remove benchmarks (#1018)

Former-commit-id: 7ab4cee4

benchmark/alexnet/1_machine_1_gpu.job

-net: "./train.net"
-resource: "./1_machine_4_gpu.resource"
-placement: "./1_machine_1_gpu.placement"
-other: "./piece_size_512.other"

benchmark/alexnet/1_machine_1_gpu.placement

-placement_group {
-  op_set {
-    op_name: "decode"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:cpu:0-1"
-  }
-}
-
-placement_group {
-  op_set {
-    op_name: "transpose"
-    op_name: "conv1"
-    op_name: "pool1"
-    op_name: "conv2"
-    op_name: "pool2"
-    op_name: "conv3"
-    op_name: "conv4"
-    op_name: "conv5"
-    op_name: "pool5"
-    op_name: "dropout1"
-    op_name: "dropout2"
-    op_name: "softmax_loss"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:gpu:0"
-  }
-}
-
-placement_group {
-  op_set {
-    op_name: "fc1"
-    op_name: "fc2"
-    op_name: "fc3"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:gpu:0"
-  }
-}

benchmark/alexnet/1_machine_2_gpu.job

-net: "./train.net"
-resource: "./1_machine_4_gpu.resource"
-placement: "./1_machine_2_gpu.placement"
-other: "./piece_size_1024.other"

benchmark/alexnet/1_machine_2_gpu.placement

-placement_group {
-  op_set {
-    op_name: "decode"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:cpu:0-1"
-  }
-}
-
-placement_group {
-  op_set {
-    op_name: "transpose"
-    op_name: "conv1"
-    op_name: "pool1"
-    op_name: "conv2"
-    op_name: "pool2"
-    op_name: "conv3"
-    op_name: "conv4"
-    op_name: "conv5"
-    op_name: "pool5"
-    op_name: "dropout1"
-    op_name: "dropout2"
-    op_name: "softmax_loss"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:gpu:0-1"
-  }
-}
-
-placement_group {
-  op_set {
-    op_name: "fc1"
-    op_name: "fc2"
-    op_name: "fc3"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:gpu:0-1"
-  }
-}

benchmark/alexnet/1_machine_4_gpu.job

-net: "./train.net"
-resource: "./1_machine_4_gpu.resource"
-placement: "./1_machine_4_gpu.placement"
-other: "./piece_size_1024.other"

benchmark/alexnet/1_machine_4_gpu.placement

-placement_group {
-  op_set {
-    op_name: "decode"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:cpu:0-1"
-  }
-}
-
-placement_group {
-  op_set {
-    op_name: "transpose"
-    op_name: "conv1"
-    op_name: "pool1"
-    op_name: "conv2"
-    op_name: "pool2"
-    op_name: "conv3"
-    op_name: "conv4"
-    op_name: "conv5"
-    op_name: "pool5"
-    op_name: "dropout1"
-    op_name: "dropout2"
-    op_name: "softmax_loss"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:gpu:0-3"
-  }
-}
-
-placement_group {
-  op_set {
-    op_name: "fc1"
-    op_name: "fc2"
-    op_name: "fc3"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:gpu:0-3"
-  }
-}

benchmark/alexnet/1_machine_4_gpu.resource

-machine {
-  addr: "127.0.0.1"
-  port: 6789
-  name: "first"
-}
-
-gpu_device_num: 4

benchmark/alexnet/piece_size_1024.other

-use_rdma: true
-globalfs_conf {
-  localfs_conf {
-  }
-}
-data_part_num: 32
-piece_size: 1024
-train_conf {
-  model_save_snapshots_path: "/tmp/snapshot"
-  num_of_batches_in_snapshot: 3000
-  total_batch_num: 30000
-  default_initializer_conf {
-    random_normal_conf {
-      mean: 0.0
-      std: 0.01
-    }
-  }
-  batch_size: 1024
-  l2: 0.0005
-  model_update_conf {
-    learning_rate: 0.01
-    naive_conf {
-    }
-    learning_rate_decay {
-      exponential_conf {
-        decay_rate: 0.1
-        decay_batches: 10000
-        staircase: true
-      }
-    }
-  }
-}

benchmark/alexnet/piece_size_512.other

-use_rdma: true
-globalfs_conf {
-  localfs_conf {
-  }
-}
-data_part_num: 32
-piece_size: 512
-train_conf {
-  model_save_snapshots_path: "/tmp/snapshot"
-  num_of_batches_in_snapshot: 3000
-  total_batch_num: 30000
-  default_initializer_conf {
-    random_normal_conf {
-      mean: 0.0
-      std: 0.01
-    }
-  }
-  batch_size: 1024
-  l2: 0.0005
-  model_update_conf {
-    learning_rate: 0.01
-    naive_conf {
-    }
-    learning_rate_decay {
-      exponential_conf {
-        decay_rate: 0.1
-        decay_batches: 10000
-        staircase: true
-      }
-    }
-  }
-}

benchmark/alexnet/report.md

-batch_size: 1024
-
-gpu num | time (one batch)
-:-------| :-------------
-1       | 549ms
-2       | 285ms
-4       | 171ms

benchmark/alexnet/train.net

-op {
-  name: "decode"
-  decode_ofrecord_conf {
-    data_dir: "/dataset/imagenet_227/train/32/"
-    blob {
-      name: "encoded"
-      data_type: kFloat
-      shape {
-        dim: 227
-        dim: 227
-        dim: 3
-      }
-      encode_case {
-        jpeg {
-        }
-      }
-    }
-    blob {
-      name: "class/label"
-      data_type: kInt32
-      shape {
-      }
-      encode_case {
-        raw {
-        }
-      }
-    }
-  }
-}
-
-op {
-  name: "transpose"
-  transpose_conf {
-    in: "decode/encoded"
-    out: "out"
-    perm: 3
-    perm: 1
-    perm: 2
-  }
-}
-
-op {
-  name: "conv1"
-  conv_2d_conf {
-    in: "transpose/out"
-    out: "out"
-    filters: 64
-    padding: "VALID"
-    data_format: "channels_first"
-    kernel_size: 11
-    kernel_size: 11
-    strides: 4
-    strides: 4
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    use_bias: true
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-
-#op {
-#  name: "lrn1"
-#  local_response_normalization_conf {
-#    in: "conv1/out"
-#    out: "out"
-#    depth_radius: 2
-#    bias: 2
-#    alpha: 0.0001
-#    beta: 0.75
-#    data_format: "channels_first"
-#  }
-#}
-
-op {
-  name: "pool1"
-  max_pooling_2d_conf {
-    in: "conv1/out"
-    out: "out"
-    padding: "VALID"
-    data_format: "channels_first"
-    pool_size: 3
-    pool_size: 3
-    strides: 2
-    strides: 2
-  }
-}
-
-op {
-  name: "conv2"
-  conv_2d_conf {
-    in: "pool1/out"
-    out: "out"
-    filters: 192
-    padding: "SAME"
-    data_format: "channels_first"
-    kernel_size: 5
-    kernel_size: 5
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    use_bias: true
-    bias_initializer {
-      constant_conf {
-        value: 0.5
-      }
-    }
-  }
-}
-
-#op {
-#  name: "lrn2"
-#  local_response_normalization_conf {
-#    in: "conv2/out"
-#    out: "out"
-#    depth_radius: 2
-#    bias: 2
-#    alpha: 0.0001
-#    beta: 0.75
-#    data_format: "channels_first"
-#  }
-#}
-
-op {
-  name: "pool2"
-  max_pooling_2d_conf {
-    in: "conv2/out"
-    out: "out"
-    padding: "VALID"
-    data_format: "channels_first"
-    pool_size: 3
-    pool_size: 3
-    strides: 2
-    strides: 2
-  }
-}
-
-op {
-  name: "conv3"
-  conv_2d_conf {
-    in: "pool2/out"
-    out: "out"
-    filters: 384
-    padding: "SAME"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    use_bias: true
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-
-op {
-  name: "conv4"
-  conv_2d_conf {
-    in: "conv3/out"
-    out: "out"
-    filters: 384
-    padding: "SAME"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    use_bias: true
-    bias_initializer {
-      constant_conf {
-        value: 0.5
-      }
-    }
-  }
-}
-
-op {
-  name: "conv5"
-  conv_2d_conf {
-    in: "conv4/out"
-    out: "out"
-    filters: 256
-    padding: "SAME"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    use_bias: true
-    bias_initializer {
-      constant_conf {
-        value: 0.5
-      }
-    }
-  }
-}
-
-op {
-  name: "pool5"
-  max_pooling_2d_conf {
-    in: "conv5/out"
-    out: "out"
-    padding: "VALID"
-    data_format: "channels_first"
-    pool_size: 3
-    pool_size: 3
-    strides: 2
-    strides: 2
-  }
-}
-
-op {
-  name: "fc1"
-  fully_connected_conf {
-    in: "pool5/out"
-    out: "out"
-    units: 4096
-    activation: kRelu
-    bias_initializer {
-      constant_conf {
-        value: 0.5
-      }
-    }
-  }
-}
-
-op {
-  name: "dropout1"
-  dropout_conf {
-    in: "fc1/out"
-    out: "out"
-    rate: 0.5
-  }
-}
-
-op {
-  name: "fc2"
-  fully_connected_conf {
-    in: "dropout1/out"
-    out: "out"
-    units: 4096
-    activation: kRelu
-    bias_initializer {
-      constant_conf {
-        value: 0.5
-      }
-    }
-  }
-}
-
-op {
-  name: "dropout2"
-  dropout_conf {
-    in: "fc2/out"
-    out: "out"
-    rate: 0.5
-  }
-}
-
-op {
-  name: "fc3"
-  fully_connected_conf {
-    in: "dropout2/out"
-    out: "out"
-    units: 1000
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-
-op {
-  name: "softmax_loss"
-  sparse_softmax_cross_entropy_loss_conf {
-    prediction: "fc3/out"
-    label: "decode/class/label"
-    loss: "loss"
-  }
-}

benchmark/vgg16/1_machine_1_gpu.job

-net: "./train.net"
-resource: "./1_machine_4_gpu.resource"
-placement: "./1_machine_1_gpu.placement"
-other: "./piece_size_16.other"

benchmark/vgg16/1_machine_1_gpu.placement

-placement_group {
-  op_set {
-    op_name: "decode"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:cpu:0-1"
-  }
-}
-
-placement_group {
-  op_set {
-    op_name: "transpose"
-    op_name: "conv1_1"
-    op_name: "conv1_2"
-    op_name: "pool1"
-    op_name: "conv2_1"
-    op_name: "conv2_2"
-    op_name: "pool2"
-    op_name: "conv3_1"
-    op_name: "conv3_2"
-    op_name: "conv3_3"
-    op_name: "pool3"
-    op_name: "conv4_1"
-    op_name: "conv4_2"
-    op_name: "conv4_3"
-    op_name: "pool4"
-    op_name: "conv5_1"
-    op_name: "conv5_2"
-    op_name: "conv5_3"
-    op_name: "pool5"
-    op_name: "fc6"
-    op_name: "drop6"
-    op_name: "fc7"
-    op_name: "drop7"
-    op_name: "fc8"
-    op_name: "softmax_loss"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:gpu:0"
-  }
-}

benchmark/vgg16/1_machine_2_gpu.job

-net: "./train.net"
-resource: "./1_machine_4_gpu.resource"
-placement: "./1_machine_2_gpu.placement"
-other: "./piece_size_16.other"

benchmark/vgg16/1_machine_2_gpu.placement

-placement_group {
-  op_set {
-    op_name: "decode"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:cpu:0-1"
-  }
-}
-
-placement_group {
-  op_set {
-    op_name: "transpose"
-    op_name: "conv1_1"
-    op_name: "conv1_2"
-    op_name: "pool1"
-    op_name: "conv2_1"
-    op_name: "conv2_2"
-    op_name: "pool2"
-    op_name: "conv3_1"
-    op_name: "conv3_2"
-    op_name: "conv3_3"
-    op_name: "pool3"
-    op_name: "conv4_1"
-    op_name: "conv4_2"
-    op_name: "conv4_3"
-    op_name: "pool4"
-    op_name: "conv5_1"
-    op_name: "conv5_2"
-    op_name: "conv5_3"
-    op_name: "pool5"
-    op_name: "fc6"
-    op_name: "drop6"
-    op_name: "fc7"
-    op_name: "drop7"
-    op_name: "fc8"
-    op_name: "softmax_loss"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:gpu:0-1"
-  }
-}

benchmark/vgg16/1_machine_4_gpu.job

-net: "./train.net"
-resource: "./1_machine_4_gpu.resource"
-placement: "./1_machine_4_gpu.placement"
-other: "./piece_size_64.other"

benchmark/vgg16/1_machine_4_gpu.placement

-placement_group {
-  op_set {
-    op_name: "decode"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:cpu:0-1"
-  }
-}
-
-placement_group {
-  op_set {
-    op_name: "transpose"
-    op_name: "conv1_1"
-    op_name: "conv1_2"
-    op_name: "pool1"
-    op_name: "conv2_1"
-    op_name: "conv2_2"
-    op_name: "pool2"
-    op_name: "conv3_1"
-    op_name: "conv3_2"
-    op_name: "conv3_3"
-    op_name: "pool3"
-    op_name: "conv4_1"
-    op_name: "conv4_2"
-    op_name: "conv4_3"
-    op_name: "pool4"
-    op_name: "conv5_1"
-    op_name: "conv5_2"
-    op_name: "conv5_3"
-    op_name: "pool5"
-    op_name: "fc6"
-    op_name: "drop6"
-    op_name: "fc7"
-    op_name: "drop7"
-    op_name: "fc8"
-    op_name: "softmax_loss"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:gpu:0-3"
-  }
-}

benchmark/vgg16/1_machine_4_gpu.resource

-machine {
-  addr: "192.168.1.12"
-  port: 7776
-  name: "first"
-}
-
-gpu_device_num: 4

benchmark/vgg16/2_machine_4_gpu.job

-net: "./train.net"
-resource: "./2_machine_4_gpu.resource"
-placement: "./2_machine_4_gpu.placement"
-other: "./hdfs_piece_size_64.other"

benchmark/vgg16/2_machine_4_gpu.placement

-placement_group {
-  op_set {
-    op_name: "decode"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:cpu:0-1"
-    device_name: "second:cpu:0-1"
-  }
-}
-
-placement_group {
-  op_set {
-    op_name: "transpose"
-    op_name: "conv1_1"
-    op_name: "conv1_2"
-    op_name: "pool1"
-    op_name: "conv2_1"
-    op_name: "conv2_2"
-    op_name: "pool2"
-    op_name: "conv3_1"
-    op_name: "conv3_2"
-    op_name: "conv3_3"
-    op_name: "pool3"
-    op_name: "conv4_1"
-    op_name: "conv4_2"
-    op_name: "conv4_3"
-    op_name: "pool4"
-    op_name: "conv5_1"
-    op_name: "conv5_2"
-    op_name: "conv5_3"
-    op_name: "pool5"
-    op_name: "softmax_loss"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:gpu:0-1"
-    device_name: "second:gpu:0-1"
-  }
-}
-
-placement_group {
-  op_set {
-    op_name: "fc6"
-    op_name: "drop6"
-    op_name: "fc7"
-    op_name: "drop7"
-    op_name: "fc8"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:gpu:0-1"
-    device_name: "second:gpu:0-1"
-  }
-}

benchmark/vgg16/2_machine_4_gpu.resource

-machine {
-  addr: "192.168.1.15"
-  port: 7767
-  name: "first"
-}
-
-machine {
-  addr: "192.168.1.16"
-  port: 7766
-  name: "second"
-}
-
-gpu_device_num: 2

benchmark/vgg16/2_machine_8_gpu.job

-net: "./train.net"
-resource: "./2_machine_8_gpu.resource"
-placement: "./2_machine_8_gpu.placement"
-other: "./hdfs_piece_size_128.other"

benchmark/vgg16/2_machine_8_gpu.placement

-placement_group {
-  op_set {
-    op_name: "decode"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:cpu:0-1"
-    device_name: "second:cpu:0-1"
-  }
-}
-
-placement_group {
-  op_set {
-    op_name: "transpose"
-    op_name: "conv1_1"
-    op_name: "conv1_2"
-    op_name: "pool1"
-    op_name: "conv2_1"
-    op_name: "conv2_2"
-    op_name: "pool2"
-    op_name: "conv3_1"
-    op_name: "conv3_2"
-    op_name: "conv3_3"
-    op_name: "pool3"
-    op_name: "conv4_1"
-    op_name: "conv4_2"
-    op_name: "conv4_3"
-    op_name: "pool4"
-    op_name: "conv5_1"
-    op_name: "conv5_2"
-    op_name: "conv5_3"
-    op_name: "pool5"
-    op_name: "softmax_loss"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:gpu:0-3"
-    device_name: "second:gpu:0-3"
-  }
-}
-
-placement_group {
-  op_set {
-    op_name: "fc6"
-    op_name: "drop6"
-    op_name: "fc7"
-    op_name: "drop7"
-    op_name: "fc8"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:gpu:0-3"
-    device_name: "second:gpu:0-3"
-  }
-}

benchmark/vgg16/2_machine_8_gpu.resource

-machine {
-  addr: "192.168.1.12"
-  port: 7766
-  name: "first"
-}
-
-machine {
-  addr: "192.168.1.14"
-  port: 7767
-  name: "second"
-}
-
-gpu_device_num: 4

benchmark/vgg16/2_machine_8_gpu_reduce.job

-net: "./train.net"
-resource: "./2_machine_8_gpu.resource"
-placement: "./2_machine_8_gpu_reduce.placement"
-other: "./hdfs_piece_size_8.other"

benchmark/vgg16/2_machine_8_gpu_reduce.placement

-placement_group {
-  op_set {
-    op_name: "decode"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:cpu:0-1"
-    device_name: "second:cpu:0-1"
-  }
-}
-
-placement_group {
-  op_set {
-    op_name: "transpose"
-    op_name: "conv1_1"
-    op_name: "conv1_2"
-    op_name: "pool1"
-    op_name: "conv2_1"
-    op_name: "conv2_2"
-    op_name: "pool2"
-    op_name: "conv3_1"
-    op_name: "conv3_2"
-    op_name: "conv3_3"
-    op_name: "pool3"
-    op_name: "conv4_1"
-    op_name: "conv4_2"
-    op_name: "conv4_3"
-    op_name: "pool4"
-    op_name: "conv5_1"
-    op_name: "conv5_2"
-    op_name: "conv5_3"
-    op_name: "pool5"
-    op_name: "fc6"
-    op_name: "drop6"
-    op_name: "fc7"
-    op_name: "drop7"
-    op_name: "fc8"
-    op_name: "softmax_loss"
-  }
-  parallel_conf {
-    policy: kDataParallel
-    device_name: "first:gpu:0-3"
-    device_name: "second:gpu:0-3"
-  }
-}

benchmark/vgg16/hdfs_piece_size_128.other

-use_rdma: true
-globalfs_conf {
-  hdfs_conf {
-    namenode: "hdfs://192.168.1.11:9000"
-  }
-}
-piece_size: 128
-data_part_num: 32
-max_data_id_length: 0
-train_conf {
-  model_save_snapshots_path: "/zhuyi/vgg_snapshots_buf"
-  num_of_batches_in_snapshot: 1000
-  staleness: 0
-  total_batch_num: 50
-  default_initializer_conf {
-    msra_conf {
-      variance_norm: kFanOut
-    }
-  }
-  batch_size: 256
-  l2: 0.0005
-  model_update_conf {
-    momentum_conf {
-      beta: 0.9
-    }
-  }
-}

benchmark/vgg16/hdfs_piece_size_32.other

-use_rdma: true
-globalfs_conf {
-  hdfs_conf {
-    namenode: "hdfs://192.168.1.11:9000"
-  }
-}
-piece_size: 32
-data_part_num: 32
-max_data_id_length: 0
-train_conf {
-  model_save_snapshots_path: "/zhuyi/vgg_snapshots_buf"
-  num_of_batches_in_snapshot: 1000
-  staleness: 0
-  total_batch_num: 50
-  default_initializer_conf {
-    msra_conf {
-      variance_norm: kFanOut
-    }
-  }
-  batch_size: 256
-  l2: 0.0005
-  model_update_conf {
-    momentum_conf {
-      beta: 0.9
-    }
-  }
-}

benchmark/vgg16/hdfs_piece_size_64.other

-use_rdma: true
-globalfs_conf {
-  hdfs_conf {
-    namenode: "hdfs://192.168.1.11:9000"
-  }
-}
-piece_size: 64
-data_part_num: 32
-max_data_id_length: 0
-train_conf {
-  model_save_snapshots_path: "/zhuyi/vgg_snapshots_buf"
-  num_of_batches_in_snapshot: 1000
-  staleness: 0
-  total_batch_num: 50
-  default_initializer_conf {
-    msra_conf {
-      variance_norm: kFanOut
-    }
-  }
-  batch_size: 256
-  l2: 0.0005
-  model_update_conf {
-    momentum_conf {
-      beta: 0.9
-    }
-  }
-}

benchmark/vgg16/hdfs_piece_size_8.other

-use_rdma: true
-globalfs_conf {
-  hdfs_conf {
-    namenode: "hdfs://192.168.1.11:9000"
-  }
-}
-piece_size: 8
-data_part_num: 8
-max_data_id_length: 0
-train_conf {
-  model_save_snapshots_path: "/zhuyi/vgg_snapshots_buf"
-  num_of_batches_in_snapshot: 1000
-  staleness: 0
-  total_batch_num: 50
-  default_initializer_conf {
-    msra_conf {
-      variance_norm: kFanOut
-    }
-  }
-  batch_size: 8
-  l2: 0.0005
-  model_update_conf {
-    momentum_conf {
-      beta: 0.9
-    }
-  }
-}

benchmark/vgg16/piece_size_16.other

-use_rdma: false
-globalfs_conf {
-  localfs_conf {
-  }
-}
-piece_size: 16
-data_part_num: 16
-max_data_id_length: 0
-train_conf {
-  model_save_snapshots_path: "/home/zhuyi/vgg_snapshots_buf"
-  num_of_batches_in_snapshot: 1000
-  staleness: 0
-  total_batch_num: 50
-  default_initializer_conf {
-    msra_conf {
-      variance_norm: kFanOut
-    }
-  }
-  batch_size: 256
-  l2: 0.0005
-  model_update_conf {
-    momentum_conf {
-      beta: 0.9
-    }
-  }
-}

benchmark/vgg16/piece_size_32.other

-use_rdma: false
-globalfs_conf {
-  localfs_conf {
-  }
-}
-piece_size: 32
-data_part_num: 32
-max_data_id_length: 0
-train_conf {
-  model_save_snapshots_path: "/home/zhuyi/vgg_snapshots_buf"
-  num_of_batches_in_snapshot: 1000
-  staleness: 0
-  total_batch_num: 50
-  default_initializer_conf {
-    msra_conf {
-      variance_norm: kFanOut
-    }
-  }
-  batch_size: 256
-  l2: 0.0005
-  model_update_conf {
-    momentum_conf {
-      beta: 0.9
-    }
-  }
-}

benchmark/vgg16/piece_size_64.other

-use_rdma: false
-globalfs_conf {
-  localfs_conf {
-  }
-}
-piece_size: 64
-data_part_num: 32
-max_data_id_length: 0
-train_conf {
-  model_save_snapshots_path: "/home/zhuyi/vgg_snapshots_buf"
-  num_of_batches_in_snapshot: 1000
-  staleness: 0
-  total_batch_num: 50
-  default_initializer_conf {
-    msra_conf {
-      variance_norm: kFanOut
-    }
-  }
-  batch_size: 256
-  l2: 0.0005
-  model_update_conf {
-    momentum_conf {
-      beta: 0.9
-    }
-  }
-}

benchmark/vgg16/report.md

-batch_size: 256
-
-gpu num | time (one batch) | piece size
-:-------| :-------------   | :---------
-1       | 3729ms           | 16
-2       | 1872ms           | 32
-4       | 950ms            | 64

benchmark/vgg16/train.net

-op {
-  name: "decode"
-  decode_ofrecord_conf {
-    data_dir: "/dataset/imagenet/train/32"
-    blob {
-      name: "encoded"
-      data_type: kFloat
-      shape {
-        dim: 224
-        dim: 224
-        dim: 3
-      }
-      encode_case {
-        jpeg {
-          preprocess {
-            resize {
-              width: 224
-              height: 224
-            }
-          }
-        }
-      }
-    }
-    blob {
-      name: "class/label"
-      data_type: kInt32
-      shape {
-      }
-      encode_case {
-        raw {
-        }
-      }
-    }
-  }
-}
-op {
-  name: "transpose"
-  transpose_conf {
-    in: "decode/encoded"
-    out: "out"
-    perm: 3
-    perm: 1
-    perm: 2
-  }
-}
-op {
-  name: "conv1_1"
-  conv_2d_conf {
-    in: "transpose/out"
-    out: "out"
-    filters: 64
-    use_bias: true
-    padding: "same"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-op {
-  name: "conv1_2"
-  conv_2d_conf {
-    in: "conv1_1/out"
-    out: "out"
-    filters: 64
-    use_bias: true
-    padding: "same"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-op {
-  name: "pool1"
-  max_pooling_2d_conf {
-    in: "conv1_2/out"
-    out: "out"
-    padding: "same"
-    data_format: "channels_first"
-    pool_size: 2
-    pool_size: 2
-    strides: 2
-    strides: 2
-  }
-}
-op {
-  name: "conv2_1"
-  conv_2d_conf {
-    in: "pool1/out"
-    out: "out"
-    filters: 128
-    use_bias: true
-    padding: "same"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-op {
-  name: "conv2_2"
-  conv_2d_conf {
-    in: "conv2_1/out"
-    out: "out"
-    filters: 128
-    use_bias: true
-    padding: "same"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-op {
-  name: "pool2"
-  max_pooling_2d_conf {
-    in: "conv2_2/out"
-    out: "out"
-    padding: "same"
-    data_format: "channels_first"
-    pool_size: 2
-    pool_size: 2
-    strides: 2
-    strides: 2
-  }
-}
-op {
-  name: "conv3_1"
-  conv_2d_conf {
-    in: "pool2/out"
-    out: "out"
-    filters: 256
-    use_bias: true
-    padding: "same"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-op {
-  name: "conv3_2"
-  conv_2d_conf {
-    in: "conv3_1/out"
-    out: "out"
-    filters: 256
-    use_bias: true
-    padding: "same"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-op {
-  name: "conv3_3"
-  conv_2d_conf {
-    in: "conv3_2/out"
-    out: "out"
-    filters: 256
-    use_bias: true
-    padding: "same"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-op {
-  name: "pool3"
-  max_pooling_2d_conf {
-    in: "conv3_3/out"
-    out: "out"
-    padding: "valid"
-    data_format: "channels_first"
-    pool_size: 2
-    pool_size: 2
-    strides: 2
-    strides: 2
-  }
-}
-op {
-  name: "conv4_1"
-  conv_2d_conf {
-    in: "pool3/out"
-    out: "out"
-    filters: 512
-    use_bias: true
-    padding: "same"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-op {
-  name: "conv4_2"
-  conv_2d_conf {
-    in: "conv4_1/out"
-    out: "out"
-    filters: 512
-    use_bias: true
-    padding: "same"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-op {
-  name: "conv4_3"
-  conv_2d_conf {
-    in: "conv4_2/out"
-    out: "out"
-    filters: 512
-    use_bias: true
-    padding: "same"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-op {
-  name: "pool4"
-  max_pooling_2d_conf {
-    in: "conv4_3/out"
-    out: "out"
-    padding: "valid"
-    data_format: "channels_first"
-    pool_size: 2
-    pool_size: 2
-    strides: 2
-    strides: 2
-  }
-}
-op {
-  name: "conv5_1"
-  conv_2d_conf {
-    in: "pool4/out"
-    out: "out"
-    filters: 512
-    use_bias: true
-    padding: "same"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-op {
-  name: "conv5_2"
-  conv_2d_conf {
-    in: "conv5_1/out"
-    out: "out"
-    filters: 512
-    use_bias: true
-    padding: "same"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-op {
-  name: "conv5_3"
-  conv_2d_conf {
-    in: "conv5_2/out"
-    out: "out"
-    filters: 512
-    use_bias: true
-    padding: "same"
-    data_format: "channels_first"
-    kernel_size: 3
-    kernel_size: 3
-    strides: 1
-    strides: 1
-    dilation_rate: 1
-    dilation_rate: 1
-    activation: kRelu
-    bias_initializer {
-      constant_conf {
-        value: 0
-      }
-    }
-  }
-}
-op {
-  name: "pool5"
-  max_pooling_2d_conf {
-    in: "conv5_3/out"
-    out: "out"
-    padding: "valid"
-    data_format: "channels_first"
-    pool_size: 2
-    pool_size: 2
-    strides: 2
-    strides: 2
-  }
-}
-# bias in fc are setted to 1.0 in some implementations
-op {
-  name: "fc6"
-  fully_connected_conf {
-    in: "pool5/out"
-    out: "out"
-    units: 4096
-    use_bias: true
-    activation: kRelu
-    weight_initializer {
-      random_normal_conf {
-        mean: 0.0
-        std: 0.01
-      }
-    }
-    bias_initializer {
-      constant_conf {
-        value: 0.0
-      }
-    }
-  }
-}
-op {
-  name: "drop6"
-  dropout_conf {
-    in: "fc6/out"
-    out: "out"
-    rate: 0.5
-  }
-}
-op {
-  name: "fc7"
-  fully_connected_conf {
-    in: "drop6/out"
-    out: "out"
-    units: 4096
-    use_bias: true
-    activation: kRelu
-    weight_initializer {
-      random_normal_conf {
-        mean: 0.0
-        std: 0.01
-      }
-    }
-    bias_initializer {
-      constant_conf {
-        value: 0.0
-      }
-    }
-  }
-}
-op {
-  name: "drop7"
-  dropout_conf {
-    in: "fc7/out"
-    out: "out"
-    rate: 0.5
-  }
-}
-op {
-  name: "fc8"
-  fully_connected_conf {
-    in: "drop7/out"
-    out: "out"
-    units: 1000
-    use_bias: true
-    weight_initializer {
-      random_normal_conf {
-        mean: 0.0
-        std: 0.001
-      }
-    }
-    bias_initializer {
-      constant_conf {
-        value: 0.0
-      }
-    }
-  }
-}
-op {
-  name: "softmax_loss"
-  sparse_softmax_cross_entropy_loss_conf {
-    prediction: "fc8/out"
-    label: "decode/class/label"
-    loss: "loss"
-  }
-}