {0};
            char value[256] = {0};
            nscan = sscanf(line, "      \"attr\": {\"%255[^\"]\": \"%255[^\"]\"}", key, value);
            if (nscan == 2)
            {
                n.attrs[key] = value;
                //                 fprintf(stderr, "# %s = %s\n", key, value);
                continue;
            }

            //      "attrs": {"__init__": "[\"zero\", {}]"},
            nscan = sscanf(line, "      \"attrs\": {\"%255[^\"]\": \"%255[^\"]\"}", key, value);
            if (nscan == 2)
            {
                n.attrs[key] = value;
                //                 fprintf(stderr, "# %s = %s\n", key, value);
                continue;
            }

            //      "param": {"p": "0.5"},
            nscan = sscanf(line, "      \"param\": {\"%255[^\"]\": \"%255[^\"]\"}", key, value);
            if (nscan == 2)
            {
                n.attrs[key] = value;
                //                 fprintf(stderr, "# %s = %s\n", key, value);
                continue;
            }

            //      "attr": {
            if (memcmp(line, "      \"attr\": {", 15) == 0)
            {
                in_attr_block = true;
                continue;
            }

            //      "attrs": {
            if (memcmp(line, "      \"attrs\": {", 16) == 0)
            {
                in_attr_block = true;
                continue;
            }

            //      "param": {
            if (memcmp(line, "      \"param\": {", 16) == 0)
            {
                in_attr_block = true;
                continue;
            }
        }

        if (in_nodes_list)
        {
            //  ],
            if (memcmp(line, "  ],", 4) == 0)
            {
                in_nodes_list = false;
                // all nodes parsed
                break;
            }

            //    {
            if (memcmp(line, "    {", 5) == 0)
            {
                n = MXNetNode();

                in_node_block = true;
                continue;
            }
        }

        //  "nodes": [
        if (memcmp(line, "  \"nodes\": [", 12) == 0)
        {
            in_nodes_list = true;
            continue;
        }
    }

    fclose(fp);

    return true;
}

static bool read_mxnet_param(const char* parampath, std::vector<MXNetParam>& params)
{
    FILE* fp = fopen(parampath, "rb");
    if (!fp)
    {
        fprintf(stderr, "fopen %s failed\n", parampath);
        return false;
    }

    size_t nread;
    uint64_t header;
    uint64_t reserved;
    nread = fread(&header, sizeof(uint64_t), 1, fp);
    if (nread != 1)
    {
        fprintf(stderr, "read header failed %zd\n", nread);
        return false;
    }
    nread = fread(&reserved, sizeof(uint64_t), 1, fp);
    if (nread != 1)
    {
        fprintf(stderr, "read reserved failed %zd\n", nread);
        return false;
    }

    // NDArray vec

    // each data
    uint64_t data_count;
    nread = fread(&data_count, sizeof(uint64_t), 1, fp);
    if (nread != 1)
    {
        fprintf(stderr, "read data_count failed %zd\n", nread);
        return false;
    }

    //     fprintf(stderr, "data count = %d\n", (int)data_count);

    for (int i = 0; i < (int)data_count; i++)
    {
        uint32_t magic; // 0xF993FAC9
        nread = fread(&magic, sizeof(uint32_t), 1, fp);
        if (nread != 1)
        {
            fprintf(stderr, "read magic failed %zd\n", nread);
            return false;
        }

        // shape
        uint32_t ndim;
        std::vector<int64_t> shape;

        if (magic == 0xF993FAC9)
        {
            int32_t stype;
            nread = fread(&stype, sizeof(int32_t), 1, fp);
            if (nread != 1)
            {
                fprintf(stderr, "read stype failed %zd\n", nread);
                return false;
            }

            nread = fread(&ndim, sizeof(uint32_t), 1, fp);
            if (nread != 1)
            {
                fprintf(stderr, "read ndim failed %zd\n", nread);
                return false;
            }

            shape.resize(ndim);
            nread = fread(&shape[0], ndim * sizeof(int64_t), 1, fp);
            if (nread != 1)
            {
                fprintf(stderr, "read shape failed %zd\n", nread);
                return false;
            }
        }
        else if (magic == 0xF993FAC8)
        {
            nread = fread(&ndim, sizeof(uint32_t), 1, fp);
            if (nread != 1)
            {
                fprintf(stderr, "read ndim failed %zd\n", nread);
                return false;
            }

            shape.resize(ndim);
            nread = fread(&shape[0], ndim * sizeof(int64_t), 1, fp);
            if (nread != 1)
            {
                fprintf(stderr, "read shape failed %zd\n", nread);
                return false;
            }
        }
        else
        {
            ndim = magic;

            shape.resize(ndim);

            std::vector<uint32_t> shape32;
            shape32.resize(ndim);
            nread = fread(&shape32[0], ndim * sizeof(uint32_t), 1, fp);
            if (nread != 1)
            {
                fprintf(stderr, "read shape failed %zd\n", nread);
                return false;
            }

            for (int j = 0; j < (int)ndim; j++)
            {
                shape[j] = shape32[j];
            }
        }

        // context
        int32_t dev_type;
        int32_t dev_id;
        nread = fread(&dev_type, sizeof(int32_t), 1, fp);
        if (nread != 1)
        {
            fprintf(stderr, "read dev_type failed %zd\n", nread);
            return false;
        }
        nread = fread(&dev_id, sizeof(int32_t), 1, fp);
        if (nread != 1)
        {
            fprintf(stderr, "read dev_id failed %zd\n", nread);
            return false;
        }

        int32_t type_flag;
        nread = fread(&type_flag, sizeof(int32_t), 1, fp);
        if (nread != 1)
        {
            fprintf(stderr, "read type_flag failed %zd\n", nread);
            return false;
        }

        // data
        size_t len = 0;
        if (shape.size() == 1) len = shape[0];
        if (shape.size() == 2) len = shape[0] * shape[1];
        if (shape.size() == 3) len = shape[0] * shape[1] * shape[2];
        if (shape.size() == 4) len = shape[0] * shape[1] * shape[2] * shape[3];

        MXNetParam p;

        p.data.resize(len);
        nread = fread(&p.data[0], len * sizeof(float), 1, fp);
        if (nread != 1)
        {
            fprintf(stderr, "read MXNetParam data failed %zd\n", nread);
            return false;
        }

        params.push_back(p);

        //         fprintf(stderr, "%u read\n", len);
    }

    // each name
    uint64_t name_count;
    nread = fread(&name_count, sizeof(uint64_t), 1, fp);
    if (nread != 1)
    {
        fprintf(stderr, "read name_count failed %zd\n", nread);
        return false;
    }

    //     fprintf(stderr, "name count = %d\n", (int)name_count);

    for (int i = 0; i < (int)name_count; i++)
    {
        uint64_t len;
        nread = fread(&len, sizeof(uint64_t), 1, fp);
        if (nread != 1)
        {
            fprintf(stderr, "read name length failed %zd\n", nread);
            return false;
        }

        MXNetParam& p = params[i];

        p.name.resize(len);
        nread = fread((char*)p.name.data(), len, 1, fp);
        if (nread != 1)
        {
            fprintf(stderr, "read MXNetParam name failed %zd\n", nread);
            return false;
        }

        // cut leading arg:
        if (memcmp(p.name.c_str(), "arg:", 4) == 0)
        {
            p.name = std::string(p.name.c_str() + 4);
        }
        if (memcmp(p.name.c_str(), "aux:", 4) == 0)
        {
            p.name = std::string(p.name.c_str() + 4);
        }

        //         fprintf(stderr, "%s read\n", p.name.c_str());
    }

    fclose(fp);

    return true;
}

static void fuse_shufflechannel(std::vector<MXNetNode>& nodes, std::vector<MXNetParam>& params, std::map<size_t, int>& node_reference, std::set<std::string>& blob_names, int& reduced_node_count)
{
    size_t node_count = nodes.size();
    for (size_t i = 0; i < node_count; i++)
    {
        const MXNetNode& n = nodes[i];

        if (n.is_weight())
            continue;

        // ShuffleChannel <= Reshape - SwapAxis - Reshape
        if (n.op == "Reshape")
        {
            if (node_reference.find(i) == node_reference.end() || node_reference[i] != 1)
                continue;

            // "shape": "(0, -4, X, -1, -2)"
            std::vector<int> shape = n.attr("shape");
            if (shape.size() != 5)
                continue;
            if (shape[0] != 0 || shape[1] != -4 || shape[3] != -1 || shape[4] != -2)
                continue;

            if (i + 2 >= node_count)
                continue;

            const MXNetNode& n2 = nodes[i + 1];
            const MXNetNode& n3 = nodes[i + 2];

            if (n2.op != "SwapAxis" || n3.op != "Reshape")
                continue;

            if (node_reference.find(i + 1) == node_reference.end() || node_reference[i + 1] != 1)
                continue;

            // "dim1": "1", "dim2": "2"
            int dim1 = n2.attr("dim1");
            int dim2 = n2.attr("dim2");
            if (dim1 != 1 || dim2 != 2)
                continue;

            // "shape": "(0, -3, -2)"
            std::vector<int> shape3 = n3.attr("shape");
            if (shape3.size() != 3)
                continue;
            if (shape3[0] != 0 || shape3[1] != -3 || shape3[2] != -2)
                continue;

            // reduce
            nodes[i].op = "noop_reducedncnn";
            nodes[i + 1].op = "noop_reducedncnn";

            node_reference.erase(node_reference.find(i));
            node_reference.erase(node_reference.find(i + 1));
            blob_names.erase(n.name);
            blob_names.erase(n2.name);

            MXNetNode new_node;
            new_node.nodes = &nodes;
            new_node.params = &params;
            new_node.op = "ShuffleChannel";
            //             new_node.name = n.name + "_" + n2.name + "_" + n3.name;
            new_node.name = n3.name;
            new_node.output_size = n3.output_size;
            char group[16];
            sprintf(group, "%d", shape[2]);
            new_node.attrs["group"] = group;
            new_node.inputs = n.inputs;
            new_node.subinputs = n.subinputs;

            nodes[i + 2] = new_node;

            reduced_node_count += 2;
            i += 2;
        }
    }
}

static void fuse_hardsigmoid_hardswish(std::vector<MXNetNode>& nodes, std::vector<MXNetParam>& params, std::map<size_t, int>& node_reference, std::set<std::string>& blob_names, int& reduced_node_count)
{
    size_t node_count = nodes.size();
    for (size_t i = 0; i < node_count; i++)
    {
        const MXNetNode& n = nodes[i];

        if (n.is_weight())
            continue;

        if (n.op == "_plus_scalar")
        {
            // HardSigmoid <= _plus_scalar(+3) - clip(0,6) - _div_scalar(/6)
            const MXNetNode& n1 = nodes[i + 1];
            const MXNetNode& n2 = nodes[i + 2];
            const MXNetNode& n3 = nodes[i + 3];

            if ((float)n.attr("scalar") != 3.f)
                continue;

            if (n1.op != "clip" || (float)n1.attr("a_min") != 0.f || (float)n1.attr("a_max") != 6.f)
                continue;

            if (n2.op != "_div_scalar" || (float)n2.attr("scalar") != 6.f)
                continue;

            // reduce
            nodes[i].op = "noop_reducedncnn";
            nodes[i + 1].op = "noop_reducedncnn";

            node_reference.erase(node_reference.find(i));
            node_reference.erase(node_reference.find(i + 1));
            blob_names.erase(n.name);
            blob_names.erase(n1.name);

            if (n3.op != "elemwise_mul" || n3.inputs[0] != n.inputs[0])
            {
                MXNetNode new_node;
                new_node.nodes = &nodes;
                new_node.params = &params;
                new_node.op = "HardSigmoid";
                new_node.name = n2.name;
                new_node.output_size = n2.output_size;
                char alpha[16], beta[16];
                sprintf(alpha, "%f", 1.f / 6.f);
                sprintf(beta, "%f", 3.f / 6.f);
                new_node.attrs["alpha"] = alpha;
                new_node.attrs["beta"] = beta;
                new_node.inputs = n.inputs;
                new_node.subinputs = n.subinputs;

                nodes[i + 2] = new_node;

                reduced_node_count += 2;
                i += 2;
            }
            else // HardSwish <= HardSigmoid - Mul
            {
                nodes[i + 2].op = "noop_reducedncnn";
                node_reference[i - 1]--;
                node_reference.erase(node_reference.find(i + 2));
                blob_names.erase(n2.name);

                MXNetNode new_node;
                new_node.nodes = &nodes;
                new_node.params = &params;
                new_node.op = "HardSwish";
                new_node.name = n3.name;
                new_node.output_size = n3.output_size;
                char alpha[16], beta[16];
                sprintf(alpha, "%f", 1.f / 6.f);
                sprintf(beta, "%f", 3.f / 6.f);
                new_node.attrs["alpha"] = alpha;
                new_node.attrs["beta"] = beta;
                new_node.inputs = n.inputs;
                new_node.subinputs = n.subinputs;

                nodes[i + 3] = new_node;

                reduced_node_count += 3;
                i += 3;
            }
        }
    }
}

int main(int argc, char** argv)
{
    if (!(argc == 3 || argc == 5))
    {
        fprintf(stderr, "Usage: %s [mxnetjson] [mxnetparam] [ncnnparam] [ncnnbin]\n", argv[0]);
        return -1;
    }

    const char* jsonpath = argv[1];
    const char* parampath = argv[2];
    const char* ncnn_prototxt = argc == 5 ? argv[3] : "ncnn.param";
    const char* ncnn_modelbin = argc == 5 ? argv[4] : "ncnn.bin";

    std::vector<MXNetNode> nodes;
    std::vector<MXNetParam> params;

    read_mxnet_json(jsonpath, nodes);
    read_mxnet_param(parampath, params);

    FILE* pp = fopen(ncnn_prototxt, "wb");
    FILE* bp = fopen(ncnn_modelbin, "wb");

    // magic
    fprintf(pp, "7767517\n");

    size_t node_count = nodes.size();

    // node reference
    std::map<size_t, int> node_reference;

    // weight node
    std::vector<int> weight_nodes;

    // sometimes mxnet produce non-unique name for activation op
    {
        std::set<std::string> known_names;
        for (size_t i = 0; i < node_count; i++)
        {
            MXNetNode& n = nodes[i];

            if (known_names.find(n.name) == known_names.end())
            {
                known_names.insert(n.name);
                continue;
            }

            // non-unique name detected, append index as suffix
            char suffix[32];
            sprintf(suffix, "_%d", (int)i);
            n.name = n.name + std::string(suffix);
        }
    }

    // global definition line
    // [layer count] [blob count]
    std::set<std::string> blob_names;
    for (size_t i = 0; i < node_count; i++)
    {
        MXNetNode& n = nodes[i];

        // assign global param reference
        n.nodes = &nodes;
        n.params = &params;

        const std::string& output_name = n.name;
        n.output_size = 1;

        if (n.op == "null")
        {
            if (n.is_weight())
            {
                weight_nodes.push_back(i);
            }
            else
            {
                if (n.has_attr("__init__"))
                {
                    // init weight param
                    MXNetParam pi;
                    pi.name = n.name;
                    pi.init = (std::string)n.attr("__init__");
                    params.push_back(pi);

                    weight_nodes.push_back(i);
                }
                else
                {
                    // null node without data, treat it as network input
                }
            }
            continue;
        }
        else if (n.op == "_contrib_MultiBoxTarget")
        {
            n.output_size = 3;
        }
        else if (n.op == "SliceChannel")
        {
            n.output_size = n.attr("num_outputs");
        }

        // distinguish weights and inputs
        std::vector<int> weights;
        std::vector<int> inputs;
        for (int j = 0; j < (int)n.inputs.size(); j++)
        {
            int input_index = n.inputs[j];
            if (nodes[input_index].is_weight())
            {
                weights.push_back(input_index);
                continue;
            }

            inputs.push_back(input_index);
        }
        n.inputs = inputs;
        n.weights = weights;

        if (n.op == "_contrib_MultiBoxDetection")
        {
            // reorder input blob
            int temp = n.inputs[0];
            n.inputs[0] = n.inputs[1];
            n.inputs[1] = temp;
        }

        // input
        for (int j = 0; j < (int)n.inputs.size(); j++)
        {
            int input_index = n.inputs[j];
            int subinput_index = n.subinputs[j];

            std::string input_name = nodes[input_index].name;
            //             fprintf(stderr, "input = %s\n", input_name.c_str());

            if (subinput_index != 0)
            {
                char subinputsuffix[256];
                sprintf(subinputsuffix, "_subncnn_%d", subinput_index);
                input_name = input_name + subinputsuffix;
            }

            blob_names.insert(input_name);

            int input_uid = input_index | (subinput_index << 16);
            if (node_reference.find(input_uid) == node_reference.end())
            {
                node_reference[input_uid] = 1;
            }
            else
            {
                node_reference[input_uid] = node_reference[input_uid] + 1;
            }
        }

        // output
        //         fprintf(stderr, "output = %s\n", output_name.c_str());
        blob_names.insert(output_name);

        for (int j = 1; j < n.output_size; j++)
        {
            char subinputsuffix[256];
            sprintf(subinputsuffix, "_%d", j);
            std::string output_name_j = output_name + subinputsuffix;
            blob_names.insert(output_name_j);
        }
    }

    //     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;
    fuse_shufflechannel(nodes, params, node_reference, blob_names, reduced_node_count);
    fuse_hardsigmoid_hardswish(nodes, params, node_reference, blob_names, reduced_node_count);

    // remove node_reference entry with reference equals to one
    int splitncnn_blob_count = 0;
    std::map<size_t, int>::iterator it = node_reference.begin();
    while (it != node_reference.end())
    {
        if (it->second == 1)
        {
            node_reference.erase(it++);
        }
        else
        {
            splitncnn_blob_count += it->second;
            //             fprintf(stderr, "%s %d\n", it->first.c_str(), it->second);
            ++it;
        }
    }

    //     fprintf(stderr, "%d %d %d %d, %d %d\n", node_count, reduced_node_count, node_reference.size(), weight_nodes.size(), blob_names.size(), splitncnn_blob_count);

    fprintf(pp, "%zu %zu\n", node_count - reduced_node_count + node_reference.size() - weight_nodes.size(), blob_names.size() + splitncnn_blob_count);

    int internal_split = 0;

    for (size_t i = 0; i < node_count; i++)
    {
        const MXNetNode& n = nodes[i];

        if (n.op == "noop_reducedncnn")
        {
            continue;
        }

        if (n.op == "null")
        {
            if (n.is_weight())
            {
                continue;
            }

            fprintf(pp, "%-16s", "Input");
        }
        else if (n.op == "_contrib_BilinearResize2D")
        {
            fprintf(pp, "%-16s", "Interp");
        }
        else if (n.op == "_contrib_MultiBoxDetection")
        {
            fprintf(pp, "%-16s", "DetectionOutput");
        }
        else if (n.op == "_contrib_MultiBoxPrior")
        {
            fprintf(pp, "%-16s", "PriorBox");
        }
        else if (n.op == "_copy")
        {
            fprintf(pp, "%-16s", "Noop");
        }
        else if (n.op == "_div_scalar")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "_maximum_scalar")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "_minimum_scalar")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "_minus_scalar")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "_mul_scalar")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "_plus_scalar")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "_power_scalar")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "_rdiv_scalar")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "_rminus_scalar")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "abs")
        {
            fprintf(pp, "%-16s", "UnaryOp");
        }
        else if (n.op == "Activation")
        {
            std::string type = n.attr("act_type");
            if (type == "relu")
            {
                fprintf(pp, "%-16s", "ReLU");
            }
            else if (type == "sigmoid")
            {
                fprintf(pp, "%-16s", "Sigmoid");
            }
            else if (type == "tanh")
            {
                fprintf(pp, "%-16s", "TanH");
            }
        }
        else if (n.op == "add_n" || n.op == "ElementWiseSum")
        {
            fprintf(pp, "%-16s", "Eltwise");
        }
        else if (n.op == "arccos")
        {
            fprintf(pp, "%-16s", "UnaryOp");
        }
        else if (n.op == "arcsin")
        {
            fprintf(pp, "%-16s", "UnaryOp");
        }
        else if (n.op == "arctan")
        {
            fprintf(pp, "%-16s", "UnaryOp");
        }
        else if (n.op == "BatchNorm")
        {
            fprintf(pp, "%-16s", "BatchNorm");
        }
        else if (n.op == "broadcast_add")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "broadcast_div")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "broadcast_mul")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "broadcast_sub")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "ceil")
        {
            fprintf(pp, "%-16s", "UnaryOp");
        }
        else if (n.op == "clip")
        {
            fprintf(pp, "%-16s", "Clip");
        }
        else if (n.op == "Concat")
        {
            fprintf(pp, "%-16s", "Concat");
        }
        else if (n.op == "Convolution")
        {
            int num_group = n.attr("num_group");
            if (num_group > 1)
            {
                fprintf(pp, "%-16s", "ConvolutionDepthWise");
            }
            else
            {
                fprintf(pp, "%-16s", "Convolution");
            }
        }
        else if (n.op == "cos")
        {
            fprintf(pp, "%-16s", "UnaryOp");
        }
        else if (n.op == "Crop")
        {
            fprintf(pp, "%-16s", "Crop");
        }
        else if (n.op == "Deconvolution")
        {
            int num_group = n.attr("num_group");
            if (num_group > 1)
            {
                fprintf(pp, "%-16s", "DeconvolutionDepthWise");
            }
            else
            {
                fprintf(pp, "%-16s", "Deconvolution");
            }
        }
        else if (n.op == "dot")
        {
            fprintf(pp, "%-16s", "Gemm");
        }
        else if (n.op == "Dropout")
        {
            fprintf(pp, "%-16s", "Dropout");
        }
        else if (n.op == "elemwise_add" || n.op == "_add" || n.op == "_plus" || n.op == "_Plus")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "elemwise_div" || n.op == "_div" || n.op == "_Div")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "elemwise_mul" || n.op == "_mul" || n.op == "_Mul")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "elemwise_sub" || n.op == "_sub" || n.op == "_minus" || n.op == "_Minus")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "Embedding")
        {
            fprintf(pp, "%-16s", "Embed");
        }
        else if (n.op == "exp")
        {
            fprintf(pp, "%-16s", "UnaryOp");
        }
        else if (n.op == "expand_dims")
        {
            fprintf(pp, "%-16s", "ExpandDims");
        }
        else if (n.op == "Flatten")
        {
            fprintf(pp, "%-16s", "Flatten");
        }
        else if (n.op == "floor")
        {
            fprintf(pp, "%-16s", "UnaryOp");
        }
        else if (n.op == "FullyConnected")
        {
            fprintf(pp, "%-16s", "InnerProduct");
        }
        else if (n.op == "HardSigmoid")
        {
            fprintf(pp, "%-16s", "HardSigmoid");
        }
        else if (n.op == "HardSwish")
        {
            fprintf(pp, "%-16s", "HardSwish");
        }
        else if (n.op == "InstanceNorm")
        {
            fprintf(pp, "%-16s", "InstanceNorm");
        }
        else if (n.op == "L2Normalization")
        {
            fprintf(pp, "%-16s", "Normalize");
        }
        else if (n.op == "LeakyReLU")
        {
            std::string type = n.attr("act_type");
            if (type == "elu")
            {
                fprintf(pp, "%-16s", "ELU");
            }
            else if (type == "leaky" || type.empty())
            {
                fprintf(pp, "%-16s", "ReLU");
            }
            else if (type == "prelu")
            {
                fprintf(pp, "%-16s", "PReLU");
            }
        }
        else if (n.op == "LinearRegressionOutput")
        {
            fprintf(pp, "%-16s", "Noop");
        }
        else if (n.op == "log")
        {
            fprintf(pp, "%-16s", "UnaryOp");
        }
        else if (n.op == "LogisticRegressionOutput")
        {
            fprintf(pp, "%-16s", "Sigmoid");
        }
        else if (n.op == "MAERegressionOutput")
        {
            fprintf(pp, "%-16s", "Noop");
        }
        else if (n.op == "max" || n.op == "mean" || n.op == "min" || n.op == "prod" || n.op == "sum")
        {
            fprintf(pp, "%-16s", "Reduction");
        }
        else if (n.op == "maximum")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "minimum")
        {
            fprintf(pp, "%-16s", "BinaryOp");
        }
        else if (n.op == "negative")
        {
            fprintf(pp, "%-16s", "UnaryOp");
        }
        else if (n.op == "Pad")
        {
            fprintf(pp, "%-16s", "Padding");
        }
        else if (n.op == "Pooling")
        {
            fprintf(pp, "%-16s", "Pooling");
        }
        else if (n.op == "reciprocal")
        {
            fprintf(pp, "%-16s", "UnaryOp");
        }
        else if (n.op == "relu")
        {
            fprintf(pp, "%-16s", "ReLU");
        }
        else if (n.op == "Reshape")
        {
            fprintf(pp, "%-16s", "Reshape");
        }
        else if (n.op == "ShuffleChannel")
        {
            fprintf(pp, "%-16s", "ShuffleChannel");
        }
        else if (n.op == "sigmoid")
        {
            fprintf(pp, "%-16s", "Sigmoid");
        }
        else if (n.op == "sin")
        {
            fprintf(pp, "%-16s", "UnaryOp");
        }
        else if (n.op == "slice")
        {
            fprintf(pp, "%-16s", "Crop");
        }
        else if (n.op == "slice_axis")
        {
            fprintf(pp, "%-16s", "Crop");
        }
        else if (n.op == "SliceChannel")
        {
            fprintf(pp, "%-16s", "Slice");
        }
        else if (n.op == "SoftmaxActivation")
        {
            fprintf(pp, "%-16s", "Softmax");
        }
        else if (n.op == "SoftmaxOutput")
        {
            fprintf(pp, "%-16s", "Softmax");
        }
        else if (n.op == "softmax")
        {
            fprintf(pp, "%-16s", "Softmax");
        }
        else if (n.op == "sqrt")
        {
            fprintf(pp, "%-16s", "UnaryOp");
        }
        else if (n.op == "square")
        {
            fprintf(pp, "%-16s", "UnaryOp");
        }
        else if (n.op == "squeeze")
        {
            fprintf(pp, "%-16s", "Squeeze");
        }
        else if (n.op == "tan")
        {
            fprintf(pp, "%-16s", "UnaryOp");
        }
        else if (n.op == "tanh")
        {
            fprintf(pp, "%-16s", "TanH");
        }
        else if (n.op == "Transpose" || n.op == "transpose")
        {
            fprintf(pp, "%-16s", "Permute");
        }
        else if (n.op == "UpSampling")
        {
            std::string sample_type = n.attr("sample_type");
            if (sample_type == "nearest")
            {
                fprintf(pp, "%-16s", "Interp");
            }
            else if (sample_type == "bilinear")
            {
                fprintf(pp, "%-16s", "DeconvolutionDepthWise");
            }
        }
        else
        {
            fprintf(stderr, "%s not supported yet!\n", n.op.c_str());
            fprintf(pp, "%-16s", n.op.c_str());
        }

        size_t input_size = n.inputs.size();
        for (int j = 0; j < (int)n.inputs.size(); j++)
        {
            int input_index = n.inputs[j];
            if (nodes[input_index].is_weight())
            {
                input_size--;
            }
        }

        if (n.op == "SoftmaxOutput" || n.op == "LogisticRegressionOutput")
        {
            // drop label
            input_size--;
        }

        fprintf(pp, " %-32s %zd %d", n.name.c_str(), input_size, n.output_size);

        for (int j = 0; j < (int)n.inputs.size(); j++)
        {
            int input_index = n.inputs[j];
            int subinput_index = n.subinputs[j];
            if (nodes[input_index].is_weight())
            {
                continue;
            }

            if (n.op == "SoftmaxOutput" || n.op == "LogisticRegressionOutput")
            {
                // drop label
                if (j == 1)
                    continue;
            }

            std::string input_name = nodes[input_index].name;

            if (subinput_index != 0)
            {
                char subinputsuffix[256];
                sprintf(subinputsuffix, "_subncnn_%d", subinput_index);
                input_name = input_name + subinputsuffix;
            }

            int input_uid = input_index | (subinput_index << 16);
            if (node_reference.find(input_uid) != node_reference.end())
            {
                int refidx = node_reference[input_uid] - 1;
                node_reference[input_uid] = refidx;

                char splitsuffix[256];
                sprintf(splitsuffix, "_splitncnn_%d", refidx);
                input_name = input_name + splitsuffix;
            }

            fprintf(pp, " %s", input_name.c_str());
        }

        fprintf(pp, " %s", n.name.c_str());
        for (int j = 1; j < n.output_size; j++)
        {
            fprintf(pp, " %s_subncnn_%d", n.name.c_str(), j);
        }

        if (n.op == "null")
        {
            // dummy input shape
            //             fprintf(pp, " 0 0 0");
        }
        else if (n.op == "_contrib_BilinearResize2D")
        {
            float scale_height = n.has_attr("scale_height") ? n.attr("scale_height") : 1.f;
            float scale_width = n.has_attr("scale_width") ? n.attr("scale_width") : 1.f;
            int height = n.has_attr("scale_height") ? 0 : n.attr("height");
            int width = n.has_attr("scale_width") ? 0 : n.attr("width");

            fprintf(pp, " 0=2");
            fprintf(pp, " 1=%e", scale_height);
            fprintf(pp, " 2=%e", scale_width);
            fprintf(pp, " 3=%d", height);
            fprintf(pp, " 4=%d", width);
        }
        else if (n.op == "_contrib_MultiBoxDetection")
        {
            float threshold = n.has_attr("threshold") ? n.attr("threshold") : 0.01f;
            float nms_threshold = n.has_attr("nms_threshold") ? n.attr("nms_threshold") : 0.5f;
            int nms_topk = n.has_attr("nms_topk") ? n.attr("nms_topk") : 300;

            fprintf(pp, " 0=-233");
            fprintf(pp, " 1=%e", nms_threshold);
            fprintf(pp, " 2=%d", nms_topk);

            int keep_top_k = 100;
            fprintf(pp, " 3=%d", keep_top_k);
            fprintf(pp, " 4=%e", 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=%e", variances[0]);
                fprintf(pp, " 6=%e", variances[1]);
                fprintf(pp, " 7=%e", variances[2]);
                fprintf(pp, " 8=%e", 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");
            fprintf(pp, " -23300=%d", (int)sizes.size());
            for (int j = 0; j < (int)sizes.size(); j++)
            {
                fprintf(pp, ",%e", 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, ",%e", aspect_ratios[j]);
            }

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

            int clip = n.attr("clip");
            fprintf(pp, " 8=%d", clip);

            // auto image size
            fprintf(pp, " 9=-233");
            fprintf(pp, " 10=-233");

            std::vector<float> steps = n.attr("steps");
            if (steps.empty() || (steps[0] == -1.f && steps[1] == -1.f))
            {
                // auto step
                fprintf(pp, " 11=-233.0");
                fprintf(pp, " 12=-233.0");
            }
            else
            {
                fprintf(pp, " 11=%e", steps[1]);
                fprintf(pp, " 12=%e", steps[0]);
            }

            std::vector<float> offsets = n.attr("offsets");
            if (offsets.empty() || (offsets[0] == 0.5f && offsets[1] == 0.5f))
            {
                fprintf(pp, " 13=0.5");
            }
            else
            {
                fprintf(stderr, "Unsupported offsets param! %g %g\n", offsets[0], offsets[1]);
            }
        }
        else if (n.op == "_copy")
        {
            // noop
        }
        else if (n.op == "_div_scalar")
        {
            int op_type = 3;
            int with_scalar = 1;
            float scalar = n.attr("scalar");
            fprintf(pp, " 0=%d", op_type);
            fprintf(pp, " 1=%d", with_scalar);
            fprintf(pp, " 2=%e", scalar);
        }
        else if (n.op == "_maximum_scalar")
        {
            int op_type = 4;
            int with_scalar = 1;
            float scalar = n.attr("scalar");
            fprintf(pp, " 0=%d", op_type);
            fprintf(pp, " 1=%d", with_scalar);
            fprintf(pp, " 2=%e", scalar);
        }
        else if (n.op == "_minimum_scalar")
        {
            int op_type = 5;
            int with_scalar = 1;
            float scalar = n.attr("scalar");
            fprintf(pp, " 0=%d", op_type);
            fprintf(pp, " 1=%d", with_scalar);
            fprintf(pp, " 2=%e", scalar);
        }
        else if (n.op == "_minus_scalar")
        {
            int op_type = 1;
            int with_scalar = 1;
            float scalar = n.attr("scalar");
            fprintf(pp, " 0=%d", op_type);
            fprintf(pp, " 1=%d", with_scalar);
            fprintf(pp, " 2=%e", scalar);
        }
        else if (n.op == "_mul_scalar")
        {
            int op_type = 2;
            int with_scalar = 1;
            float scalar = n.attr("scalar");
            fprintf(pp, " 0=%d", op_type);
            fprintf(pp, " 1=%d", with_scalar);
            fprintf(pp, " 2=%e", scalar);
        }
        else if (n.op == "_plus_scalar")
        {
            int op_type = 0;
            int with_scalar = 1;
            float scalar = n.attr("scalar");
            fprintf(pp, " 0=%d", op_type);
            fprintf(pp, " 1=%d", with_scalar);
            fprintf(pp, " 2=%e", scalar);
        }
        else if (n.op == "_power_scalar")
        {
            int op_type = 6;
            int with_scalar = 1;
            float scalar = n.attr("scalar");
            fprintf(pp, " 0=%d", op_type);
            fprintf(pp, " 1=%d", with_scalar);
            fprintf(pp, " 2=%e", scalar);
        }
        else if (n.op == "_rdiv_scalar")
        {
            int op_type = 8;
            int with_scalar = 1;
            float scalar = n.attr("scalar");
            fprintf(pp, " 0=%d", op_type);
            fprintf(pp, " 1=%d", with_scalar);
            fprintf(pp, " 2=%e", scalar);
        }
        else if (n.op == "_rminus_scalar")
        {
            int op_type = 7;
            int with_scalar = 1;
            float scalar = n.attr("scalar");
            fprintf(pp, " 0=%d", op_type);
            fprintf(pp, " 1=%d", with_scalar);
            fprintf(pp, " 2=%e", scalar);
        }
        else if (n.op == "abs")
        {
            int op_type = 0;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "Activation")
        {
            std::string type = n.attr("act_type");
            if (type == "relu")
            {
                //                 fprintf(pp, " 0=%e", 0.f);
            }
        }
        else if (n.op == "add_n" || n.op == "ElementWiseSum")
        {
            int op_type = 1;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "arccos")
        {
            int op_type = 13;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "arcsin")
        {
            int op_type = 12;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "arctan")
        {
            int op_type = 14;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "BatchNorm")
        {
            float eps = 1e-3f;
            if (n.has_attr("eps"))
            {
                eps = n.attr("eps");
            }

            std::vector<float> slope_data = n.weight(0);
            std::vector<float> bias_data = n.weight(1);

            int channels = static_cast<int>(slope_data.size());

            std::vector<float> mean_data = n.weight(2, channels);
            std::vector<float> var_data = n.weight(3, channels);

            for (int j = 0; j < (int)var_data.size(); j++)
            {
                var_data[j] += eps;
            }

            fprintf(pp, " 0=%d", channels);

            int fix_gamma = n.has_attr("fix_gamma") ? n.attr("fix_gamma") : 0;
            if (fix_gamma)
            {
                // slope data are all 0 here, force set 1
                for (int j = 0; j < channels; j++)
                {
                    slope_data[j] = 1.f;
                }
            }

            fwrite(slope_data.data(), sizeof(float), slope_data.size(), bp);
            fwrite(mean_data.data(), sizeof(float), mean_data.size(), bp);
            fwrite(var_data.data(), sizeof(float), var_data.size(), bp);
            fwrite(bias_data.data(), sizeof(float), bias_data.size(), bp);
        }
        else if (n.op == "broadcast_add")
        {
            int op_type = 0;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "broadcast_div")
        {
            int op_type = 3;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "broadcast_mul")
        {
            int op_type = 2;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "broadcast_sub")
        {
            int op_type = 1;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "ceil")
        {
            int op_type = 3;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "clip")
        {
            float min = n.attr("a_min");
            float max = n.attr("a_max");
            fprintf(pp, " 0=%e", min);
            fprintf(pp, " 1=%e", max);
        }
        else if (n.op == "Concat")
        {
            int dim = n.has_attr("dim") ? n.attr("dim") : 1;
            fprintf(pp, " 0=%d", dim - 1);
        }
        else if (n.op == "Convolution")
        {
            int num_filter = n.attr("num_filter");
            std::vector<int> kernel = n.attr("kernel");
            std::vector<int> dilate = n.attr("dilate");
            std::vector<int> stride = n.attr("stride");
            std::vector<int> pad = n.attr("pad");
            int no_bias = n.attr("no_bias");
            int num_group = n.attr("num_group");

            std::vector<float> weight_data = n.weight(0);
            std::vector<float> bias_data = n.weight(1);

            fprintf(pp, " 0=%d", num_filter);
            if (kernel.size() == 1)
            {
                fprintf(pp, " 1=%d", kernel[0]);
            }
            else if (kernel.size() == 2)
            {
                fprintf(pp, " 1=%d", kernel[1]);
                fprintf(pp, " 11=%d", kernel[0]);
            }

            if (dilate.size() == 1)
            {
                fprintf(pp, " 2=%d", dilate[0]);
            }
            else if (dilate.size() == 2)
            {
                fprintf(pp, " 2=%d", dilate[1]);
                fprintf(pp, " 12=%d", dilate[0]);
            }

            if (stride.size() == 1)
            {
                fprintf(pp, " 3=%d", stride[0]);
            }
            else if (stride.size() == 2)
            {
                fprintf(pp, " 3=%d", stride[1]);
                fprintf(pp, " 13=%d", stride[0]);
            }

            if (pad.size() == 1)
            {
                fprintf(pp, " 4=%d", pad[0]);
            }
            else if (pad.size() == 2)
            {
                fprintf(pp, " 4=%d", pad[1]);
                fprintf(pp, " 14=%d", pad[0]);
            }

            fprintf(pp, " 5=%d", no_bias == 1 ? 0 : 1);
            fprintf(pp, " 6=%d", (int)weight_data.size());
            if (num_group > 1)
            {
                fprintf(pp, " 7=%d", num_group);
            }

            int quantize_tag = 0;
            fwrite(&quantize_tag, sizeof(int), 1, bp);
            fwrite(weight_data.data(), sizeof(float), weight_data.size(), bp);
            fwrite(bias_data.data(), sizeof(float), bias_data.size(), bp);
        }
        else if (n.op == "cos")
        {
            int op_type = 10;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "Crop")
        {
            int num_args = n.attr("num_args");
            std::vector<int> offset = n.attr("offset");

            int woffset = 0;
            int hoffset = 0;
            if (offset.size() == 2)
            {
                woffset = offset[1];
                hoffset = offset[0];
            }

            fprintf(pp, " 0=%d", woffset);
            fprintf(pp, " 1=%d", hoffset);
            fprintf(pp, " 2=0");

            if (num_args == 1)
            {
                std::vector<int> h_w = n.attr("h_w");
                fprintf(pp, " 3=%d", h_w[1]);
                fprintf(pp, " 4=%d", h_w[0]);
                fprintf(pp, " 5=0");
            }
        }
        else if (n.op == "Deconvolution")
        {
            int num_filter = n.attr("num_filter");
            std::vector<int> kernel = n.attr("kernel");
            std::vector<int> dilate = n.attr("dilate");
            std::vector<int> stride = n.attr("stride");
            std::vector<int> pad = n.attr("pad");
            std::vector<int> adj = n.attr("adj");
            std::vector<int> target_shape = n.attr("target_shape");
            int no_bias = n.attr("no_bias");
            int num_group = n.attr("num_group");

            std::vector<float> weight_data = n.weight(0);
            std::vector<float> bias_data = n.weight(1);

            fprintf(pp, " 0=%d", num_filter);
            if (kernel.size() == 1)
            {
                fprintf(pp, " 1=%d", kernel[0]);
            }
            else if (kernel.size() == 2)
            {
                fprintf(pp, " 1=%d", kernel[1]);
                fprintf(pp, " 11=%d", kernel[0]);
            }

            if (dilate.size() == 1)
            {
                fprintf(pp, " 2=%d", dilate[0]);
            }
            else if (dilate.size() == 2)
            {
                fprintf(pp, " 2=%d", dilate[1]);
                fprintf(pp, " 12=%d", dilate[0]);
            }

            if (stride.size() == 1)
            {
                fprintf(pp, " 3=%d", stride[0]);
            }
            else if (stride.size() == 2)
            {
                fprintf(pp, " 3=%d", stride[1]);
                fprintf(pp, " 13=%d", stride[0]);
            }

            if (target_shape.size() == 0)
            {
                if (pad.size() == 1)
                {
                    fprintf(pp, " 4=%d", pad[0]);
                }
                else if (pad.size() == 2)
                {
                    fprintf(pp, " 4=%d", pad[1]);
                    fprintf(pp, " 14=%d", pad[0]);
                }

                if (adj.size() == 1)
                {
                    fprintf(pp, " 18=%d", adj[0]);
                }
                else if (adj.size() == 2)
                {
                    fprintf(pp, " 18=%d", adj[1]);
                    fprintf(pp, " 19=%d", adj[0]);
                }
            }
            else
            {
                fprintf(pp, " 4=-233");

                if (target_shape.size() == 1)
                {
                    fprintf(pp, " 20=%d", target_shape[0]);
                }
                else if (target_shape.size() == 2)
                {
                    fprintf(pp, " 20=%d", target_shape[1]);
                    fprintf(pp, " 21=%d", target_shape[0]);
                }
            }

            fprintf(pp, " 5=%d", no_bias == 1 ? 0 : 1);
            fprintf(pp, " 6=%d", (int)weight_data.size());
            if (num_group > 1)
            {
                fprintf(pp, " 7=%d", num_group);
            }

            int quantize_tag = 0;
            fwrite(&quantize_tag, sizeof(int), 1, bp);

            int maxk = 0;
            if (kernel.size() == 2)
            {
                maxk = kernel[1] * kernel[0];
            }
            else
            {
                maxk = kernel[0] * kernel[0];
            }
            for (int g = 0; g < num_group; g++)
            {
                // reorder weight from inch-outch to outch-inch
                int num_filter_g = num_filter / num_group;
                int num_input = static_cast<int>(weight_data.size() / maxk / num_filter_g / num_group);
                const float* weight_data_ptr = weight_data.data() + g * maxk * num_filter_g * num_input;
                for (int k = 0; k < num_filter_g; k++)
                {
                    for (int j = 0; j < num_input; j++)
                    {
                        fwrite(weight_data_ptr + (j * num_filter_g + k) * maxk, sizeof(float), maxk, bp);
                    }
                }
            }

            fwrite(bias_data.data(), sizeof(float), bias_data.size(), bp);
        }
        else if (n.op == "dot")
        {
            int transpose_a = n.attr("transpose_a");
            int transpose_b = n.attr("transpose_b");
            fprintf(pp, " 0=1.0"); // alpha
            fprintf(pp, " 1=1.0"); // beta
            fprintf(pp, " 2=%d", transpose_a);
            fprintf(pp, " 3=%d", transpose_b);
        }
        else if (n.op == "Dropout")
        {
            //             float p = n.attr("p");
            //             fprintf(pp, " 0=%d", p);
        }
        else if (n.op == "elemwise_add" || n.op == "_add" || n.op == "_plus" || n.op == "_Plus")
        {
            int op_type = 0;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "elemwise_div" || n.op == "_div" || n.op == "_Div")
        {
            int op_type = 3;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "elemwise_mul" || n.op == "_mul" || n.op == "_Mul")
        {
            int op_type = 2;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "elemwise_sub" || n.op == "_sub" || n.op == "_minus" || n.op == "_Minus")
        {
            int op_type = 1;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "Embedding")
        {
            int input_dim = n.attr("input_dim");
            int output_dim = n.attr("output_dim");

            std::vector<float> weight_data = n.weight(0);

            fprintf(pp, " 0=%d", output_dim);
            fprintf(pp, " 1=%d", input_dim);
            fprintf(pp, " 3=%d", (int)weight_data.size());

            int quantize_tag = 0;
            fwrite(&quantize_tag, sizeof(int), 1, bp);
            fwrite(weight_data.data(), sizeof(float), weight_data.size(), bp);
        }
        else if (n.op == "exp")
        {
            int op_type = 7;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "expand_dims")
        {
            int axis = n.attr("axis");

            fprintf(pp, " -23303=1,%d", axis);
        }
        else if (n.op == "Flatten")
        {
            // no param
        }
        else if (n.op == "floor")
        {
            int op_type = 2;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "FullyConnected")
        {
            int num_hidden = n.attr("num_hidden");
            int no_bias = n.attr("no_bias");
            //             int flatten = n.attr("flatten");

            // TODO flatten

            std::vector<float> weight_data = n.weight(0);
            std::vector<float> bias_data = n.weight(1);

            fprintf(pp, " 0=%d", num_hidden);
            fprintf(pp, " 1=%d", no_bias == 1 ? 0 : 1);
            fprintf(pp, " 2=%d", (int)weight_data.size());

            int quantize_tag = 0;
            fwrite(&quantize_tag, sizeof(int), 1, bp);
            fwrite(weight_data.data(), sizeof(float), weight_data.size(), bp);
            fwrite(bias_data.data(), sizeof(float), bias_data.size(), bp);
        }
        else if (n.op == "HardSigmoid")
        {
            float alpha = n.attr("alpha");
            float beta = n.attr("beta");

            fprintf(pp, " 0=%e", alpha);
            fprintf(pp, " 1=%e", beta);
        }
        else if (n.op == "HardSwish")
        {
            float alpha = n.attr("alpha");
            float beta = n.attr("beta");

            fprintf(pp, " 0=%e", alpha);
            fprintf(pp, " 1=%e", beta);
        }
        else if (n.op == "InstanceNorm")
        {
            float eps = n.has_attr("eps") ? n.attr("eps") : 0.001f;

            std::vector<float> gamma_data = n.weight(0);
            std::vector<float> beta_data = n.weight(1);

            fprintf(pp, " 0=%d", (int)gamma_data.size());
            fprintf(pp, " 1=%e", eps);

            fwrite(gamma_data.data(), sizeof(float), gamma_data.size(), bp);
            fwrite(beta_data.data(), sizeof(float), beta_data.size(), bp);
        }
        else if (n.op == "L2Normalization")
        {
            std::string mode = n.attr("mode");
            float eps = n.has_attr("eps") ? n.attr("eps") : 1e-10f;

            int across_spatial = 0;
            int across_channel = 1;
            int channel_shared = 1;
            int scale_data_size = 1;

            if (mode == "instance")
            {
                across_spatial = 1;
                across_channel = 1;
            }
            else if (mode == "channel")
            {
                across_spatial = 0;
                across_channel = 1;
            }
            else if (mode == "spatial")
            {
                across_spatial = 1;
                across_channel = 0;
            }

            fprintf(pp, " 0=%d", across_spatial);
            fprintf(pp, " 4=%d", across_channel);
            fprintf(pp, " 1=%d", channel_shared);
            fprintf(pp, " 2=%e", eps);
            fprintf(pp, " 3=%d", scale_data_size);

            const float scale_data[1] = {1.f};
            fwrite(scale_data, sizeof(float), 1, bp);
        }
        else if (n.op == "LeakyReLU")
        {
            std::string type = n.attr("act_type");
            if (type == "elu")
            {
                float slope = n.has_attr("slope") ? n.attr("slope") : 0.25f;
                fprintf(pp, " 0=%e", slope);
            }
            else if (type == "leaky" || type.empty())
            {
                float slope = n.has_attr("slope") ? n.attr("slope") : 0.25f;
                fprintf(pp, " 0=%e", slope);
            }
            else if (type == "prelu")
            {
                std::vector<float> weight_data = n.weight(0);

                fprintf(pp, " 0=%d", (int)weight_data.size());

                fwrite(weight_data.data(), sizeof(float), weight_data.size(), bp);
            }
        }
        else if (n.op == "LinearRegressionOutput")
        {
            // noop
        }
        else if (n.op == "log")
        {
            int op_type = 8;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "LogisticRegressionOutput")
        {
            // noop
        }
        else if (n.op == "MAERegressionOutput")
        {
            // noop
        }
        else if (n.op == "max" || n.op == "mean" || n.op == "min" || n.op == "prod" || n.op == "sum")
        {
            int operation = -233;
            if (n.op == "max") operation = 4;
            if (n.op == "mean") operation = 3;
            if (n.op == "min") operation = 5;
            if (n.op == "prod") operation = 6;
            if (n.op == "sum") operation = 0;

            std::vector<int> axis = n.attr("axis");
            int keepdims = n.attr("keepdims");

            fprintf(pp, " 0=%d", operation);
            if (axis.empty())
            {
                // if axis not set, reduce all axis by default
                fprintf(pp, " 1=%d", 1);
            }
            else
            {
                // if axis set, reduce according to axis
                fprintf(pp, " 1=%d", 0);
                fprintf(pp, " -23303=%zd", axis.size());
                for (size_t j = 0; j < axis.size(); j++)
                {
                    if (axis[j] == 0 || axis[j] > 4 || axis[j] < -3)
                        fprintf(stderr, "Unsupported reduction axis !\n");
                    fprintf(pp, ",%d", axis[j] > 0 ? axis[j] - 1 : axis[j]);
                }
            }
            fprintf(pp, " 4=%d", keepdims);
            fprintf(pp, " 5=1");
        }
        else if (n.op == "maximum")
        {
            int op_type = 4;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "minimum")
        {
            int op_type = 5;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "negative")
        {
            int op_type = 1;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "Pad")
        {
            std::string mode = n.attr("mode");
            std::vector<int> pad_width = n.attr("pad_width");
            float constant_value = n.attr("constant_value");

            int type = 0;
            if (mode == "constant")
            {
                type = 0;
            }
            else if (mode == "edge")
            {
                type = 1;
            }
            else if (mode == "reflect")
            {
                type = 2;
            }

            if (pad_width.size() != 8)
            {
                fprintf(stderr, "Unsupported pad_width !\n");
            }

            int channel_before = pad_width[2];
            int channel_after = pad_width[3];
            int top = pad_width[4];
            int bottom = pad_width[5];
            int left = pad_width[6];
            int right = pad_width[7];

            fprintf(pp, " 0=%d", top);
            fprintf(pp, " 1=%d", bottom);
            fprintf(pp, " 2=%d", left);
            fprintf(pp, " 3=%d", right);
            fprintf(pp, " 4=%d", type);
            fprintf(pp, " 5=%e", constant_value);
            fprintf(pp, " 7=%d", channel_before);
            fprintf(pp, " 8=%d", channel_after);
        }
        else if (n.op == "Pooling")
        {
            std::string pool_type = n.attr("pool_type");
            std::vector<int> kernel = n.attr("kernel");
            std::vector<int> stride = n.attr("stride");
            std::vector<int> pad = n.attr("pad");
            std::string pooling_convention = n.attr("pooling_convention");
            int global_pool = n.attr("global_pool");

            int pool = 0;
            if (pool_type == "max")
            {
                pool = 0;
            }
            else if (pool_type == "avg")
            {
                pool = 1;
            }

            int pad_mode = 1;
            if (pooling_convention == "valid")
            {
                pad_mode = 1;
            }
            else if (pooling_convention == "full")
            {
                pad_mode = 0;
            }

            fprintf(pp, " 0=%d", pool);

            if (kernel.size() == 1)
            {
                fprintf(pp, " 1=%d", kernel[0]);
            }
            else if (kernel.size() == 2)
            {
                fprintf(pp, " 1=%d", kernel[1]);
                fprintf(pp, " 11=%d", kernel[0]);
            }

            if (stride.size() == 1)
            {
                fprintf(pp, " 2=%d", stride[0]);
            }
            else if (stride.size() == 2)
            {
                fprintf(pp, " 2=%d", stride[1]);
                fprintf(pp, " 12=%d", stride[0]);
            }

            if (pad.size() == 1)
            {
                fprintf(pp, " 3=%d", pad[0]);
            }
            else if (pad.size() == 2)
            {
                fprintf(pp, " 3=%d", pad[1]);
                fprintf(pp, " 13=%d", pad[0]);
            }

            fprintf(pp, " 4=%d", global_pool);
            fprintf(pp, " 5=%d", pad_mode);

            if (pool_type == "avg")
            {
                int avgpool_count_include_pad = n.has_attr("count_include_pad") ? n.attr("count_include_pad") : 0;
                fprintf(pp, " 6=%d", avgpool_count_include_pad);
            }
        }
        else if (n.op == "reciprocal")
        {
            int op_type = 15;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "relu")
        {
            // no param
        }
        else if (n.op == "Reshape")
        {
            std::vector<int> shape = n.attr("shape");

            if (shape.size() == 1)
            {
                fprintf(pp, " 0=%d", shape[0]); // should never reach here
            }
            else if (shape.size() == 2)
            {
                fprintf(pp, " 0=%d", shape[1]);
            }
            else if (shape.size() == 3)
            {
                fprintf(pp, " 0=%d", shape[2]);
                fprintf(pp, " 1=%d", shape[1]);
            }
            else if (shape.size() == 4)
            {
                fprintf(pp, " 0=%d", shape[3]);
                fprintf(pp, " 1=%d", shape[2]);
                fprintf(pp, " 2=%d", shape[1]);
            }
            else if (shape.size() == 5)
            {
                fprintf(pp, " 0=%d", shape[4] * shape[3]);
                fprintf(pp, " 1=%d", shape[2]);
                fprintf(pp, " 2=%d", shape[1]);
            }
        }
        else if (n.op == "ShuffleChannel")
        {
            int group = n.attr("group");
            fprintf(pp, " 0=%d", group);
        }
        else if (n.op == "sigmoid")
        {
            // no param
        }
        else if (n.op == "sin")
        {
            int op_type = 9;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "slice")
        {
            std::vector<int> begin = n.attr("begin");
            std::vector<int> end = n.attr("end");
            std::vector<int> step = n.attr("step"); // TODO

            // skip N-dim
            begin.erase(begin.begin());
            end.erase(end.begin());
            if (step.size() != 0)
                step.erase(step.begin());

            // assert step == 1
            for (size_t j = 0; j < step.size(); j++)
            {
                if (step[j] != 1)
                    fprintf(stderr, "Unsupported slice step !\n");
            }

            fprintf(pp, " -23309=%d", (int)begin.size());
            for (size_t j = 0; j < begin.size(); j++)
            {
                fprintf(pp, ",%d", begin[j]);
            }
            fprintf(pp, " -23310=%d", (int)end.size());
            for (size_t j = 0; j < end.size(); j++)
            {
                fprintf(pp, ",%d", end[j]);
            }
        }
        else if (n.op == "slice_axis")
        {
            int axis = n.attr("axis");
            int begin = n.attr("begin");
            int end = n.has_attr("end") ? n.attr("end") : INT_MAX;

            if (axis == 0 || axis > 3 || axis < -3)
                fprintf(stderr, "Unsupported slice_axis axes !\n");

            if (axis > 0)
                axis = axis - 1; // -1 for skip N-dim

            fprintf(pp, " -23309=1,%d", begin);
            fprintf(pp, " -23310=1,%d", end);
            fprintf(pp, " -23311=1,%d", axis);
        }
        else if (n.op == "SliceChannel")
        {
            int num_outputs = n.attr("num_outputs");
            int squeeze_axis = n.attr("squeeze_axis"); // TODO
            if (squeeze_axis)
            {
                fprintf(stderr, "Unsupported SliceChannel squeeze_axis !\n");
            }

            fprintf(pp, " -23300=%d", num_outputs);
            for (int j = 0; j < num_outputs; j++)
            {
                fprintf(pp, ",-233");
            }
        }
        else if (n.op == "SoftmaxActivation")
        {
            std::string mode = n.attr("mode");
            if (mode != "channel")
            {
                fprintf(stderr, "Unsupported SoftmaxActivation mode !\n");
            }
            fprintf(pp, " 1=1");
        }
        else if (n.op == "SoftmaxOutput")
        {
            fprintf(pp, " 1=1");
        }
        else if (n.op == "softmax")
        {
            fprintf(pp, " 1=1");
        }
        else if (n.op == "sqrt")
        {
            int op_type = 5;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "square")
        {
            int op_type = 4;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "squeeze")
        {
            std::vector<int> axis = n.attr("axis");

            if (axis.empty())
            {
                fprintf(pp, " 0=1");
                fprintf(pp, " 1=1");
                fprintf(pp, " 2=1");
            }
            else
            {
                fprintf(pp, " -23303=%zd", axis.size());
                for (size_t j = 0; j < axis.size(); j++)
                {
                    fprintf(pp, ",%d", axis[j]);
                }
            }
        }
        else if (n.op == "tan")
        {
            int op_type = 11;
            fprintf(pp, " 0=%d", op_type);
        }
        else if (n.op == "tanh")
        {
            // no param
        }
        else if (n.op == "Transpose" || n.op == "transpose")
        {
            std::vector<int> axes = n.attr("axes");

            if (axes.size() == 3)
            {
                if (axes[1] == 2 && axes[2] == 1)
                    fprintf(pp, " 0=1"); // h w c
                else
                    fprintf(stderr, "Unsupported transpose type !\n");
            }
            else if (axes.size() == 4)
            {
                if (axes[1] == 1 && axes[2] == 2 && axes[3] == 3)
                    fprintf(pp, " 0=0"); // w h c
                else if (axes[1] == 1 && axes[2] == 3 && axes[3] == 2)
                    fprintf(pp, " 0=1"); // h w c
                else if (axes[1] == 2 && axes[2] == 1 && axes[3] == 3)
                    fprintf(pp, " 0=2"); // w c h
                else if (axes[1] == 2 && axes[2] == 3 && axes[3] == 1)
                    fprintf(pp, " 0=3"); // c w h
                else if (axes[1] == 3 && axes[2] == 1 && axes[3] == 2)
                    fprintf(pp, " 0=4"); // h c w
                else if (axes[1] == 3 && axes[2] == 2 && axes[3] == 1)
                    fprintf(pp, " 0=5"); // c h w
            }
            else if (axes.size() == 5)
            {
                if (axes[1] == 1 && axes[2] == 2 && axes[3] == 3 && axes[4] == 4)
                    fprintf(pp, " 0=0"); // wx h c
                else if (axes[1] == 1 && axes[2] == 3 && axes[3] == 4 && axes[4] == 2)
                    fprintf(pp, " 0=1"); // h wx c
                else if (axes[1] == 2 && axes[2] == 1 && axes[3] == 3 && axes[4] == 4)
                    fprintf(pp, " 0=2"); // wx c h
                else if (axes[1] == 2 && axes[2] == 3 && axes[3] == 4 && axes[4] == 1)
                    fprintf(pp, " 0=3"); // c wx h
                else if (axes[1] == 3 && axes[2] == 4 && axes[3] == 1 && axes[4] == 2)
                    fprintf(pp, " 0=4"); // h c wx
                else if (axes[1] == 3 && axes[2] == 4 && axes[3] == 2 && axes[4] == 1)
                    fprintf(pp, " 0=5"); // c h wx
                else
                    fprintf(stderr, "Unsupported transpose type !\n");
            }
            else
            {
                fprintf(stderr, "Unsupported transpose type !\n");
            }
        }
        else if (n.op == "UpSampling")
        {
            int scale = n.attr("scale");
            std::string sample_type = n.attr("sample_type");

            if (sample_type == "nearest")
            {
                fprintf(pp, " 0=1");
                fprintf(pp, " 1=%e", (float)scale);
                fprintf(pp, " 2=%e", (float)scale);
            }
            else if (sample_type == "bilinear")
            {
                // DeconvolutionDepthWise
                int num_filter = n.attr("num_filter");

                std::vector<float> weight_data = n.weight(0);

                int kernel = scale * 2 - scale % 2;
                int stride = scale;
                int pad = (scale - 1) / 2;

                fprintf(pp, " 0=%d", num_filter);
                fprintf(pp, " 1=%d", kernel);
                fprintf(pp, " 2=1");
                fprintf(pp, " 3=%d", stride);
                fprintf(pp, " 4=%d", pad);
                fprintf(pp, " 5=0");
                fprintf(pp, " 6=%d", (int)weight_data.size());
                fprintf(pp, " 7=%d", num_filter);

                int quantize_tag = 0;
                fwrite(&quantize_tag, sizeof(int), 1, bp);

                fwrite(weight_data.data(), sizeof(float), weight_data.size(), bp);
            }
        }
        else
        {
            // TODO op specific params
            std::map<std::string, std::string>::const_iterator attr_it = n.attrs.begin();
            for (; attr_it != n.attrs.end(); attr_it++)
            {
                fprintf(stderr, "# %s=%s\n", attr_it->first.c_str(), attr_it->second.c_str());
                //                 fprintf(pp, " %s=%s", attr_it->first.c_str(), attr_it->second.c_str());
            }
        }

        fprintf(pp, "\n");

        for (int j = 0; j < n.output_size; j++)
        {
            int input_uid = i | (j << 16);
            if (node_reference.find(input_uid) != node_reference.end())
            {
                int refcount = node_reference[input_uid];
                if (refcount > 1)
                {
                    std::string output_name = n.name;

                    char splitname[256];
                    sprintf(splitname, "splitncnn_%d", internal_split);
                    fprintf(pp, "%-16s %-32s %d %d", "Split", splitname, 1, refcount);
                    if (j == 0)
                    {
                        fprintf(pp, " %s", output_name.c_str());
                    }
                    else
                    {
                        fprintf(pp, " %s_subncnn_%d", output_name.c_str(), j);
                    }

                    for (int k = 0; k < refcount; k++)
                    {
                        if (j == 0)
                        {
                            fprintf(pp, " %s_splitncnn_%d", output_name.c_str(), k);
                        }
                        else
                        {
                            fprintf(pp, " %s_subncnn_%d_splitncnn_%d", output_name.c_str(), j, k);
                        }
                    }
                    fprintf(pp, "\n");

                    internal_split++;
                }
            }
        }
    }

    fclose(pp);
    fclose(bp);

    return 0;
}