ncnn/tools/pnnx/src/ir.cpp

// Tencent is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2021 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 "ir.h"

#include <limits.h>
#include <stdint.h>
#include <string.h>
#include <algorithm>
#include <fstream>
#include <sstream>
#include <string>
#include <stack>

#if BUILD_PNNX
#include <torch/script.h>
#endif

#include "storezip.h"

namespace pnnx {

static bool type_is_integer(int type)
{
    if (type == 1) return false;
    if (type == 2) return false;
    if (type == 3) return false;
    if (type == 4) return true;
    if (type == 5) return true;
    if (type == 6) return true;
    if (type == 7) return true;
    if (type == 8) return true;
    if (type == 9) return true;
    if (type == 10) return false;
    if (type == 11) return false;
    if (type == 12) return false;
    return false;
}

static const char* type_to_string(int type)
{
    if (type == 1) return "f32";
    if (type == 2) return "f64";
    if (type == 3) return "f16";
    if (type == 4) return "i32";
    if (type == 5) return "i64";
    if (type == 6) return "i16";
    if (type == 7) return "i8";
    if (type == 8) return "u8";
    if (type == 9) return "bool";
    if (type == 10) return "cp64";
    if (type == 11) return "cp128";
    if (type == 12) return "cp32";
    return "null";
}

static const char* type_to_numpy_string(int type)
{
    if (type == 1) return "float32";
    if (type == 2) return "float64";
    if (type == 3) return "float16";
    if (type == 4) return "int32";
    if (type == 5) return "int64";
    if (type == 6) return "int16";
    if (type == 7) return "int8";
    if (type == 8) return "uint8";
    if (type == 9) return "bool8";
    if (type == 10) return "csingle";
    if (type == 11) return "cdouble";
    if (type == 12) return "chalf";
    return "null";
}

static const char* type_to_dtype_string(int type)
{
    if (type == 1) return "torch.float";
    if (type == 2) return "torch.double";
    if (type == 3) return "torch.half";
    if (type == 4) return "torch.int";
    if (type == 5) return "torch.long";
    if (type == 6) return "torch.short";
    if (type == 7) return "torch.int8";
    if (type == 8) return "torch.uint8";
    if (type == 9) return "torch.bool";
    if (type == 10) return "torch.complex64";
    if (type == 11) return "torch.complex128";
    if (type == 12) return "torch.complex32";
    return "null";
}

static size_t type_to_elemsize(int type)
{
    if (type == 1) return 4;
    if (type == 2) return 8;
    if (type == 3) return 2;
    if (type == 4) return 4;
    if (type == 5) return 8;
    if (type == 6) return 2;
    if (type == 7) return 1;
    if (type == 8) return 1;
    if (type == 9) return 1;
    if (type == 10) return 8;
    if (type == 11) return 16;
    if (type == 12) return 4;
    return 0; // null
}

static int string_to_type(const char* s)
{
    if (strcmp(s, "f32") == 0) return 1;
    if (strcmp(s, "f64") == 0) return 2;
    if (strcmp(s, "f16") == 0) return 3;
    if (strcmp(s, "i32") == 0) return 4;
    if (strcmp(s, "i64") == 0) return 5;
    if (strcmp(s, "i16") == 0) return 6;
    if (strcmp(s, "i8") == 0) return 7;
    if (strcmp(s, "u8") == 0) return 8;
    if (strcmp(s, "bool") == 0) return 9;
    if (strcmp(s, "cp64") == 0) return 10;
    if (strcmp(s, "cp128") == 0) return 11;
    if (strcmp(s, "cp32") == 0) return 12;
    return 0; // null
}

#if BUILD_PNNX
int get_at_tensor_type(const at::ScalarType& st)
{
    if (st == c10::ScalarType::Float) return 1;
    if (st == c10::ScalarType::Double) return 2;
    if (st == c10::ScalarType::Half) return 3;
    if (st == c10::ScalarType::Int) return 4;
    if (st == c10::ScalarType::QInt32) return 4;
    if (st == c10::ScalarType::Long) return 5;
    if (st == c10::ScalarType::Short) return 6;
    if (st == c10::ScalarType::Char) return 7;
    if (st == c10::ScalarType::QInt8) return 7;
    if (st == c10::ScalarType::Byte) return 8;
    if (st == c10::ScalarType::QUInt8) return 8;
    if (st == c10::ScalarType::Bool) return 9;
    if (st == c10::ScalarType::ComplexFloat) return 10;
    if (st == c10::ScalarType::ComplexDouble) return 11;
    if (st == c10::ScalarType::ComplexHalf) return 12;
    return 0; // unknown type
}

Parameter::Parameter(const torch::jit::Node* value_node)
{
    type = 0;

    if (value_node->kind() == c10::prim::Constant)
    {
        if (value_node->output()->type()->kind() == c10::TypeKind::NoneType)
        {
            type = 0;
            return;
        }

        if (!value_node->hasAttribute(torch::jit::attr::value))
        {
            fprintf(stderr, "no attribute value\n");
            value_node->dump();
            return;
        }

        switch (value_node->output()->type()->kind())
        {
        case c10::TypeKind::NoneType:
        {
            type = 0;
            break;
        }
        case c10::TypeKind::BoolType:
        {
            type = 1;
            b = value_node->i(torch::jit::attr::value);
            break;
        }
        case c10::TypeKind::IntType:
        {
            type = 2;
            int64_t i64 = value_node->i(torch::jit::attr::value);
            if (i64 == std::numeric_limits<int64_t>::max()) i64 = INT_MAX;
            if (i64 == std::numeric_limits<int64_t>::max()) i64 = INT_MIN;
            i = (int)i64;
            break;
        }
        case c10::TypeKind::FloatType:
        {
            type = 3;
            f = (float)value_node->f(torch::jit::attr::value);
            break;
        }
        case c10::TypeKind::StringType:
        {
            type = 4;
            s = value_node->s(torch::jit::attr::value);
            break;
        }
        case c10::TypeKind::DeviceObjType:
        {
            type = 4;
            s = value_node->s(torch::jit::attr::value);
            break;
        }
        case c10::TypeKind::TensorType:
        {
            at::Tensor t = value_node->t(torch::jit::attr::value);

            if (t.dim() == 0)
            {
                if (t.scalar_type() == c10::ScalarType::Long)
                {
                    type = 2;
                    int64_t i64 = t.item<int64_t>();
                    if (i64 == std::numeric_limits<int64_t>::max()) i64 = INT_MAX;
                    if (i64 == std::numeric_limits<int64_t>::max()) i64 = INT_MIN;
                    i = (int)i64;
                }
                else if (t.scalar_type() == c10::ScalarType::Int)
                {
                    type = 2;
                    i = t.item<int>();
                }
                else if (t.scalar_type() == c10::ScalarType::Double)
                {
                    type = 3;
                    f = (float)t.item<double>();
                }
                else if (t.scalar_type() == c10::ScalarType::Float)
                {
                    type = 3;
                    f = t.item<float>();
                }
                else
                {
                    fprintf(stderr, "unknown Parameter value kind %s of TensorType, t.dim = 0\n", value_node->kind().toDisplayString());
                }
            }
            else
            {
                const int ndim = (int)t.dim();

                type = 8;
                fprintf(stderr, "unknown Parameter value kind %s of TensorType, t.dim = %d\n", value_node->kind().toDisplayString(), ndim);
            }

            break;
        }
        default:
        {
            fprintf(stderr, "unknown Parameter value kind %s\n", c10::typeKindToString(value_node->output()->type()->kind()));
            break;
        }
        }
    }
    else if (value_node->kind() == c10::prim::ListConstruct)
    {
        switch (value_node->output()->type()->cast<c10::ListType>()->getElementType()->kind())
        {
        case c10::TypeKind::IntType:
        {
            type = 5;
            for (const auto& x : value_node->inputs())
            {
                ai.push_back((int)x->node()->i(torch::jit::attr::value));
            }
            break;
        }
        case c10::TypeKind::FloatType:
        {
            type = 6;
            for (const auto& x : value_node->inputs())
            {
                af.push_back((float)x->node()->f(torch::jit::attr::value));
            }
            break;
        }
        case c10::TypeKind::StringType:
        {
            type = 7;
            for (const auto& x : value_node->inputs())
            {
                as.push_back(x->node()->s(torch::jit::attr::value));
            }
            break;
        }
        default:
        {
            fprintf(stderr, "unknown Parameter value list element kind %s\n", c10::typeKindToString(value_node->output()->type()->cast<c10::ListType>()->getElementType()->kind()));
            break;
        }
        }
    }
    else
    {
        fprintf(stderr, "unknown Parameter value_node kind %s\n", value_node->kind().toDisplayString());
    }
}

Parameter::Parameter(const torch::jit::Value* value)
    : Parameter(value->node())
{
}
#endif // BUILD_PNNX

bool operator==(const Parameter& lhs, const Parameter& rhs)
{
    if (lhs.type != rhs.type)
        return false;

    if (lhs.type == 0)
        return true;

    if (lhs.type == 1 && lhs.b == rhs.b)
        return true;

    if (lhs.type == 2 && lhs.i == rhs.i)
        return true;

    if (lhs.type == 3 && lhs.f == rhs.f)
        return true;

    if (lhs.type == 4 && lhs.s == rhs.s)
        return true;

    if (lhs.type == 5 && lhs.ai == rhs.ai)
        return true;

    if (lhs.type == 6 && lhs.af == rhs.af)
        return true;

    if (lhs.type == 7 && lhs.as == rhs.as)
        return true;

    return false;
}

#if BUILD_PNNX
Attribute::Attribute(const at::Tensor& t)
{
    type = get_at_tensor_type(t.scalar_type());

    const int ndim = (int)t.dim();

    if (ndim == 0)
    {
        shape = {1};

        data.resize(type_to_elemsize(type));

        if (t.scalar_type() == c10::ScalarType::Long)
        {
            int64_t i = t.item<int64_t>();
            memcpy((void*)data.data(), (const void*)&i, data.size());
        }
        else if (t.scalar_type() == c10::ScalarType::Int)
        {
            int i = t.item<int>();
            memcpy((void*)data.data(), (const void*)&i, data.size());
        }
        else if (t.scalar_type() == c10::ScalarType::Double)
        {
            double f = t.item<double>();
            memcpy((void*)data.data(), (const void*)&f, data.size());
        }
        else if (t.scalar_type() == c10::ScalarType::Float)
        {
            float f = t.item<float>();
            memcpy((void*)data.data(), (const void*)&f, data.size());
        }
        else
        {
            fprintf(stderr, "unknown Attribute tensor scalar type %d\n", type);
        }

        return;
    }

    shape.resize(ndim);
    for (int i = 0; i < ndim; i++)
        shape[i] = t.size(i);

    if (shape.size() > 0)
    {
        int size = shape[0];
        for (size_t i = 1; i < shape.size(); i++)
        {
            size *= shape[i];
        }

        data.resize(size * type_to_elemsize(type));
        memcpy((void*)data.data(), (const void*)t.cpu().contiguous().data_ptr(), data.size());
    }
}
#endif // BUILD_PNNX

Attribute::Attribute(const std::initializer_list<int>& _shape, const std::vector<float>& t)
{
    type = 1;
    shape = _shape;

    if (shape.size() > 0)
    {
        int size = shape[0];
        for (size_t i = 1; i < shape.size(); i++)
        {
            size *= shape[i];
        }

        data.resize(size * type_to_elemsize(type));
        memcpy((void*)data.data(), (const void*)t.data(), data.size());
    }
}

bool operator==(const Attribute& lhs, const Attribute& rhs)
{
    if (lhs.type != rhs.type)
        return false;

    if (lhs.type == 0)
        return true;

    if (lhs.shape != rhs.shape)
        return false;

    if (lhs.data != rhs.data)
        return false;

    return true;
}

Attribute operator+(const Attribute& a, const Attribute& b)
{
    Attribute c;

    if (a.type != b.type)
    {
        fprintf(stderr, "concat attribute type mismatch\n");
        return c;
    }

    if (a.shape.size() != b.shape.size())
    {
        fprintf(stderr, "concat attribute shape rank mismatch\n");
        return c;
    }

    for (int i = 1; i < (int)a.shape.size(); i++)
    {
        if (a.shape[i] != b.shape[i])
        {
            fprintf(stderr, "concat attribute shape mismatch\n");
            return c;
        }
    }

    c.type = a.type;
    c.shape = a.shape;
    c.shape[0] += b.shape[0]; // concat the first dim

    c.data.resize(a.data.size() + b.data.size());
    memcpy(c.data.data(), a.data.data(), a.data.size());
    memcpy(c.data.data() + a.data.size(), b.data.data(), b.data.size());

    return c;
}

Parameter Parameter::parse_from_string(const std::string& value)
{
    Parameter p;
    p.type = 0;

    if (value == "None" || value == "()" || value == "[]")
    {
        return p;
    }

    if (value == "True" || value == "False")
    {
        // bool
        p.type = 1;
        p.b = value == "True";
        return p;
    }

    if (value[0] == '(' || value[0] == '[')
    {
        // list
        std::string lc = value.substr(1, value.size() - 2);
        std::istringstream lcss(lc);

        while (!lcss.eof())
        {
            std::string elem;
            std::getline(lcss, elem, ',');

            if ((elem[0] != '-' && (elem[0] < '0' || elem[0] > '9')) || (elem[0] == '-' && (elem[1] < '0' || elem[1] > '9')))
            {
                // string
                p.type = 7;
                p.as.push_back(elem);
            }
            else if (elem.find('.') != std::string::npos || elem.find('e') != std::string::npos)
            {
                // float
                p.type = 6;
                p.af.push_back(std::stof(elem));
            }
            else
            {
                // integer
                p.type = 5;
                p.ai.push_back(std::stoi(elem));
            }
        }
        return p;
    }

    if ((value[0] != '-' && (value[0] < '0' || value[0] > '9')) || (value[0] == '-' && (value[1] < '0' || value[1] > '9')))
    {
        // string
        p.type = 4;
        p.s = value;
        return p;
    }

    if (value.find('.') != std::string::npos || value.find('e') != std::string::npos)
    {
        // float
        p.type = 3;
        p.f = std::stof(value);
        return p;
    }

    // integer
    p.type = 2;
    p.i = std::stoi(value);
    return p;
}

Graph::Graph()
{
}

Graph::~Graph()
{
    for (auto x : ops)
        delete x;

    for (auto x : operands)
        delete x;

    ops.clear();
    operands.clear();
}

Graph::Graph(const Graph& /*rhs*/)
{
}

Graph& Graph::operator=(const Graph& /*rhs*/)
{
    return *this;
}

static void load_parameter(Operator* op, const std::string& key, const std::string& value)
{
    op->params[key] = Parameter::parse_from_string(value);
}

static void load_input_key(Operator* op, const std::string& key, const std::string& value)
{
    op->inputnames.resize(op->inputs.size());

    for (size_t i = 0; i < op->inputs.size(); i++)
    {
        const Operand* oprand = op->inputs[i];
        if (oprand->name == value)
        {
            op->inputnames[i] = key;
            break;
        }
    }
}

static void load_shape(Operator* op, const std::string& key, const std::string& value)
{
    Operand* operand = 0;
    for (auto r : op->inputs)
    {
        if (r->name == key)
        {
            operand = r;
            break;
        }
    }

    if (!operand)
    {
        for (auto r : op->outputs)
        {
            if (r->name == key)
            {
                operand = r;
                break;
            }
        }
    }

    if (!operand)
    {
        fprintf(stderr, "no such operand %s for operator %s\n", key.c_str(), op->name.c_str());
        return;
    }

    // type
    std::string typestr = value.substr(value.find_last_of(')') + 1);
    operand->type = string_to_type(typestr.c_str());

    // shape
    std::string lc = value.substr(1, value.find_last_of(')') - 1);
    std::istringstream lcss(lc);

    operand->shape.clear();
    while (!lcss.eof())
    {
        std::string elem;
        std::getline(lcss, elem, ',');

        if (elem == "?")
        {
            operand->shape.push_back(-1);
        }
        else
        {
            int i = std::stoi(elem);
            operand->shape.push_back(i);
        }
    }
}

static void load_attribute(Operator* op, const std::string& key, const std::string& value, StoreZipReader& szr)
{
    Attribute& a = op->attrs[key];

    // type
    std::string typestr = value.substr(value.find_last_of(')') + 1);
    a.type = string_to_type(typestr.c_str());

    if (a.type == 0)
        return;

    // shape
    std::string lc = value.substr(1, value.find_last_of(')') - 1);
    std::istringstream lcss(lc);

    a.shape.clear();
    while (!lcss.eof())
    {
        std::string elem;
        std::getline(lcss, elem, ',');

        int i = std::stoi(elem);
        a.shape.push_back(i);
    }

    if (a.shape.empty())
        return;

    // data
    size_t size = 1;
    for (int i : a.shape)
    {
        size *= i;
    }

    size_t bytesize = size * type_to_elemsize(a.type);

    std::string filename = op->name + "." + key;

    size_t filesize = szr.get_file_size(filename);

    if (filesize == 0)
    {
        // no such file
        return;
    }

    if (filesize != bytesize)
    {
        fprintf(stderr, "file size not match expect %lu but got %lu\n", bytesize, filesize);
    }

    a.data.resize(bytesize);
    szr.read_file(filename, (char*)a.data.data());
}

int Graph::load(const std::string& parampath, const std::string& binpath)
{
    std::ifstream is(parampath, std::ios::in | std::ios::binary);
    if (!is.good())
    {
        fprintf(stderr, "open failed\n");
        return -1;
    }

    StoreZipReader szr;
    if (szr.open(binpath) != 0)
    {
        fprintf(stderr, "open failed\n");
        return -1;
    }

    int magic = 0;
    {
        std::string line;
        std::getline(is, line);
        std::istringstream iss(line);

        iss >> magic;
    }

    int operator_count = 0;
    int operand_count = 0;
    {
        std::string line;
        std::getline(is, line);
        std::istringstream iss(line);

        iss >> operator_count >> operand_count;
    }

    for (int i = 0; i < operator_count; i++)
    {
        std::string line;
        std::getline(is, line);
        std::istringstream iss(line);

        std::string type;
        std::string name;
        int input_count = 0;
        int output_count = 0;

        iss >> type >> name >> input_count >> output_count;

        Operator* op = new_operator(type, name);

        for (int j = 0; j < input_count; j++)
        {
            std::string operand_name;
            iss >> operand_name;

            Operand* r = get_operand(operand_name);
            r->consumers.push_back(op);
            op->inputs.push_back(r);
        }

        for (int j = 0; j < output_count; j++)
        {
            std::string operand_name;
            iss >> operand_name;

            Operand* r = new_operand(operand_name);
            r->producer = op;
            op->outputs.push_back(r);
        }

        // key=value
        while (!iss.eof())
        {
            std::string param;
            iss >> param;

            std::string key;
            std::string value;
            std::istringstream pss(param);
            std::getline(pss, key, '=');
            std::getline(pss, value);

            if (key[0] == '@')
            {
                // attribute
                load_attribute(op, key.substr(1), value, szr);
            }
            else if (key[0] == '$')
            {
                // operand input key
                load_input_key(op, key.substr(1), value);
            }
            else if (key[0] == '#')
            {
                // operand shape
                load_shape(op, key.substr(1), value);
            }
            else
            {
                // parameter
                load_parameter(op, key, value);
            }
        }
    }

    return 0;
}

int Graph::save(const std::string& parampath, const std::string& binpath)
{
    FILE* paramfp = fopen(parampath.c_str(), "wb");
    if (!paramfp)
    {
        fprintf(stderr, "fopen %s failed\n", parampath.c_str());
        return -1;
    }

    StoreZipWriter szw;
    if (szw.open(binpath) != 0)
    {
        fprintf(stderr, "open failed\n");
        return -1;
    }

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

    // op count and oprand count
    fprintf(paramfp, "%d %d\n", (int)ops.size(), (int)operands.size());

    for (const Operator* op : ops)
    {
        fprintf(paramfp, "%-24s %-24s %d %d", op->type.c_str(), op->name.c_str(), (int)op->inputs.size(), (int)op->outputs.size());

        for (const Operand* oprand : op->inputs)
        {
            fprintf(paramfp, " %s", oprand->name.c_str());
        }

        for (const Operand* oprand : op->outputs)
        {
            fprintf(paramfp, " %s", oprand->name.c_str());
        }

        for (const auto& it : op->params)
        {
            fprintf(paramfp, " %s=", it.first.c_str());

            const Parameter& param = it.second;
            if (param.type == 0)
            {
                fprintf(paramfp, "None");
            }
            if (param.type == 1)
            {
                if (param.b)
                    fprintf(paramfp, "True");
                else
                    fprintf(paramfp, "False");
            }
            if (param.type == 2)
            {
                fprintf(paramfp, "%d", param.i);
            }
            if (param.type == 3)
            {
                fprintf(paramfp, "%e", param.f);
            }
            if (param.type == 4)
            {
                fprintf(paramfp, "%s", param.s.c_str());
            }
            if (param.type == 5)
            {
                fprintf(paramfp, "(");
                for (size_t i = 0; i < param.ai.size(); i++)
                {
                    fprintf(paramfp, "%d", param.ai[i]);
                    if (i + 1 != param.ai.size())
                        fprintf(paramfp, ",");
                }
                fprintf(paramfp, ")");
            }
            if (param.type == 6)
            {
                fprintf(paramfp, "(");
                for (size_t i = 0; i < param.af.size(); i++)
                {
                    fprintf(paramfp, "%e", param.af[i]);
                    if (i + 1 != param.af.size())
                        fprintf(paramfp, ",");
                }
                fprintf(paramfp, ")");
            }
            if (param.type == 7)
            {
                fprintf(paramfp, "(");
                for (size_t i = 0; i < param.as.size(); i++)
                {
                    fprintf(paramfp, "%s", param.as[i].c_str());
                    if (i + 1 != param.as.size())
                        fprintf(paramfp, ",");
                }
                fprintf(paramfp, ")");
            }
        }

        for (const auto& it : op->attrs)
        {
            fprintf(paramfp, " @%s=", it.first.c_str());

            const Attribute& attr = it.second;
            fprintf(paramfp, "(");
            for (int i = 0; i < (int)attr.shape.size() - 1; i++)
            {
                fprintf(paramfp, "%d,", attr.shape[i]);
            }
            if (attr.shape.size() > 0)
                fprintf(paramfp, "%d", attr.shape[attr.shape.size() - 1]);
            fprintf(paramfp, ")");

            fprintf(paramfp, type_to_string(attr.type));

            std::string filename = op->name + "." + it.first;
            szw.write_file(filename, attr.data.data(), attr.data.size());
        }

        if (op->inputnames.size() == op->inputs.size())
        {
            for (size_t i = 0; i < op->inputs.size(); i++)
            {
                if (op->inputnames[i].empty())
                    continue;

                const Operand* oprand = op->inputs[i];
                fprintf(paramfp, " $%s=%s", op->inputnames[i].c_str(), oprand->name.c_str());
            }
        }

        for (const Operand* oprand : op->inputs)
        {
            if (oprand->shape.empty())
                continue;

            fprintf(paramfp, " #%s=", oprand->name.c_str());

            fprintf(paramfp, "(");
            for (int i = 0; i < (int)oprand->shape.size() - 1; i++)
            {
                if (oprand->shape[i] == -1)
                    fprintf(paramfp, "?,");
                else
                    fprintf(paramfp, "%d,", oprand->shape[i]);
            }
            if (oprand->shape.size() > 0)
            {
                if (oprand->shape[oprand->shape.size() - 1] == -1)
                    fprintf(paramfp, "?");
                else
                    fprintf(paramfp, "%d", oprand->shape[oprand->shape.size() - 1]);
            }
            fprintf(paramfp, ")");

            fprintf(paramfp, type_to_string(oprand->type));
        }

        for (const Operand* oprand : op->outputs)
        {
            if (oprand->shape.empty())
                continue;

            fprintf(paramfp, " #%s=", oprand->name.c_str());

            fprintf(paramfp, "(");
            for (int i = 0; i < (int)oprand->shape.size() - 1; i++)
            {
                if (oprand->shape[i] == -1)
                    fprintf(paramfp, "?,");
                else
                    fprintf(paramfp, "%d,", oprand->shape[i]);
            }
            if (oprand->shape.size() > 0)
            {
                if (oprand->shape[oprand->shape.size() - 1] == -1)
                    fprintf(paramfp, "?");
                else
                    fprintf(paramfp, "%d", oprand->shape[oprand->shape.size() - 1]);
            }
            fprintf(paramfp, ")");

            fprintf(paramfp, type_to_string(oprand->type));
        }

        fprintf(paramfp, "\n");
    }

    fclose(paramfp);

    return 0;
}

static std::string sanitize_identifier(const std::string& s)
{
    std::string ss = s;
    for (size_t i = 0; i < ss.size(); i++)
    {
        if (ss[i] == '.' || ss[i] == ':')
            ss[i] = '_';
    }

    return ss;
}

static std::string expand_expression(const Operator* op)
{
    std::string expr = op->params.at("expr").s;

    // split into tokens
    std::vector<std::string> tokens;
    {
        std::string t;
        for (size_t i = 0; i < expr.size(); i++)
        {
            char ch = expr[i];

            if (ch == '[') // list
            {
                t += ch;
                tokens.push_back(t);
                t.clear();
            }
            else if (ch == '(' || ch == ')' || ch == ',' || ch == ']')
            {
                if (!t.empty())
                {
                    tokens.push_back(t);
                    t.clear();
                }
            }
            else
            {
                t += ch;
            }
        }

        if (!t.empty())
        {
            tokens.push_back(t);
        }
    }

    // scan and stack
    std::stack<std::string> exprstack;
    for (int i = (int)tokens.size() - 1; i >= 0; i--)
    {
        const std::string& t = tokens[i];

        if (t == "size")
        {
            std::string a = exprstack.top();
            exprstack.pop();
            std::string b = exprstack.top();
            exprstack.pop();

            std::string r = a + ".size(" + b + ")";
            exprstack.push(r);
        }
        else if (t == "int"
                 || t == "abs"
                 || t == "acos"
                 || t == "acosh"
                 || t == "asin"
                 || t == "asinh"
                 || t == "atan"
                 || t == "atanh"
                 || t == "ceil"
                 || t == "cos"
                 || t == "cosh"
                 || t == "exp"
                 || t == "floor"
                 || t == "log"
                 || t == "log10"
                 || t == "neg"
                 || t == "reciprocal"
                 || t == "rsqrt"
                 || t == "sign"
                 || t == "sin"
                 || t == "sinh"
                 || t == "sqrt"
                 || t == "square"
                 || t == "tan"
                 || t == "tanh"
                 || t == "trunc")
        {
            std::string unaryop;
            if (t == "int") unaryop = "int";
            if (t == "abs") unaryop = "torch.abs";
            if (t == "acos") unaryop = "torch.acos";
            if (t == "acosh") unaryop = "torch.acosh";
            if (t == "asin") unaryop = "torch.asin";
            if (t == "asinh") unaryop = "torch.asinh";
            if (t == "atan") unaryop = "torch.atan";
            if (t == "atanh") unaryop = "torch.atanh";
            if (t == "ceil") unaryop = "torch.ceil";
            if (t == "cos") unaryop = "torch.cos";
            if (t == "cosh") unaryop = "torch.cosh";
            if (t == "exp") unaryop = "torch.exp";
            if (t == "floor") unaryop = "torch.floor";
            if (t == "log") unaryop = "torch.log";
            if (t == "log10") unaryop = "torch.log10";
            if (t == "neg") unaryop = "torch.neg";
            if (t == "reciprocal") unaryop = "torch.reciprocal";
            if (t == "rsqrt") unaryop = "torch.rsqrt";
            if (t == "sign") unaryop = "torch.sign";
            if (t == "sin") unaryop = "torch.sin";
            if (t == "sinh") unaryop = "torch.sinh";
            if (t == "sqrt") unaryop = "torch.sqrt";
            if (t == "square") unaryop = "torch.square";
            if (t == "tan") unaryop = "torch.tan";
            if (t == "tanh") unaryop = "torch.tanh";
            if (t == "trunc") unaryop = "torch.trunc";

            std::string a = exprstack.top();
            exprstack.pop();

            std::string r = unaryop + "(" + a + ")";
            exprstack.push(r);
        }
        else if (t == "atan2"
                 || t == "pow")
        {
            std::string binaryop;
            if (t == "atan2") binaryop = "torch.atan2";
            if (t == "pow") binaryop = "torch.pow";

            std::string a = exprstack.top();
            exprstack.pop();
            std::string b = exprstack.top();
            exprstack.pop();

            std::string r = binaryop + "(" + a + ", " + b + ")";
            exprstack.push(r);
        }
        else if (t == "add" || t == "sub" || t == "mul" || t == "div" || t == "floor_divide" || t == "and" || t == "or" || t == "xor" || t == "lshift" || t == "rshift")
        {
            std::string binaryop;
            if (t == "add") binaryop = "+";
            if (t == "sub") binaryop = "-";
            if (t == "mul") binaryop = "*";
            if (t == "div") binaryop = "/";
            if (t == "floor_divide") binaryop = "//";
            if (t == "and") binaryop = "&";
            if (t == "or") binaryop = "|";
            if (t == "xor") binaryop = "^";
            if (t == "lshift") binaryop = "<<";
            if (t == "rshift") binaryop = ">>";

            std::string a = exprstack.top();
            exprstack.pop();
            std::string b = exprstack.top();
            exprstack.pop();

            std::string r = std::string("(") + a + " " + binaryop + " " + b + ")";
            exprstack.push(r);
        }
        else if (t == "[") // list
        {
            std::vector<std::string> elements;
            while (!exprstack.empty())
            {
                std::string a = exprstack.top();
                exprstack.pop();

                elements.push_back(a);
            }

            std::string r = "[";
            for (int j = 0; j < (int)elements.size() - 1; j++)
            {
                r += elements[j];
                if (j + 1 != (int)elements.size())
                    r += ", ";
            }
            if (!elements.empty())
            {
                r += elements[elements.size() - 1];
            }
            r += "]";

            exprstack.push(r);
        }
        else if (t[0] == '@')
        {
            int input_index = std::stoi(t.substr(1));
            std::string varid = std::string("v_") + sanitize_identifier(op->inputs[input_index]->name);
            exprstack.push(varid);
        }
        else
        {
            // literal
            exprstack.push(t);
        }
    }

    std::string r = exprstack.top();
    exprstack.pop();

    return r;
}

static std::string make_slice_expression(const Operator* op)
{
    for (size_t j = 0; j < op->inputnames.size(); j++)
    {
        fprintf(stderr, "make_slice_expression %s %s\n", op->inputnames[j].c_str(), op->inputs[j]->name.c_str());
    }

    std::vector<int> dims;
    if (op->params.find("dims") != op->params.end())
    {
        dims = op->params.at("dims").ai;
    }
    else
    {
        dims.push_back(op->params.at("dim").i);
    }

    std::string r;

    int last_dim = -1;
    const int ndim = (int)dims.size();
    for (int i = 0; i < ndim; i++)
    {
        int dim = dims[i];
        for (int j = last_dim + 1; j < dim; j++)
        {
            r += ":,";
        }
        last_dim = dim;

        if (op->params.find("starts") != op->params.end())
        {
            std::vector<int> starts = op->params.at("starts").ai;
            int start = starts[i];

            if (start != 0)
                r += std::to_string(start);
        }
        else
        {
            fprintf(stderr, "find start\n");
            // find start
            for (size_t j = 0; j < op->inputnames.size(); j++)
            {
                if (op->inputnames[j] == "start")
                {
                    r += std::string("v_") + sanitize_identifier(op->inputs[j]->name);

                    fprintf(stderr, "find start %s\n", op->inputs[j]->name.c_str());
                    break;
                }
            }
        }

        r += ':';

        if (op->params.find("ends") != op->params.end())
        {
            std::vector<int> ends = op->params.at("ends").ai;
            int end = ends[i];
            if (end != INT_MAX)
                r += std::to_string(end);
        }
        else
        {
            // find end
            for (size_t j = 0; j < op->inputnames.size(); j++)
            {
                if (op->inputnames[j] == "end")
                {
                    r += std::string("v_") + sanitize_identifier(op->inputs[j]->name);
                    break;
                }
            }
        }

        if (op->params.find("steps") != op->params.end())
        {
            std::vector<int> steps = op->params.at("steps").ai;
            int step = steps[i];
            if (step != 1)
            {
                r += ':';
                r += std::to_string(step);
            }
        }
        else
        {
            // find step
            for (size_t j = 0; j < op->inputnames.size(); j++)
            {
                if (op->inputnames[j] == "step")
                {
                    r += ':';
                    r += std::string("v_") + sanitize_identifier(op->inputs[j]->name);
                    break;
                }
            }
        }

        if (i + 1 != ndim)
            r += ',';
    }

    return r;
}

static std::string make_index_expression(const Operator* op)
{
    fprintf(stderr, "make_index_expression %s\n", op->name.c_str());

    std::string index_expr = op->params.at("expr").s;

    // strip out-most [ ] pair
    index_expr = index_expr.substr(1, index_expr.size() - 2);

    // None,None,   ->   ...,
    bool leading_none = false;
    while (index_expr.substr(0, 5) == "None,")
    {
        leading_none = true;
        index_expr = index_expr.substr(5);
    }
    if (leading_none)
    {
        index_expr = "...," + index_expr;
    }

    return index_expr;
}

int Graph::python(const std::string& pypath, const std::string& pnnxbinpath)
{
    FILE* pyfp = fopen(pypath.c_str(), "wb");
    if (!pyfp)
    {
        fprintf(stderr, "fopen %s failed\n", pypath.c_str());
        return -1;
    }

    fprintf(pyfp, "import os\n");
    fprintf(pyfp, "import numpy as np\n");
    fprintf(pyfp, "import tempfile, zipfile\n");
    fprintf(pyfp, "import torch\n");
    fprintf(pyfp, "import torch.nn as nn\n");
    fprintf(pyfp, "import torch.nn.functional as F\n");
    fprintf(pyfp, "try:\n");
    fprintf(pyfp, "    import torchvision\n");
    fprintf(pyfp, "except:\n");
    fprintf(pyfp, "    pass\n");

    fprintf(pyfp, "\n");

    fprintf(pyfp, "class Model(nn.Module):\n");
    fprintf(pyfp, "    def __init__(self):\n");
    fprintf(pyfp, "        super(Model, self).__init__()\n");

    fprintf(pyfp, "\n");

    // module
    {
        for (const Operator* op : ops)
        {
            if (op->type.substr(0, 3) != "nn." && op->type.substr(0, 16) != "torchvision.ops.")
                continue;

            fprintf(pyfp, "        self.%s = %s(", sanitize_identifier(op->name).c_str(), op->type.c_str());

            int param_count = op->params.size();
            if (op->type == "nn.quantized.Conv2d" || op->type == "nn.quantized.Linear")
            {
                param_count -= 2; // ignore scale and zero_point
            }

            int param_index = 0;
            for (const auto& it : op->params)
            {
                if (op->type == "nn.quantized.Conv2d" || op->type == "nn.quantized.Linear")
                {
                    if (it.first == "scale" || it.first == "zero_point")
                        continue;
                }

                fprintf(pyfp, "%s=", it.first.c_str());

                const Parameter& param = it.second;
                if (param.type == 0)
                {
                    fprintf(pyfp, "None");
                }
                if (param.type == 1)
                {
                    if (param.b)
                        fprintf(pyfp, "True");
                    else
                        fprintf(pyfp, "False");
                }
                if (param.type == 2)
                {
                    fprintf(pyfp, "%d", param.i);
                }
                if (param.type == 3)
                {
                    fprintf(pyfp, "%f", param.f);
                }
                if (param.type == 4)
                {
                    if (param.s.substr(0, 6) == "torch.")
                    {
                        fprintf(pyfp, "%s", param.s.c_str());
                    }
                    else
                    {
                        fprintf(pyfp, "\'%s\'", param.s.c_str());
                    }
                }
                if (param.type == 5)
                {
                    fprintf(pyfp, "(");
                    for (size_t i = 0; i < param.ai.size(); i++)
                    {
                        fprintf(pyfp, "%d", param.ai[i]);
                        if (i + 1 != param.ai.size() || param.ai.size() == 1)
                            fprintf(pyfp, ",");
                    }
                    fprintf(pyfp, ")");
                }
                if (param.type == 6)
                {
                    fprintf(pyfp, "(");
                    for (size_t i = 0; i < param.af.size(); i++)
                    {
                        fprintf(pyfp, "%f", param.af[i]);
                        if (i + 1 != param.af.size() || param.af.size() == 1)
                            fprintf(pyfp, ",");
                    }
                    fprintf(pyfp, ")");
                }
                if (param.type == 7)
                {
                    fprintf(pyfp, "(");
                    for (size_t i = 0; i < param.as.size(); i++)
                    {
                        if (param.as[i].substr(0, 6) == "torch.")
                        {
                            fprintf(pyfp, "%s", param.as[i].c_str());
                        }
                        else
                        {
                            fprintf(pyfp, "\'%s\'", param.as[i].c_str());
                        }
                        if (i + 1 != param.as.size() || param.as.size() == 1)
                            fprintf(pyfp, ",");
                    }
                    fprintf(pyfp, ")");
                }

                param_index++;
                if (param_index != param_count)
                    fprintf(pyfp, ", ");
            }

            fprintf(pyfp, ")\n");
        }
    }

    fprintf(pyfp, "\n");

    // load weights
    {
        fprintf(pyfp, "        archive = zipfile.ZipFile('%s', 'r')\n", pnnxbinpath.c_str());

        for (const Operator* op : ops)
        {
            if (op->type.substr(0, 3) != "nn." && op->type.substr(0, 16) != "torchvision.ops.")
                continue;

            if (op->type == "nn.quantized.Conv2d" || op->type == "nn.quantized.Linear")
            {
                for (const auto& it : op->attrs)
                {
                    if (it.first == "weight" || it.first == "bias")
                    {
                        fprintf(pyfp, "        self_%s_%s = self.load_pnnx_bin_as_parameter(archive, '%s.%s', (", sanitize_identifier(op->name).c_str(), it.first.c_str(), op->name.c_str(), it.first.c_str());
                    }
                    else
                    {
                        // unknown attr
                        continue;
                    }

                    const Attribute& attr = it.second;
                    for (size_t i = 0; i < attr.shape.size(); i++)
                    {
                        fprintf(pyfp, "%d", attr.shape[i]);
                        if (i + 1 != attr.shape.size())
                            fprintf(pyfp, ",");
                    }

                    fprintf(pyfp, "), '%s', requires_grad=False)\n", type_to_numpy_string(attr.type));
                }

                fprintf(pyfp, "        self.%s.set_weight_bias(self_%s_weight, self_%s_bias)\n", sanitize_identifier(op->name).c_str(), sanitize_identifier(op->name).c_str(), sanitize_identifier(op->name).c_str());
                fprintf(pyfp, "        self.%s.scale = %f\n", sanitize_identifier(op->name).c_str(), op->params.at("scale").f);
                fprintf(pyfp, "        self.%s.zero_point = %d\n", sanitize_identifier(op->name).c_str(), op->params.at("zero_point").i);

                continue;
            }

            for (const auto& it : op->attrs)
            {
                if (it.first == "running_mean" || it.first == "running_var")
                {
                    fprintf(pyfp, "        self.%s.%s = self.load_pnnx_bin_as_tensor(archive, '%s.%s', (", sanitize_identifier(op->name).c_str(), it.first.c_str(), op->name.c_str(), it.first.c_str());
                }
                else
                {
                    fprintf(pyfp, "        self.%s.%s = self.load_pnnx_bin_as_parameter(archive, '%s.%s', (", sanitize_identifier(op->name).c_str(), it.first.c_str(), op->name.c_str(), it.first.c_str());
                }

                const Attribute& attr = it.second;
                for (size_t i = 0; i < attr.shape.size(); i++)
                {
                    fprintf(pyfp, "%d", attr.shape[i]);
                    if (i + 1 != attr.shape.size())
                        fprintf(pyfp, ",");
                }

                if (attr.type == 1 || attr.type == 2 || attr.type == 3)
                {
                    fprintf(pyfp, "), '%s')\n", type_to_numpy_string(attr.type));
                }
                else
                {
                    fprintf(pyfp, "), '%s', requires_grad=False)\n", type_to_numpy_string(attr.type));
                }
            }
        }

        for (const Operator* op : ops)
        {
            if (op->type != "pnnx.Attribute")
                continue;

            const std::string& key = op->attrs.begin()->first;
            const Attribute& attr = op->attrs.begin()->second;

            bool is_running_mean_var = false;
            {
                const Operand* r = op->outputs[0];
                if (r->consumers.size() == 1)
                {
                    const Operator* op2 = r->consumers[0];
                    if (op2->type == "F.batch_norm" || op2->type == "F.instance_norm")
                    {
                        if (r == op2->inputs[1] || r == op2->inputs[2])
                        {
                            is_running_mean_var = true;
                        }
                    }
                }
            }

            if (is_running_mean_var)
            {
                fprintf(pyfp, "        self.%s_%s = self.load_pnnx_bin_as_tensor(archive, '%s.%s', (", sanitize_identifier(op->name).c_str(), sanitize_identifier(key).c_str(), op->name.c_str(), key.c_str());
            }
            else
            {
                fprintf(pyfp, "        self.%s_%s = self.load_pnnx_bin_as_parameter(archive, '%s.%s', (", sanitize_identifier(op->name).c_str(), sanitize_identifier(key).c_str(), op->name.c_str(), key.c_str());
            }

            for (size_t i = 0; i < attr.shape.size(); i++)
            {
                fprintf(pyfp, "%d", attr.shape[i]);
                if (i + 1 != attr.shape.size())
                    fprintf(pyfp, ",");
            }

            if (attr.type == 1 || attr.type == 2 || attr.type == 3)
            {
                fprintf(pyfp, "), '%s')\n", type_to_numpy_string(attr.type));
            }
            else
            {
                fprintf(pyfp, "), '%s', requires_grad=False)\n", type_to_numpy_string(attr.type));
            }
        }

        fprintf(pyfp, "        archive.close()\n");
    }

    fprintf(pyfp, "\n");

    // utility function
    {
        fprintf(pyfp, "    def load_pnnx_bin_as_parameter(self, archive, key, shape, dtype, requires_grad=True):\n");
        fprintf(pyfp, "        return nn.Parameter(self.load_pnnx_bin_as_tensor(archive, key, shape, dtype), requires_grad)\n");
        fprintf(pyfp, "\n");
        fprintf(pyfp, "    def load_pnnx_bin_as_tensor(self, archive, key, shape, dtype):\n");
        fprintf(pyfp, "        _, tmppath = tempfile.mkstemp()\n");
        fprintf(pyfp, "        tmpf = open(tmppath, 'wb')\n");
        fprintf(pyfp, "        with archive.open(key) as keyfile:\n");
        fprintf(pyfp, "            tmpf.write(keyfile.read())\n");
        fprintf(pyfp, "        tmpf.close()\n");
        fprintf(pyfp, "        m = np.memmap(tmppath, dtype=dtype, mode='r', shape=shape).copy()\n");
        fprintf(pyfp, "        os.remove(tmppath)\n");
        fprintf(pyfp, "        return torch.from_numpy(m)\n");
    }

    fprintf(pyfp, "\n");

    // def forward
    {
        fprintf(pyfp, "    def forward(self");

        for (const Operator* op : ops)
        {
            if (op->type != "pnnx.Input")
                continue;

            fprintf(pyfp, ", v_%s", sanitize_identifier(op->outputs[0]->name).c_str());
        }

        fprintf(pyfp, "):\n");
    }

    // forward body
    {
        for (const Operator* op : ops)
        {
            if (op->type == "pnnx.Input" || op->type == "pnnx.Output")
                continue;

            fprintf(pyfp, "        ");

            if (op->type == "pnnx.Expression")
            {
                // expr
                for (size_t i = 0; i < op->outputs.size(); i++)
                {
                    fprintf(pyfp, "v_%s", sanitize_identifier(op->outputs[i]->name).c_str());
                    if (i + 1 != op->outputs.size())
                        fprintf(pyfp, ", ");
                }
                std::string expanded_expr = expand_expression(op);
                fprintf(pyfp, " = %s\n", expanded_expr.c_str());
            }
            else if (op->type == "pnnx.Attribute")
            {
                const std::string& key = op->attrs.begin()->first;
                fprintf(pyfp, "v_%s = self.%s_%s\n", sanitize_identifier(op->outputs[0]->name).c_str(), sanitize_identifier(op->name).c_str(), sanitize_identifier(key).c_str());
            }
            else if (op->type == "Tensor.slice")
            {
                // slice expr
                std::string slice_expr = make_slice_expression(op);
                fprintf(pyfp, "v_%s = v_%s[%s]\n", sanitize_identifier(op->outputs[0]->name).c_str(), sanitize_identifier(op->inputs[0]->name).c_str(), slice_expr.c_str());
            }
            else if (op->type == "Tensor.slice_copy")
            {
                // slice copy expr
                std::string slice_expr = make_slice_expression(op);
                fprintf(pyfp, "v_%s = v_%s\n", sanitize_identifier(op->outputs[0]->name).c_str(), sanitize_identifier(op->inputs[0]->name).c_str());
                fprintf(pyfp, "        v_%s[%s] = v_%s\n", sanitize_identifier(op->outputs[0]->name).c_str(), slice_expr.c_str(), sanitize_identifier(op->inputs[1]->name).c_str());
            }
            else if (op->type == "Tensor.index")
            {
                // index expr
                if (op->inputs.size() == 2)
                {
                    std::string expanded_expr = expand_expression(op->inputs[1]->producer);
                    fprintf(pyfp, "v_%s = v_%s[%s]\n", sanitize_identifier(op->outputs[0]->name).c_str(), sanitize_identifier(op->inputs[0]->name).c_str(), expanded_expr.c_str());
                }
                else
                {
                    std::string index_expr = make_index_expression(op);
                    fprintf(pyfp, "v_%s = v_%s[%s]\n", sanitize_identifier(op->outputs[0]->name).c_str(), sanitize_identifier(op->inputs[0]->name).c_str(), index_expr.c_str());
                }
            }
            else if (op->type == "Tensor.view" || op->type == "Tensor.reshape")
            {
                // view reshape
                fprintf(pyfp, "v_%s = v_%s.%s(", sanitize_identifier(op->outputs[0]->name).c_str(), sanitize_identifier(op->inputs[0]->name).c_str(), op->type.substr(7).c_str());
                if (op->inputs.size() == 2)
                {
                    fprintf(pyfp, "*v_%s", sanitize_identifier(op->inputs[1]->name).c_str());
                }
                else
                {
                    const std::vector<int>& shape = op->params.at("shape").ai;
                    for (size_t i = 0; i < shape.size(); i++)
                    {
                        fprintf(pyfp, "%d", shape[i]);
                        if (i + 1 != shape.size())
                            fprintf(pyfp, ", ");
                    }
                }
                fprintf(pyfp, ")\n");
            }
            else if (op->type == "Tensor.repeat")
            {
                // view reshape
                fprintf(pyfp, "v_%s = v_%s.%s(", sanitize_identifier(op->outputs[0]->name).c_str(), sanitize_identifier(op->inputs[0]->name).c_str(), op->type.substr(7).c_str());
                if (op->inputs.size() == 2)
                {
                    fprintf(pyfp, "*v_%s", sanitize_identifier(op->inputs[1]->name).c_str());
                }
                else
                {
                    const std::vector<int>& sizes = op->params.at("sizes").ai;
                    for (size_t i = 0; i < sizes.size(); i++)
                    {
                        fprintf(pyfp, "%d", sizes[i]);
                        if (i + 1 != sizes.size())
                            fprintf(pyfp, ", ");
                    }
                }
                fprintf(pyfp, ")\n");
            }
            else if (op->type == "torch.cat" || op->type == "torch.stack")
            {
                // cat
                fprintf(pyfp, "v_%s = %s(", sanitize_identifier(op->outputs[0]->name).c_str(), op->type.c_str());
                if (op->inputs.size() == 1)
                {
                    fprintf(pyfp, "v_%s", sanitize_identifier(op->inputs[0]->name).c_str());
                }
                else
                {
                    fprintf(pyfp, "(");
                    for (size_t i = 0; i < op->inputs.size(); i++)
                    {
                        fprintf(pyfp, "v_%s", sanitize_identifier(op->inputs[i]->name).c_str());
                        if (i + 1 != op->inputs.size())
                            fprintf(pyfp, ", ");
                    }
                    fprintf(pyfp, ")");
                }
                fprintf(pyfp, ", dim=%d", op->params.at("dim").i);
                fprintf(pyfp, ")\n");
            }
            else if (op->type == "torch.einsum")
            {
                // einsum
                fprintf(pyfp, "v_%s = %s(", sanitize_identifier(op->outputs[0]->name).c_str(), op->type.c_str());

                fprintf(pyfp, "\'%s\'", op->params.at("equation").s.c_str());

                for (size_t i = 0; i < op->inputs.size(); i++)
                {
                    fprintf(pyfp, ", v_%s", sanitize_identifier(op->inputs[i]->name).c_str());
                }
                fprintf(pyfp, ")\n");
            }
            else if (op->type == "prim::TupleUnpack")
            {
                for (size_t i = 0; i < op->outputs.size(); i++)
                {
                    fprintf(pyfp, "v_%s", sanitize_identifier(op->outputs[i]->name).c_str());
                    if (i + 1 != op->outputs.size())
                        fprintf(pyfp, ", ");
                }
                fprintf(pyfp, " = v_%s\n", sanitize_identifier(op->inputs[0]->name).c_str());
            }
            else if (op->type == "prim::TupleConstruct")
            {
                fprintf(pyfp, "v_%s", sanitize_identifier(op->outputs[0]->name).c_str());
                fprintf(pyfp, " = (");
                for (size_t i = 0; i < op->inputs.size(); i++)
                {
                    fprintf(pyfp, "v_%s, ", sanitize_identifier(op->inputs[i]->name).c_str());
                }
                fprintf(pyfp, ")\n");
            }
            else if (op->type == "prim::ListUnpack")
            {
                for (size_t i = 0; i < op->outputs.size(); i++)
                {
                    fprintf(pyfp, "v_%s", sanitize_identifier(op->outputs[i]->name).c_str());
                    if (i + 1 != op->outputs.size())
                        fprintf(pyfp, ", ");
                }
                fprintf(pyfp, " = v_%s\n", sanitize_identifier(op->inputs[0]->name).c_str());
            }
            else if (op->type == "prim::ListConstruct")
            {
                fprintf(pyfp, "v_%s", sanitize_identifier(op->outputs[0]->name).c_str());
                fprintf(pyfp, " = [");
                for (size_t i = 0; i < op->inputs.size(); i++)
                {
                    fprintf(pyfp, "v_%s", sanitize_identifier(op->inputs[i]->name).c_str());
                    if (i + 1 != op->inputs.size())
                        fprintf(pyfp, ", ");
                }
                fprintf(pyfp, "]\n");
            }
            else if (op->type == "nn.LSTM")
            {
                if (op->outputs.size() == 1)
                {
                    fprintf(pyfp, "v_%s, _", sanitize_identifier(op->outputs[0]->name).c_str());
                }
                else
                {
                    fprintf(pyfp, "v_%s, (v_%s, v_%s)", sanitize_identifier(op->outputs[0]->name).c_str(), sanitize_identifier(op->outputs[1]->name).c_str(), sanitize_identifier(op->outputs[2]->name).c_str());
                }
                fprintf(pyfp, " = self.%s(", sanitize_identifier(op->name).c_str());
                fprintf(pyfp, "v_%s", sanitize_identifier(op->inputs[0]->name).c_str());
                if (op->inputs.size() == 3)
                {
                    fprintf(pyfp, ", (v_%s, v_%s)", sanitize_identifier(op->inputs[1]->name).c_str(), sanitize_identifier(op->inputs[2]->name).c_str());
                }
                fprintf(pyfp, ")\n");
            }
            else if (op->type == "nn.MultiheadAttention")
            {
                if (op->outputs.size() == 1)
                {
                    fprintf(pyfp, "v_%s, _", sanitize_identifier(op->outputs[0]->name).c_str());
                }
                else
                {
                    for (size_t i = 0; i < op->outputs.size(); i++)
                    {
                        fprintf(pyfp, "v_%s", sanitize_identifier(op->outputs[i]->name).c_str());
                        if (i + 1 != op->outputs.size())
                            fprintf(pyfp, ", ");
                    }
                }
                fprintf(pyfp, " = self.%s(", sanitize_identifier(op->name).c_str());
                if (op->inputs.size() == 1)
                {
                    std::string in0 = sanitize_identifier(op->inputs[0]->name);
                    fprintf(pyfp, "v_%s, v_%s, v_%s", in0.c_str(), in0.c_str(), in0.c_str());
                }
                else if (op->inputs.size() == 2)
                {
                    std::string in0 = sanitize_identifier(op->inputs[0]->name);
                    std::string in1 = sanitize_identifier(op->inputs[1]->name);
                    fprintf(pyfp, "v_%s, v_%s, v_%s", in0.c_str(), in1.c_str(), in1.c_str());
                }
                else
                {
                    for (size_t i = 0; i < op->inputs.size(); i++)
                    {
                        fprintf(pyfp, "v_%s", sanitize_identifier(op->inputs[i]->name).c_str());
                        if (i + 1 != op->inputs.size())
                            fprintf(pyfp, ", ");
                    }
                }
                fprintf(pyfp, ")\n");
            }
            else if (op->type.substr(0, 3) == "nn." || op->type.substr(0, 16) == "torchvision.ops.")
            {
                // self.xxx()
                for (size_t i = 0; i < op->outputs.size(); i++)
                {
                    fprintf(pyfp, "v_%s", sanitize_identifier(op->outputs[i]->name).c_str());
                    if (i + 1 != op->outputs.size())
                        fprintf(pyfp, ", ");
                }
                fprintf(pyfp, " = self.%s(", sanitize_identifier(op->name).c_str());
                for (size_t i = 0; i < op->inputs.size(); i++)
                {
                    fprintf(pyfp, "v_%s", sanitize_identifier(op->inputs[i]->name).c_str());
                    if (i + 1 != op->inputs.size())
                        fprintf(pyfp, ", ");
                }
                fprintf(pyfp, ")\n");
            }
            else if (op->type.find("::") != std::string::npos || op->type.find(".") != std::string::npos)
            {
                // direct
                for (size_t i = 0; i < op->outputs.size(); i++)
                {
                    fprintf(pyfp, "v_%s", sanitize_identifier(op->outputs[i]->name).c_str());
                    if (i + 1 != op->outputs.size())
                        fprintf(pyfp, ", ");
                }

                if (op->type.substr(0, 7) == "Tensor.")
                {
                    fprintf(pyfp, " = v_%s.%s(", sanitize_identifier(op->inputs[0]->name).c_str(), op->type.substr(7).c_str());

                    for (size_t i = 1; i < op->inputs.size(); i++)
                    {
                        fprintf(pyfp, "v_%s, ", sanitize_identifier(op->inputs[i]->name).c_str());
                    }
                }
                else
                {
                    fprintf(pyfp, " = %s(", op->type.c_str());

                    if (op->inputnames.size() == op->inputs.size())
                    {
                        for (size_t i = 0; i < op->inputs.size(); i++)
                        {
                            if (!op->inputnames[i].empty())
                                continue;

                            fprintf(pyfp, "v_%s", sanitize_identifier(op->inputs[i]->name).c_str());
                            if (i + 1 != op->inputs.size())
                                fprintf(pyfp, ", ");
                        }

                        for (size_t i = 0; i < op->inputs.size(); i++)
                        {
                            if (op->inputnames[i].empty())
                                continue;

                            fprintf(pyfp, "%s=v_%s", op->inputnames[i].c_str(), sanitize_identifier(op->inputs[i]->name).c_str());
                            if (i + 1 != op->inputs.size())
                                fprintf(pyfp, ", ");
                        }
                    }
                    else
                    {
                        for (size_t i = 0; i < op->inputs.size(); i++)
                        {
                            fprintf(pyfp, "v_%s", sanitize_identifier(op->inputs[i]->name).c_str());
                            if (i + 1 != op->inputs.size())
                                fprintf(pyfp, ", ");
                        }
                    }
                }

                int i = 0;
                for (const auto& it : op->params)
                {
                    if (op->type.substr(0, 7) == "Tensor." && i == 0)
                    {
                        fprintf(pyfp, "%s=", it.first.c_str());
                    }
                    else if (op->inputs.empty() && i == 0)
                    {
                        fprintf(pyfp, "%s=", it.first.c_str());
                    }
                    else
                    {
                        fprintf(pyfp, ", %s=", it.first.c_str());
                    }

                    i++;

                    const Parameter& param = it.second;
                    if (param.type == 0)
                    {
                        fprintf(pyfp, "None");
                    }
                    if (param.type == 1)
                    {
                        if (param.b)
                            fprintf(pyfp, "True");
                        else
                            fprintf(pyfp, "False");
                    }
                    if (param.type == 2)
                    {
                        fprintf(pyfp, "%d", param.i);
                    }
                    if (param.type == 3)
                    {
                        fprintf(pyfp, "%f", param.f);
                    }
                    if (param.type == 4)
                    {
                        if (param.s.substr(0, 6) == "torch.")
                        {
                            fprintf(pyfp, "%s", param.s.c_str());
                        }
                        else
                        {
                            fprintf(pyfp, "\'%s\'", param.s.c_str());
                        }
                    }
                    if (param.type == 5)
                    {
                        fprintf(pyfp, "(");
                        for (size_t i = 0; i < param.ai.size(); i++)
                        {
                            fprintf(pyfp, "%d", param.ai[i]);
                            if (i + 1 != param.ai.size() || param.ai.size() == 1)
                                fprintf(pyfp, ",");
                        }
                        fprintf(pyfp, ")");
                    }
                    if (param.type == 6)
                    {
                        fprintf(pyfp, "(");
                        for (size_t i = 0; i < param.af.size(); i++)
                        {
                            fprintf(pyfp, "%f", param.af[i]);
                            if (i + 1 != param.af.size() || param.af.size() == 1)
                                fprintf(pyfp, ",");
                        }
                        fprintf(pyfp, ")");
                    }
                    if (param.type == 7)
                    {
                        fprintf(pyfp, "(");
                        for (size_t i = 0; i < param.as.size(); i++)
                        {
                            if (param.as[i].substr(0, 6) == "torch.")
                            {
                                fprintf(pyfp, "%s", param.as[i].c_str());
                            }
                            else
                            {
                                fprintf(pyfp, "\'%s\'", param.as[i].c_str());
                            }
                            if (i + 1 != param.as.size() || param.as.size() == 1)
                                fprintf(pyfp, ",");
                        }
                        fprintf(pyfp, ")");
                    }
                }

                fprintf(pyfp, ")\n");
            }
            else
            {
                fprintf(stderr, "todo %s\n", op->type.c_str());
            }
        }
    }

    // return
    {
        fprintf(pyfp, "        return ");

        int output_count = 0;
        {
            for (const Operator* op : ops)
            {
                if (op->type == "pnnx.Output")
                    output_count++;
            }
        }

        int output_index = 0;
        for (const Operator* op : ops)
        {
            if (op->type != "pnnx.Output")
                continue;

            fprintf(pyfp, "v_%s", sanitize_identifier(op->inputs[0]->name).c_str());
            if (output_index + 1 != output_count)
                fprintf(pyfp, ", ");

            output_index++;
        }

        fprintf(pyfp, "\n");
    }

    fprintf(pyfp, "\n");

    // export torchscript
    {
        fprintf(pyfp, "def export_torchscript():\n");
        fprintf(pyfp, "    net = Model()\n");
        fprintf(pyfp, "    net.eval()\n");
        fprintf(pyfp, "\n");
        fprintf(pyfp, "    torch.manual_seed(0)\n");

        std::vector<std::string> input_names;
        for (const Operator* op : ops)
        {
            if (op->type != "pnnx.Input")
                continue;

            const Operand* r = op->outputs[0];
            std::string input_name = std::string("v_") + sanitize_identifier(r->name);
            if (type_is_integer(r->type))
            {
                fprintf(pyfp, "    %s = torch.randint(10, (", input_name.c_str());
                for (size_t i = 0; i < r->shape.size(); i++)
                {
                    fprintf(pyfp, "%d", r->shape[i]);
                    if (i + 1 != r->shape.size() || r->shape.size() == 1)
                        fprintf(pyfp, ", ");
                }
                fprintf(pyfp, "), dtype=%s)\n", type_to_dtype_string(r->type));
            }
            else
            {
                fprintf(pyfp, "    %s = torch.rand(", input_name.c_str());
                for (size_t i = 0; i < r->shape.size(); i++)
                {
                    fprintf(pyfp, "%d, ", r->shape[i]);
                }
                fprintf(pyfp, "dtype=%s)\n", type_to_dtype_string(r->type));
            }

            input_names.push_back(input_name);
        }

        fprintf(pyfp, "\n");

        if (input_names.size() == 1)
        {
            fprintf(pyfp, "    mod = torch.jit.trace(net, %s)\n", input_names[0].c_str());
        }
        else
        {
            fprintf(pyfp, "    mod = torch.jit.trace(net, (");

            for (size_t i = 0; i < input_names.size(); i++)
            {
                fprintf(pyfp, "%s", input_names[i].c_str());
                if (i + 1 != input_names.size())
                    fprintf(pyfp, ", ");
            }

            fprintf(pyfp, "))\n");
        }

        fprintf(pyfp, "    mod.save(\"%s.pt\")\n", pypath.c_str());
    }

    fprintf(pyfp, "\n");

    // export onnx
    {
        fprintf(pyfp, "def export_onnx():\n");
        fprintf(pyfp, "    net = Model()\n");
        fprintf(pyfp, "    net.eval()\n");
        fprintf(pyfp, "\n");
        fprintf(pyfp, "    torch.manual_seed(0)\n");

        std::vector<std::string> input_names;
        for (const Operator* op : ops)
        {
            if (op->type != "pnnx.Input")
                continue;

            const Operand* r = op->outputs[0];
            std::string input_name = std::string("v_") + sanitize_identifier(r->name);
            if (type_is_integer(r->type))
            {
                fprintf(pyfp, "    %s = torch.randint(10, (", input_name.c_str());
                for (size_t i = 0; i < r->shape.size(); i++)
                {
                    fprintf(pyfp, "%d", r->shape[i]);
                    if (i + 1 != r->shape.size() || r->shape.size() == 1)
                        fprintf(pyfp, ", ");
                }
                fprintf(pyfp, "), dtype=%s)\n", type_to_dtype_string(r->type));
            }
            else
            {
                fprintf(pyfp, "    %s = torch.rand(", input_name.c_str());
                for (size_t i = 0; i < r->shape.size(); i++)
                {
                    fprintf(pyfp, "%d, ", r->shape[i]);
                }
                fprintf(pyfp, "dtype=%s)\n", type_to_dtype_string(r->type));
            }

            input_names.push_back(input_name);
        }

        fprintf(pyfp, "\n");

        // torch.onnx._export(net, v_0, "test_swin_t.onnx", export_params=True, opset_version=14, input_names=['in0'], output_names=['out0'])

        if (input_names.size() == 1)
        {
            fprintf(pyfp, "    torch.onnx._export(net, %s", input_names[0].c_str());
        }
        else
        {
            fprintf(pyfp, "    torch.onnx._export(net, (");

            for (size_t i = 0; i < input_names.size(); i++)
            {
                fprintf(pyfp, "%s", input_names[i].c_str());
                if (i + 1 != input_names.size())
                    fprintf(pyfp, ", ");
            }

            fprintf(pyfp, ")");
        }

        fprintf(pyfp, ", \"%s.onnx\", export_params=True, operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK, opset_version=13", pypath.c_str());

        fprintf(pyfp, ", input_names=[");
        {
            int input_count = 0;
            {
                for (const Operator* op : ops)
                {
                    if (op->type == "pnnx.Input")
                        input_count++;
                }
            }

            int input_index = 0;
            for (const Operator* op : ops)
            {
                if (op->type != "pnnx.Input")
                    continue;

                fprintf(pyfp, "'in%d'", input_index);
                if (input_index + 1 != input_count)
                    fprintf(pyfp, ", ");

                input_index++;
            }
        }
        fprintf(pyfp, "]");

        fprintf(pyfp, ", output_names=[");
        {
            int output_count = 0;
            {
                for (const Operator* op : ops)
                {
                    if (op->type == "pnnx.Output")
                        output_count++;
                }
            }

            int output_index = 0;
            for (const Operator* op : ops)
            {
                if (op->type != "pnnx.Output")
                    continue;

                fprintf(pyfp, "'out%d'", output_index);
                if (output_index + 1 != output_count)
                    fprintf(pyfp, ", ");

                output_index++;
            }
        }
        fprintf(pyfp, "]");

        fprintf(pyfp, ")\n");
    }

    fprintf(pyfp, "\n");

    // test inference
    {
        fprintf(pyfp, "def test_inference():\n");
        fprintf(pyfp, "    net = Model()\n");
        fprintf(pyfp, "    net.eval()\n");
        fprintf(pyfp, "\n");
        fprintf(pyfp, "    torch.manual_seed(0)\n");

        std::vector<std::string> input_names;
        for (const Operator* op : ops)
        {
            if (op->type != "pnnx.Input")
                continue;

            const Operand* r = op->outputs[0];
            std::string input_name = std::string("v_") + sanitize_identifier(r->name);
            if (type_is_integer(r->type))
            {
                fprintf(pyfp, "    %s = torch.randint(10, (", input_name.c_str());
                for (size_t i = 0; i < r->shape.size(); i++)
                {
                    fprintf(pyfp, "%d", r->shape[i]);
                    if (i + 1 != r->shape.size() || r->shape.size() == 1)
                        fprintf(pyfp, ", ");
                }
                fprintf(pyfp, "), dtype=%s)\n", type_to_dtype_string(r->type));
            }
            else
            {
                fprintf(pyfp, "    %s = torch.rand(", input_name.c_str());
                for (size_t i = 0; i < r->shape.size(); i++)
                {
                    fprintf(pyfp, "%d, ", r->shape[i]);
                }
                fprintf(pyfp, "dtype=%s)\n", type_to_dtype_string(r->type));
            }

            input_names.push_back(input_name);
        }

        fprintf(pyfp, "\n");

        if (input_names.size() == 1)
        {
            fprintf(pyfp, "    return net(%s)\n", input_names[0].c_str());
        }
        else
        {
            fprintf(pyfp, "    return net(");

            for (size_t i = 0; i < input_names.size(); i++)
            {
                fprintf(pyfp, "%s", input_names[i].c_str());
                if (i + 1 != input_names.size())
                    fprintf(pyfp, ", ");
            }

            fprintf(pyfp, ")\n");
        }
    }

    fclose(pyfp);

    return 0;
}

int Graph::parse(const std::string& param)
{
    std::istringstream is(param);
    if (!is.good())
    {
        fprintf(stderr, "open failed\n");
        return -1;
    }

    int magic = 0;
    {
        std::string line;
        std::getline(is, line);
        std::istringstream iss(line);

        iss >> magic;
    }

    int operator_count = 0;
    int operand_count = 0;
    {
        std::string line;
        std::getline(is, line);
        std::istringstream iss(line);

        iss >> operator_count >> operand_count;
    }

    for (int i = 0; i < operator_count; i++)
    {
        std::string line;
        std::getline(is, line);
        std::istringstream iss(line);

        std::string type;
        std::string name;
        int input_count = 0;
        int output_count = 0;

        iss >> type >> name >> input_count >> output_count;

        Operator* op = new_operator(type, name);

        for (int j = 0; j < input_count; j++)
        {
            std::string operand_name;
            iss >> operand_name;

            Operand* r = get_operand(operand_name);
            r->consumers.push_back(op);
            op->inputs.push_back(r);
        }

        for (int j = 0; j < output_count; j++)
        {
            std::string operand_name;
            iss >> operand_name;

            Operand* r = new_operand(operand_name);
            r->producer = op;
            op->outputs.push_back(r);
        }

        // key=value
        while (!iss.eof())
        {
            std::string param;
            iss >> param;

            std::string key;
            std::string value;
            std::istringstream pss(param);
            std::getline(pss, key, '=');
            std::getline(pss, value);

            if (key[0] == '@')
            {
                // attribute
                //                 load_attribute(op, key.substr(1), value, szr);
            }
            else if (key[0] == '$')
            {
                // operand input key
                //                 load_input_key(op, key.substr(1), value);
            }
            else if (key[0] == '#')
            {
                // operand shape
                load_shape(op, key.substr(1), value);
            }
            else
            {
                // parameter
                load_parameter(op, key, value);
            }
        }
    }

    return 0;
}

void Operand::remove_consumer(const Operator* c)
{
    auto it = std::find(consumers.begin(), consumers.end(), c);
    consumers.erase(it);
}

Operator* Graph::new_operator(const std::string& type, const std::string& name)
{
    Operator* op = new Operator;
    op->type = type;
    op->name = name;
    ops.push_back(op);
    return op;
}

Operator* Graph::new_operator_before(const std::string& type, const std::string& name, const Operator* cur)
{
    Operator* op = new Operator;
    op->type = type;
    op->name = name;
    ops.insert(std::find(ops.begin(), ops.end(), cur), op);
    return op;
}

Operator* Graph::new_operator_after(const std::string& type, const std::string& name, const Operator* cur)
{
    Operator* op = new Operator;
    op->type = type;
    op->name = name;
    ops.insert(std::find(ops.begin(), ops.end(), cur) + 1, op);
    return op;
}

#if BUILD_PNNX
Operand* Graph::new_operand(const torch::jit::Value* v)
{
    Operand* r = new Operand;
    r->name = v->debugName();

    auto pt = v->type()->cast<c10::TensorType>();
    if (pt)
    {
        if (pt->scalarType().has_value() && pt->dim().has_value())
        {
            r->type = get_at_tensor_type(pt->scalarType().value());
            const int ndim = (int)pt->dim().value();
            r->shape.resize(ndim);
            for (int i = 0; i < ndim; i++)
            {
                if (pt->sizes()[i].has_value())
                    r->shape[i] = (int)pt->sizes()[i].value();
                else
                    r->shape[i] = -1;
            }
        }
    }

    operands.push_back(r);
    return r;
}
#endif // BUILD_PNNX

Operand* Graph::new_operand(const std::string& name)
{
    Operand* r = new Operand;
    r->name = name;
    operands.push_back(r);
    return r;
}

Operand* Graph::get_operand(const std::string& name)
{
    for (Operand* r : operands)
    {
        if (r->name == name)
            return r;
    }

    return 0;
}

const Operand* Graph::get_operand(const std::string& name) const
{
    for (const Operand* r : operands)
    {
        if (r->name == name)
            return r;
    }

    return 0;
}

} // namespace pnnx