Sensei

Sensei is a logistic regression engine. It is intended to be run on one machine on medium-sized data.

It is currently capable of:

• reading datasets in LIBSVM format (with binary features only)
• highly configurable training
• automated feature exploration
• automated feature pruning
• saving models for later training or scoring
• data scoring

Build

To build Sensei you need to install the following prerequisites:

• C++11 compatibile compiler
• cmake
• gflags
• glog
• protobuf
• sparsehash

If you use Ubuntu you can install all the prerequisites by running the following command:

sudo apt-get install g++ \
                     cmake \
                     protobuf-compiler \
                     libprotobuf-dev \
                     libgflags-dev \
                     libgoogle-glog-dev \
                     libsparsehash-dev

With all the necessary dependencies installed you can build Sensei by running:

cmake . && make

To build a debug version of the binary, you can run:

cmake -DCMAKE_BUILD_TYPE=Debug . && make

Test

To build Sensei unit tests you need to install the following additional prerequisites:

• gtest

If gtest is installed on your machine, Sensei will use the installed library. Otherwise, you can put gtest as a subdirectory in Sensei top-level directory. You can do so by running:

wget https://googletest.googlecode.com/files/gtest-1.7.0.zip && unzip gtest-1.7.0.zip

To build and run Sensei unit tests run:

cmake . && make && ctest

Usage

Sensei is configured and controlled by commands defined in sensei/config.proto and sensei/common_config.proto. The commands must be provided in text protobufer format. You can read more about specific command options in comments in aforementioned files.

To execute Sensei with a specified config run:

./sensei_bin --config_files path/to/config_file

You can find sample configuration files in the examples directory.

Sample config file

command_list {
  command {
    read_data {
      data_reader {
        format: LIBSVM
        training_set {
          files_glob: "input.libsvm"
        }
        feature_spec {
          product {
            prefix: ""
          }
          product {
          }
        }
      }
    }
  }
  command {
    fit_model_weights {
      iterations: 100
    }
  }
  command {
    write_model {
      set {
        format: TEXT
        output_model_path: "model.txt"
      }
    }
  }
  command {
    write_model {
      write {
      }
    }
  }
}