name: "alexnet"
train_steps: 450000
test_steps: 500
test_freq: 1000
disp_freq: 20
checkpoint_freq: 100000
checkpoint_after: 100000
gpu: 0
#debug: true
#checkpoint_path: "examples/alexnet/checkpoint/step10000-worker0"
train_one_batch {
  alg: kBP
}
updater{
  type: kSGD
  weight_decay: 0.0005
  momentum: 0.9
  learning_rate {
    type: kStep
    base_lr: 0.01
    step_conf {
      gamma: 0.1
      change_freq: 100000
    }
  }
}
neuralnet {
  layer{
    name: "data"
    type: kRecordInput
    store_conf {
      backend: "kvfile"
      path :"/data/dataset/imagenet/train_record.bin"
      mean_file: "/data/dataset/imagenet/image_mean.bin"
      batchsize: 256
      #random_skip: 1000
      shape: 3
      shape: 256
      shape: 256
    }
    include: kTrain
  }
  layer{
    name: "data"
    type: kRecordInput
    store_conf {
      backend: "kvfile"
      path :"/data/dataset/imagenet/val_record.bin"
      mean_file: "/data/dataset/imagenet/image_mean.bin"
      batchsize: 100
      shape: 3
      shape: 256
      shape: 256
    }
    include: kTest
  }
  layer{
    name: "image"
    type: kImagePreprocess
    rgbimage_conf {
      cropsize: 227
      mirror: true
    }
#    partition_dim: 0
    srclayers: "data"
  }
  layer{
    name: "conv1"
    type: kCudnnConv
    srclayers: "image"
    convolution_conf {
      num_filters: 96
      kernel: 11
      stride: 4
    }
#    partition_dim: 0
    param {
      name: "w1"
      init {
        type: kGaussian
        std: 0.01
      }
    }
    param {
      name: "b1"
      lr_scale: 2
      wd_scale: 0
      init {
        type: kConstant
        value: 0
      }
    }
  }
  layer {
    name: "relu1"
    type: kCudnnActivation
    activation_conf {
      type: RELU
    }
    share_src_blobs: true
    srclayers: "conv1"
#    partition_dim: 0
  }
  layer {
    name: "pool1"
    type: kCudnnPool
    pooling_conf {
      pool: MAX
      kernel: 3
      stride: 2
    }
    srclayers: "relu1"
#    partition_dim: 0
  }
  layer {
    name: "norm1"
    type: kCudnnLRN
    lrn_conf {
      local_size: 5
      alpha: 0.0001
      beta: 0.75
      knorm: 1
    }
    srclayers: "pool1"
#    partition_dim: 0
  }

  layer{
    name: "conv2"
    type: kCudnnConv
    srclayers: "norm1"
    convolution_conf {
      num_filters: 256
      kernel: 5
      pad: 2
    }
#    partition_dim: 0
    param {
      name: "w2"
      init {
        type: kGaussian
        std: 0.01
      }
    }
    param {
      name: "b2"
      lr_scale: 2
      wd_scale: 0
      init {
        type: kConstant
        value: 1
      }
    }
  }
  layer {
    name: "relu2"
    type: kCudnnActivation
    activation_conf {
      type: RELU
    }
    share_src_blobs: true
    srclayers: "conv2"
#    partition_dim: 0
  }
  layer {
    name: "pool2"
    type: kCudnnPool
    pooling_conf {
      pool: MAX
      kernel: 3
      stride: 2
    }
    srclayers: "relu2"
#    partition_dim: 0
  }

  layer {
    name: "norm2"
    type: kCudnnLRN
    lrn_conf {
      local_size: 5
      alpha: 0.0001
      beta: 0.75
      knorm: 1
    }
    srclayers: "pool2"
#    partition_dim: 0
  }
  layer{
    name: "conv3"
    type: kCudnnConv
    srclayers: "norm2"
    convolution_conf {
      num_filters: 384
      kernel: 3
      pad: 1
    }
#    partition_dim: 0
    param {
      name: "w3"
      init {
        type: kGaussian
        std: 0.01
      }
    }
    param {
      name: "b3"
      lr_scale: 2
      wd_scale: 0
      init {
        type: kConstant
        value: 0
      }
    }
  }
  layer {
    name: "relu3"
    type: kCudnnActivation
    activation_conf {
      type: RELU
    }
    share_src_blobs: true
    srclayers: "conv3"
#    partition_dim: 0
  }
  layer{
    name: "conv4"
    type: kCudnnConv
    srclayers: "relu3"
    convolution_conf {
      num_filters: 384
      kernel: 3
      pad: 1
    }
#    partition_dim: 0
    param {
      name: "w4"
      init {
        type: kGaussian
        std: 0.01
      }
    }
    param {
      name: "b4"
      lr_scale:2
      wd_scale:0
      init {
        type: kConstant
        value: 1
      }
    }
  }
  layer {
    name: "relu4"
    type: kCudnnActivation
    activation_conf {
      type: RELU
    }
    share_src_blobs: true
    srclayers: "conv4"
#    partition_dim: 0
  }
  layer{
    name: "conv5"
    type: kCudnnConv
    srclayers: "relu4"
    convolution_conf {
      num_filters: 256
      kernel: 3
      pad: 1
    }
#    partition_dim: 0
    param {
      name: "w5"
      init {
        type: kGaussian
        std: 0.01
      }
    }
    param {
      name: "b5"
      lr_scale: 2
      wd_scale: 0
      init {
        type: kConstant
        value: 1
      }
    }
  }
  layer {
    name: "relu5"
    type: kCudnnActivation
    activation_conf {
      type: RELU
    }
    share_src_blobs: true
    srclayers: "conv5"
#    partition_dim: 0
  }
  layer {
    name: "pool5"
    type: kCudnnPool
    pooling_conf {
      pool: MAX
      kernel: 3
      stride: 2
    }
    srclayers: "relu5"
#    partition_dim: 0
  }
  layer {
    name: "ip6"
    type: kInnerProduct
    innerproduct_conf {
      num_output: 4096
    }
    param {
      name: "w6"
      init {
        type: kGaussian
        std: 0.005
      }
    }
    param {
      name: "b6"
      lr_scale: 2
      wd_scale: 0
      init {
        type: kConstant
        value: 1
      }
    }
    srclayers: "pool5"
#    partition_dim: 1
  }
  layer {
    name: "relu6"
    type: kCudnnActivation
    activation_conf {
      type: RELU
    }
    share_src_blobs: true
    srclayers: "ip6"
#    partition_dim: 1
  }
  layer {
    name: "drop6"
    type: kDropout
    srclayers: "relu6"
#    partition_dim: 1
  }
  layer {
    name: "ip7"
    type: kInnerProduct
    innerproduct_conf {
      num_output: 4096
    }
#    partition_dim: 1
    param {
      name: "w7"
      init {
        type: kGaussian
        std: 0.005
      }
    }
    param {
      name: "b7"
      lr_scale: 2
      wd_scale: 0
      init {
        type: kConstant
        value: 1
      }
    }
    srclayers: "drop6"
  }
  layer {
    name: "relu7"
    type: kCudnnActivation
    activation_conf {
      type: RELU
    }
    share_src_blobs: true
    srclayers: "ip7"
#    partition_dim: 1
  }
  layer {
    name: "drop7"
    type: kDropout
    srclayers: "relu7"
#    partition_dim: 1
  }
  layer {
    name: "ip8"
    type: kInnerProduct
    innerproduct_conf {
      num_output: 1000
    }
#    partition_dim: 1
    param {
      name: "w8"
      init {
        type: kGaussian
        std: 0.01
      }
    }
    param {
      name: "b8"
      lr_scale: 2
      wd_scale: 0
      init {
        type: kConstant
        value: 0
      }
    }
    srclayers: "drop7"
  }
  layer {
    name: "loss"
    type: kCudnnSoftmaxLoss
    softmaxloss_conf {
      topk:1
    }
    srclayers: "ip8"
    srclayers: "data"
    include: kTrain
  }
  layer {
   name : "softmax"
   type: kCudnnSoftmax
   srclayers: "ip8"
   include: kTest
  }
  layer {
   name : "accuracy"
   type: kAccuracy
   srclayers: "softmax"
   srclayers: "data"
   include: kTest
  }
}
cluster {
  nworker_groups: 1
  nserver_groups: 1
  nworkers_per_group: 1
  nworkers_per_procs: 1
  workspace: "examples/alexnet"
}