loadcaffe.cpp

//
//  main.cpp
//  loadcaffe
//
//  Created by Sergey Zagoruyko on 28/11/14.
//  Copyright (c) 2014 Sergey Zagoruyko. All rights reserved.
//

#include <iostream>
#include <fstream>
#include <fcntl.h>
#include <unistd.h>
#include <string.h>
#include <TH/TH.h>
#include <locale>

#include <google/protobuf/io/coded_stream.h>
#include <google/protobuf/io/zero_copy_stream_impl.h>
#include <google/protobuf/text_format.h>

#include "build/caffe.pb.h"

using google::protobuf::io::FileInputStream;
using google::protobuf::io::FileOutputStream;
using google::protobuf::io::ZeroCopyInputStream;
using google::protobuf::io::CodedInputStream;
using google::protobuf::Message;


extern "C" {
void loadBinary(void** handle, const char* prototxt_name, const char* binary_name);
void convertProtoToLua(void** handle, const char* lua_name, const char* cuda_package);
void convertProtoToLuaV1(const caffe::NetParameter &netparam, const char* lua_name, const char* cuda_package);
void convertProtoToLuaV2(const caffe::NetParameter &netparam, const char* lua_name, const char* cuda_package);
void loadModule(const void** handle, const char* name, THFloatTensor* weight, THFloatTensor* bias);
void loadModuleV2(const caffe::NetParameter* netparam, const char* name, THFloatTensor* weight, THFloatTensor* bias);
void loadModuleV1(const caffe::NetParameter* netparam, const char* name, THFloatTensor* weight, THFloatTensor* bias);
void destroyBinary(void** handle);
}


bool ReadProtoFromTextFile(const char* filename, Message* proto) {
    int fd = open(filename, O_RDONLY);
    if(fd < 0)
      return false;
    
    FileInputStream* input = new FileInputStream(fd);
    bool success = google::protobuf::TextFormat::Parse(input, proto);
    delete input;
    close(fd);
    return success;
}


bool ReadProtoFromBinaryFile(const char* filename, Message* proto) {
    int fd = open(filename, O_RDONLY);
    if(fd < 0)
      return false;
    
    ZeroCopyInputStream* raw_input = new FileInputStream(fd);
    CodedInputStream* coded_input = new CodedInputStream(raw_input);
    coded_input->SetTotalBytesLimit(1073741824, 536870912);
    
    bool success = proto->ParseFromCodedStream(coded_input);
    
    delete coded_input;
    delete raw_input;
    close(fd);
    return success;
}


enum PACKAGE_TYPE {
  CCN2, NN, CUDNN
};


void convertProtoToLua(void** handle, const char* lua_name, const char* cuda_package)
{
  std::locale::global(std::locale());
  const caffe::NetParameter netparam = *(const caffe::NetParameter*)handle[1];
  if (netparam.layers_size() > 0)
      convertProtoToLuaV1(netparam, lua_name, cuda_package);
  else
      convertProtoToLuaV2(netparam, lua_name, cuda_package);
}


void convertProtoToLuaV1(const caffe::NetParameter &netparam, const char* lua_name, const char* cuda_package)
{
  PACKAGE_TYPE cuda_package_type = CCN2;
  if(std::string(cuda_package) == "ccn2")
    cuda_package_type = CCN2;
  else if(std::string(cuda_package) == "nn")
    cuda_package_type = NN;
  else if(std::string(cuda_package) == "cudnn")
    cuda_package_type = CUDNN;

  std::ofstream ofs (lua_name);

  ofs << "require '" << cuda_package << "'\n";
  ofs << "local model = {}\n";
  if(std::string(cuda_package)=="ccn2")
    ofs<< "table.insert(model, {'torch_transpose_dwhb', nn.Transpose({1,4},{1,3},{1,2})})\n";

  int num_output = netparam.input_dim_size();
  for (int i=0; i<netparam.layers_size(); ++i)
  {
    std::vector<std::pair<std::string, std::string>> lines;
    auto& layer = netparam.layers(i);
    switch(layer.type())
    {
      case caffe::V1LayerParameter::CONVOLUTION:
      {
        auto &param = layer.convolution_param();
        int groups = param.group() == 0 ? 1 : param.group();
        int nInputPlane = layer.blobs(0).channels()*groups;
        int nOutputPlane = layer.blobs(0).num();
        //int nOutputPlane = param.num_output();
        num_output = nOutputPlane;
        int kW = param.kernel_w();
        int kH = param.kernel_h();
        int dW = param.stride_w();
        int dH = param.stride_h();
        if(kW==0 || kH==0)
        {
          kW = param.kernel_size();
          kH = kW;
        }
        if(dW==0 || dH==0)
        {
          dW = param.stride();
          dH = dW;
        }
        int pad_w = param.pad_w();
        int pad_h = param.pad_h();
        if(pad_w==0 || pad_h==0)
        {
          pad_w = param.pad();
          pad_h = pad_w;
        }
        if(cuda_package_type == CCN2)
        {
          if(kW != kH || dW != dH || pad_w != pad_h)
          {
            std::cout << "ccn2 only supports square images!\n";
            break;
          }
          char buf[1024];
          sprintf(buf, "ccn2.SpatialConvolution(%d, %d, %d, %d, %d, %d)",
              nInputPlane, nOutputPlane, kW, dW, pad_w, groups);
          lines.emplace_back(layer.name(), buf);
        }
        else if(cuda_package_type == NN)
        {
          if(groups != 1)
          {
            std::cout << "nn supports no groups!\n";
            break;
          }
          char buf[1024];
          sprintf(buf, "nn.SpatialConvolutionMM(%d, %d, %d, %d, %d, %d, %d, %d)",
              nInputPlane, nOutputPlane, kW, kH, dW, dH, pad_w, pad_h);
          lines.emplace_back(layer.name(), buf);
        }
        else
        {
          char buf[1024];
          sprintf(buf, "cudnn.SpatialConvolution(%d, %d, %d, %d, %d, %d, %d, %d, %d)",
              nInputPlane, nOutputPlane, kW, kH, dW, dH, pad_w, pad_h, groups);
          lines.emplace_back(layer.name(), buf);
        }
        break;
      }
      case caffe::V1LayerParameter::POOLING:
      {
        auto &param = layer.pooling_param();
        int kW = param.kernel_w();
        int kH = param.kernel_h();
        int dW = param.stride_w();
        int dH = param.stride_h();
        int padW = param.pad_w();
        int padH = param.pad_h();
        if(kW==0 || kH==0)
        {
          kW = param.kernel_size();
          kH = kW;
        }
        if(dW==0 || dH==0)
        {
          dW = param.stride();
          dH = dW;
        }
        if(padW==0 && padH==0)
        {
          padW = param.pad();
          padH = padW;
        }

        char buf[1024];
        switch(cuda_package_type)
        {
          case CCN2: // ceil mode by default
            if(param.pool() == caffe::PoolingParameter::MAX)
              sprintf(buf, "ccn2.SpatialMaxPooling(%d, %d)", kW, dW);
            else if(param.pool() == caffe::PoolingParameter::AVE)
              sprintf(buf, "ccn2.SpatialAvgPooling(%d, %d)", kW, dW);
            else if(param.pool() == caffe::PoolingParameter::STOCHASTIC)
              THError("Stochastic pooling is not implemented in DHWB format");
            break;
          case CUDNN:
            if(param.pool() == caffe::PoolingParameter::MAX)
              sprintf(buf, "cudnn.SpatialMaxPooling(%d, %d, %d, %d, %d, %d):ceil()", kW, kH, dW, dH, padW, padH);
            else if(param.pool() == caffe::PoolingParameter::AVE)
              sprintf(buf, "cudnn.SpatialAveragePooling(%d, %d, %d, %d, %d, %d):ceil()", kW, kH, dW, dH, padW, padH);
            else if(param.pool() == caffe::PoolingParameter::STOCHASTIC)
              sprintf(buf, "inn.SpatialStochasticPooling(%d, %d, %d, %d)", kW, kH, dW, dH);
            break;
          case NN:
            if(param.pool() == caffe::PoolingParameter::MAX)
              sprintf(buf, "nn.SpatialMaxPooling(%d, %d, %d, %d, %d, %d):ceil()", kW, kH, dW, dH, padW, padH);
            else if(param.pool() == caffe::PoolingParameter::AVE)
              sprintf(buf, "inn.SpatialAveragePooling(%d, %d, %d, %d)", kW, kH, dW, dH); // padding is not supported yet
            else if(param.pool() == caffe::PoolingParameter::STOCHASTIC)
              sprintf(buf, "inn.SpatialStochasticPooling(%d, %d, %d, %d)", kW, kH, dW, dH);
            break;
        }
        lines.emplace_back(layer.name(), buf);
        break;
      }
      case caffe::V1LayerParameter::RELU:
      {
        switch(cuda_package_type)
        {
          case CUDNN:
            lines.emplace_back(layer.name(), "cudnn.ReLU(true)");
            break;
          default:
            lines.emplace_back(layer.name(), "nn.ReLU(true)");
            break;
        }
        break;
      }
      case caffe::V1LayerParameter::TANH:
      {
        switch(cuda_package_type)
        {
          case CUDNN:
            lines.emplace_back(layer.name(), "cudnn.Tanh(true)");
            break;
          default:
            lines.emplace_back(layer.name(), "nn.Tanh()");
            break;
        }
        break;
      }
      case caffe::V1LayerParameter::SIGMOID:
      {
        switch(cuda_package_type)
        {
          case CUDNN:
            lines.emplace_back(layer.name(), "cudnn.Sigmoid(true)");
            break;
          default:
            lines.emplace_back(layer.name(), "nn.Sigmoid()");
            break;
        }
        break;
      }
      case caffe::V1LayerParameter::LRN:
      {
        auto &param = layer.lrn_param();
        int local_size = param.local_size();
        float alpha = param.alpha();
        float beta = param.beta();
        float k = param.k();
        char buf[1024];
        if(std::string(cuda_package) == "ccn2")
          sprintf(buf, "ccn2.SpatialCrossResponseNormalization(%d, %.6f, %.4f, %f)", local_size, alpha, beta, k);
        else
          sprintf(buf, "inn.SpatialCrossResponseNormalization(%d, %.6f, %.4f, %f)", local_size, alpha, beta, k);
        lines.emplace_back(layer.name(), buf);
        break;
      }
      case caffe::V1LayerParameter::INNER_PRODUCT:
      {
        auto &param = layer.inner_product_param();
        int nInputPlane = layer.blobs(0).width();
        int nOutputPlane = param.num_output();
        char buf[1024];
        sprintf(buf, "nn.Linear(%d, %d)", nInputPlane, nOutputPlane);
        if(num_output != nInputPlane)
        {
          if(std::string(cuda_package) == "ccn2")
            lines.emplace_back("torch_transpose_bdwh", "nn.Transpose({4,1},{4,2},{4,3})");
          lines.emplace_back("torch_view", "nn.View(-1):setNumInputDims(3)");
        }
        lines.emplace_back(layer.name(), buf);
        num_output = nOutputPlane;
        break;
      }
      case caffe::V1LayerParameter::DROPOUT:
      {
        char buf[1024];
        sprintf(buf, "nn.Dropout(%f)", layer.dropout_param().dropout_ratio());
        lines.emplace_back(layer.name(), buf);
        break;
      }
      case caffe::V1LayerParameter::SOFTMAX_LOSS:
      {
        lines.emplace_back(layer.name(), "nn.SoftMax()");
        break;
      }
      case caffe::V1LayerParameter::SOFTMAX:
      {
        lines.emplace_back(layer.name(), "nn.SoftMax()");
        break;
      }
      default:
      {
        std::cout << "MODULE " << layer.name() << " UNDEFINED\n";
        break;
      }
    }
    if(!lines.empty())
      for(auto& it: lines)
        ofs << "table.insert(model, {'" << it.first << "', " << it.second << "})\n";
    else
    {
      ofs << "-- warning: module '" << layer.name() << " [type " << layer.type() << "]" << "' not found\n";
      std::cout << "warning: module '" << layer.name() << " [type " << layer.type() << "]" << "' not found\n";
    }
  }
  ofs << "return model";
}


void convertProtoToLuaV2(const caffe::NetParameter &netparam, const char* lua_name, const char* cuda_package)
{
  PACKAGE_TYPE cuda_package_type = CCN2;
  if(std::string(cuda_package) == "ccn2")
    cuda_package_type = CCN2;
  else if(std::string(cuda_package) == "nn")
    cuda_package_type = NN;
  else if(std::string(cuda_package) == "cudnn")
    cuda_package_type = CUDNN;

  std::ofstream ofs (lua_name);

  ofs << "require '" << cuda_package << "'\n";
  ofs << "local model = {}\n";
  if(std::string(cuda_package)=="ccn2")
    ofs<< "table.insert(model, {'torch_transpose_dwhb', nn.Transpose({1,4},{1,3},{1,2})})\n";

  int num_output = netparam.input_shape_size() * 4;

  for (int i=0; i<netparam.layer_size(); ++i)
  {
    std::vector<std::pair<std::string, std::string>> lines;
    auto& layer = netparam.layer(i);
    if(layer.type() == "Convolution")
    {
      auto &param = layer.convolution_param();
      int groups = param.group() == 0 ? 1 : param.group();
      int nInputPlane = layer.blobs(0).shape().dim(1)*groups;
      int nOutputPlane = layer.blobs(0).shape().dim(0);
      //int nOutputPlane = param.num_output();
      num_output = nOutputPlane;
      int kW = param.kernel_w();
      int kH = param.kernel_h();
      int dW = param.stride_w();
      int dH = param.stride_h();
      if(kW==0 || kH==0)
      {
        kW = param.kernel_size();
        kH = kW;
      }
      if(dW==0 || dH==0)
      {
        dW = param.stride();
        dH = dW;
      }
      int pad_w = param.pad_w();
      int pad_h = param.pad_h();
      if(pad_w==0 || pad_h==0)
      {
        pad_w = param.pad();
        pad_h = pad_w;
      }
      if(cuda_package_type == CCN2)
      {
        if(kW != kH || dW != dH || pad_w != pad_h)
        {
          std::cout << "ccn2 only supports square images!\n";
          break;
        }
        char buf[1024];
        sprintf(buf, "ccn2.SpatialConvolution(%d, %d, %d, %d, %d, %d)",
            nInputPlane, nOutputPlane, kW, dW, pad_w, groups);
        lines.emplace_back(layer.name(), buf);
      }
      else if(cuda_package_type == NN)
      {
        if(groups != 1)
        {
          std::cout << "nn supports no groups!\n";
          break;
        }
        char buf[1024];
        sprintf(buf, "nn.SpatialConvolutionMM(%d, %d, %d, %d, %d, %d, %d, %d)",
            nInputPlane, nOutputPlane, kW, kH, dW, dH, pad_w, pad_h);
        lines.emplace_back(layer.name(), buf);
      }
      else
      {
        char buf[1024];
        sprintf(buf, "cudnn.SpatialConvolution(%d, %d, %d, %d, %d, %d, %d, %d, %d)",
            nInputPlane, nOutputPlane, kW, kH, dW, dH, pad_w, pad_h, groups);
        lines.emplace_back(layer.name(), buf);
      }
    }
    if(layer.type() == "Pooling")
    {
      auto &param = layer.pooling_param();
      std::string ptype = param.pool() == caffe::PoolingParameter::MAX ? "Max" : "Avg";
      int kW = param.kernel_w();
      int kH = param.kernel_h();
      int dW = param.stride_w();
      int dH = param.stride_h();
      int padW = param.pad_w();
      int padH = param.pad_h();
      if(kW==0 || kH==0)
      {
        kW = param.kernel_size();
        kH = kW;
      }
      if(dW==0 || dH==0)
      {
        dW = param.stride();
        dH = dW;
      }
      if(padW==0 && padH==0)
      {
        padW = param.pad();
        padH = padW;
      }

      char buf[1024];
      switch(cuda_package_type)
      {
        case CCN2: // ceil mode by default
          if(param.pool() == caffe::PoolingParameter::MAX)
            sprintf(buf, "ccn2.SpatialMaxPooling(%d, %d)", kW, dW);
          else if(param.pool() == caffe::PoolingParameter::AVE)
            sprintf(buf, "ccn2.SpatialAvgPooling(%d, %d)", kW, dW);
          else if(param.pool() == caffe::PoolingParameter::STOCHASTIC)
            THError("Stochastic pooling is not implemented in DHWB format");
          break;
        case CUDNN:
          if(param.pool() == caffe::PoolingParameter::MAX)
            sprintf(buf, "cudnn.SpatialMaxPooling(%d, %d, %d, %d, %d, %d):ceil()", kW, kH, dW, dH, padW, padH);
          else if(param.pool() == caffe::PoolingParameter::AVE)
            sprintf(buf, "cudnn.SpatialAveragePooling(%d, %d, %d, %d, %d, %d):ceil()", kW, kH, dW, dH, padW, padH);
          else if(param.pool() == caffe::PoolingParameter::STOCHASTIC)
            sprintf(buf, "inn.SpatialStochasticPooling(%d, %d, %d, %d)", kW, kH, dW, dH);
          break;
        case NN:
          if(param.pool() == caffe::PoolingParameter::MAX)
            sprintf(buf, "nn.SpatialMaxPooling(%d, %d, %d, %d, %d, %d):ceil()", kW, kH, dW, dH, padW, padH);
          else if(param.pool() == caffe::PoolingParameter::AVE)
            sprintf(buf, "inn.SpatialAveragePooling(%d, %d, %d, %d)", kW, kH, dW, dH); // padding is not supported yet
          else if(param.pool() == caffe::PoolingParameter::STOCHASTIC)
            sprintf(buf, "inn.SpatialStochasticPooling(%d, %d, %d, %d)", kW, kH, dW, dH);
          break;
      }
      lines.emplace_back(layer.name(), buf);
    }
    if(layer.type() == "ReLU")
    {
      if(cuda_package_type == CUDNN)
	lines.emplace_back(layer.name(), "cudnn.ReLU(true)");
      else
        lines.emplace_back(layer.name(), "nn.ReLU(true)");
    }
    if(layer.type() == "Sigmoid")
    {
      if(cuda_package_type == CUDNN)
	lines.emplace_back(layer.name(), "cudnn.Sigmoid(true)");
      else
        lines.emplace_back(layer.name(), "nn.Sigmoid()");
    }
    if(layer.type() == "Tanh")
    {
      if(cuda_package_type == CUDNN)
	lines.emplace_back(layer.name(), "cudnn.Tanh(true)");
      else
        lines.emplace_back(layer.name(), "nn.Tanh()");
    }
    if(layer.type() == "LRN")
    {
      auto &param = layer.lrn_param();
      int local_size = param.local_size();
      float alpha = param.alpha();
      float beta = param.beta();
      float k = param.k();
      char buf[1024];
      if(std::string(cuda_package) == "ccn2")
        sprintf(buf, "ccn2.SpatialCrossResponseNormalization(%d, %.6f, %.4f, %f)", local_size, alpha, beta, k);
      else
        sprintf(buf, "inn.SpatialCrossResponseNormalization(%d, %.6f, %.4f, %f)", local_size, alpha, beta, k);
      lines.emplace_back(layer.name(), buf);
    }
    if(layer.type() == "InnerProduct")
    {
      auto &param = layer.inner_product_param();
      int nInputPlane = layer.blobs(0).shape().dim(1);
      int nOutputPlane = param.num_output();
      char buf[1024];
      sprintf(buf, "nn.Linear(%d, %d)", nInputPlane, nOutputPlane);
      if(num_output != nInputPlane)
      {
        if(std::string(cuda_package) == "ccn2")
          lines.emplace_back("torch_transpose_bdwh", "nn.Transpose({4,1},{4,2},{4,3})");
        lines.emplace_back("torch_view", "nn.View(-1):setNumInputDims(3)");
      }
      lines.emplace_back(layer.name(), buf);
      num_output = nOutputPlane;
    }
    if(layer.type() == "Dropout")
    {
      char buf[1024];
      sprintf(buf, "nn.Dropout(%f)", layer.dropout_param().dropout_ratio());
      lines.emplace_back(layer.name(), buf);
    }
    if(layer.type()=="SoftmaxWithLoss")
    {
      lines.emplace_back(layer.name(), "nn.SoftMax()");
    }
    if(layer.type()=="Softmax")
    {
      lines.emplace_back(layer.name(), "nn.SoftMax()");
    }

    if(!lines.empty())
      for(auto& it: lines)
        ofs << "table.insert(model, {'" << it.first << "', " << it.second << "})\n";
    else
    {
      ofs << "-- warning: module '" << layer.name() << " [type " << layer.type() << "]" << "' not found\n";
      std::cout << "warning: module '" << layer.name() << " [type " << layer.type() << "]" << "' not found\n";
    }
  }
  ofs << "return model";
}


void loadBinary(void** handle, const char* prototxt_name, const char* binary_name)
{
  caffe::NetParameter* netparam = new caffe::NetParameter();
  ReadProtoFromTextFile(prototxt_name, netparam);
  bool success = ReadProtoFromBinaryFile(binary_name, netparam);
  if(success)
  {
    std::cout << "Successfully loaded " << binary_name << std::endl;
    handle[1] = netparam;
  }
  else
    std::cout << "Couldn't load " << binary_name << std::endl;
}

void destroyBinary(void** handle)
{
  const caffe::NetParameter* netparam2 = (const caffe::NetParameter*)handle[1];
  delete netparam2;
}

void loadModule(const void** handle, const char* name, THFloatTensor* weight, THFloatTensor* bias)
{
  if(handle == NULL)
  {
    std::cout << "network not loaded!\n";
    return;
  }

  const caffe::NetParameter* netparam = (const caffe::NetParameter*)handle[1];

  if (netparam->layers_size() > 0)
      loadModuleV1(netparam, name, weight, bias);
  else
      loadModuleV2(netparam, name, weight, bias);
}


void loadModuleV1(const caffe::NetParameter* netparam, const char* name, THFloatTensor* weight, THFloatTensor* bias)
{
  int n = netparam->layers_size();
  for(int i=0; i<n; ++i)
  {
    auto &layer = netparam->layers(i);
    if(std::string(name) == layer.name())
    {
      int nInputPlane = layer.blobs(0).channels();
      int nOutputPlane = layer.blobs(0).num();
      int kW = layer.blobs(0).width();
      int kH = layer.blobs(0).height();
      printf("%s: %d %d %d %d\n", name, nOutputPlane, nInputPlane, kW, kH);

      THFloatTensor_resize4d(weight, nOutputPlane, nInputPlane, kW, kH);
      memcpy(THFloatTensor_data(weight), layer.blobs(0).data().data(), sizeof(float)*nOutputPlane*nInputPlane*kW*kH);

      THFloatTensor_resize1d(bias, layer.blobs(1).data_size());
      memcpy(THFloatTensor_data(bias), layer.blobs(1).data().data(), sizeof(float)*layer.blobs(1).data_size());
    }
  }
}


void loadModuleV2(const caffe::NetParameter* netparam, const char* name, THFloatTensor* weight, THFloatTensor* bias)
{
  int n = netparam->layer_size();
  for(int i=0; i<n; ++i)
  {
    auto &layer = netparam->layer(i);
    if(std::string(name) == layer.name())
    {
      int nInputPlane = layer.blobs(0).shape().dim(1);
      int nOutputPlane = layer.blobs(0).shape().dim(0);
      int kW = layer.blobs(0).shape().dim(3);
      int kH = layer.blobs(0).shape().dim(2);
      if(layer.type() == "InnerProduct")
      {
        printf("%s: %d %d %d %d\n", name, 1, 1, nInputPlane, nOutputPlane);
        THFloatTensor_resize4d(weight, 1, 1, nInputPlane, nOutputPlane);
        memcpy(THFloatTensor_data(weight), layer.blobs(0).data().data(), sizeof(float)*nOutputPlane*nInputPlane);
      }
      else
      {
        printf("%s: %d %d %d %d\n", name, nOutputPlane, nInputPlane, kW, kH);
        THFloatTensor_resize4d(weight, nOutputPlane, nInputPlane, kW, kH);
        memcpy(THFloatTensor_data(weight), layer.blobs(0).data().data(), sizeof(float)*nOutputPlane*nInputPlane*kW*kH);
      }
      THFloatTensor_resize1d(bias, layer.blobs(1).data_size());
      memcpy(THFloatTensor_data(bias), layer.blobs(1).data().data(), sizeof(float)*layer.blobs(1).data_size());
    }
  }
}