name:"char-rnn"
train_steps: 100000
disp_freq: 100
#debug: true
gpu: 0
train_one_batch {
  alg: kBPTT
}

updater {
  type: kRMSProp
  rmsprop_conf {
    rho: 0.95
  }
  learning_rate {
    type: kStep
    base_lr: 0.002
    step_conf {
      gamma: 0.97
      change_freq: 2000
    }
  }
  clip_low: -5
  clip_high: 5
}

neuralnet {
  unroll_len: 50
  layer {
    name: "data"
    type: kCharRNN
    unroll_len: 1
    char_rnn_conf {
      path: "examples/char-rnn/linux_input.txt"
      vocab_path:"examples/char-rnn/vocab.txt"
      batchsize: 50
      unroll_len: 50
    }
  }
  layer {
    name: "onehot"
    type: kOneHot
    srclayers: "data"
    unroll_conn_type: kUnrollOneToAll
    onehot_conf {
      vocab_size: 101
    }
  }

  layer {
    name: "label"
    type: kRNNLabel
    srclayers: "data"
    unroll_conn_type: kUnrollOneToAll
  }

  layer {
    name: "gru1"
    type: kGRU
    srclayers: "onehot"
    gru_conf {
      dim_hidden: 512
    }
    param {
      name: "z_hx"
      init {
        type: kUniform
        low: -0.08
        high: 0.08
      }
    }
    param {
      name: "r_hx"
      init {
        type: kUniform
        low: -0.08
        high: 0.08
      }
    }
    param {
      name: "c_hx"
      init {
        type: kUniform
        low: -0.08
        high: 0.08
      }
    }
    param {
      name: "z_hh"
      init {
        type: kUniform
        low: -0.08
        high: 0.08
      }
    }
    param {
      name: "r_hh"
      init {
        type: kUniform
        low: -0.08
        high: 0.08
      }
    }
    param {
      name: "c_hh"
      init {
        type: kUniform
        low: -0.08
        high: 0.08
      }
    }
    param {
      name: "z_b"
      init {
        type: kUniform
        low: -0.08
        high: 0.08
      }
    }
    param {
      name: "r_b"
      init {
        type: kUniform
        low: -0.08
        high: 0.08
      }
    }
    param {
      name: "c_b"
      init {
        type: kUniform
        low: -0.08
        high: 0.08
      }
    }

  }
#  layer {
#    name: "gru2"
#    type: kGRU
#    srclayers: "gru1"
#    gru_conf {
#      dim_hidden: 512
#    }
#    param {
#      name: "z_hx2"
#      init {
#        type: kUniform
#        low: -0.08
#        high: 0.08
#      }
#    }
#    param {
#      name: "r_hx2"
#      init {
#        type: kUniform
#        low: -0.08
#        high: 0.08
#      }
#    }
#    param {
#      name: "c_hx2"
#      init {
#        type: kUniform
#        low: -0.08
#        high: 0.08
#      }
#    }
#    param {
#      name: "z_hh2"
#      init {
#        type: kUniform
#        low: -0.08
#        high: 0.08
#      }
#    }
#    param {
#      name: "r_hh2"
#      init {
#        type: kUniform
#        low: -0.08
#        high: 0.08
#      }
#    }
#    param {
#      name: "c_hh2"
#      init {
#        type: kUniform
#        low: -0.08
#        high: 0.08
#      }
#    }
#    param {
#      name: "z_b2"
#      init {
#        type: kUniform
#        low: -0.08
#        high: 0.08
#      }
#    }
#    param {
#      name: "r_b2"
#      init {
#        type: kUniform
#        low: -0.08
#        high: 0.08
#      }
#    }
#    param {
#      name: "c_b2"
#      init {
#        type: kUniform
#        low: -0.08
#        high: 0.08
#      }
#    }
#  }
#
  layer {
    name: "ip1"
    type: kInnerProduct
    srclayers: "gru1"
    innerproduct_conf {
      num_output: 101
    }
    param {
      name: "w"
      init {
        type: kUniform
        low: -0.08
        high: 0.08
      }
    }
    param {
      name: "b"
      init {
        type: kUniform
        low: -0.08
        high: 0.08
      }
    }
  }
  layer {
    name: "loss"
    type: kSoftmaxLoss
    srclayers: "ip1"
    srclayers: "label"
  }
}

cluster {
  workspace: "examples/char-rnn/"
}