caffe（conversation architecture for faster feature embeding）伯克利视觉和学习中心基于c++/cuda/python实现卷积神经网络框架
    前身decaf，作者贾扬清
    caffe优点
    1）速度快
    2）适合特征提取
    3）开源
    4）一套工具集，实现模型训练，预测，微调等
    5）代码组织良好
    caffe依赖包解析
    1）protobuffer 由谷歌开发实现内存与非易失存储介质交换协议接口
    2）boost c++ 准标准库
    3）GFLAGS在caffe起命令行参数解析的作用
    4）GLOG库谷歌开发记录应用程序日志的实用库
    5）BLAS 基本线性代数子程序
    6）HDF5 高效存储和分发科学数据的新型数据格式
    7）opencv 开源计算机视觉库
    8）LMDB 内存映射型数据库管理器
    9）sanppy 压缩和解压缩的c++库
    caffe实现
    1）数据生成lmdb（提高io利用率）
    2）模型文件
    train.prototxt

    1. name: "ssdJacintoNetV2"
    2. layer {
    3.   name: "data"
    4.   type: "AnnotatedData"
    5.   top: "data"
    6.   top: "label"
    7.   include {
    8.     phase: TRAIN
    9.   }
    10.   transform_param {
    11.     mirror: true
    12.     mean_value: 0
    13.     mean_value: 0
    14.     mean_value: 0
    15.     force_color: false
    16.     resize_param {
    17.       prob: 1
    18.       resize_mode: WARP
    19.       height: 160 #160
    20.       width: 320  #320
    21.       interp_mode: LINEAR
    22.       interp_mode: AREA
    23.       interp_mode: NEAREST
    24.       interp_mode: CUBIC
    25.       interp_mode: LANCZOS4
    26.     }
    27.     emit_constraint {
    28.       emit_type: CENTER
    29.     }
    30.     distort_param {
    31.       brightness_prob: 0.5
    32.       brightness_delta: 32
    33.       contrast_prob: 0.5
    34.       contrast_lower: 0.5
    35.       contrast_upper: 1.5
    36.       hue_prob: 0.5
    37.       hue_delta: 18
    38.       saturation_prob: 0.5
    39.       saturation_lower: 0.5
    40.       saturation_upper: 1.5
    41.       random_order_prob: 0.0
    42.     }
    43.     expand_param {
    44.       prob: 0.5
    45.       max_expand_ratio: 4.0
    46.     }
    47.   }
    48.   data_param {
    49.     source: "/home/caffe/examples/bsd_ti_5l_320/bsd_left/lmdb/bsd_left_trainval_lmdb"
    50.     batch_size: 16
    51.     backend: LMDB
    52.   }
    53.   annotated_data_param {
    54.     batch_sampler {
    55.       max_sample: 1
    56.       max_trials: 1
    57.     }
    58.     batch_sampler {
    59.       sampler {
    60.         min_scale: 0.3
    61.         max_scale: 1.0
    62.         min_aspect_ratio: 0.5
    63.         max_aspect_ratio: 2.0
    64.       }
    65.       sample_constraint {
    66.         min_jaccard_overlap: 0.1
    67.       }
    68.       max_sample: 1
    69.       max_trials: 50
    70.     }
    71.     batch_sampler {
    72.       sampler {
    73.         min_scale: 0.3
    74.         max_scale: 1.0
    75.         min_aspect_ratio: 0.5
    76.         max_aspect_ratio: 2.0
    77.       }
    78.       sample_constraint {
    79.         min_jaccard_overlap: 0.3
    80.       }
    81.       max_sample: 1
    82.       max_trials: 50
    83.     }
    84.     batch_sampler {
    85.       sampler {
    86.         min_scale: 0.3
    87.         max_scale: 1.0
    88.         min_aspect_ratio: 0.5
    89.         max_aspect_ratio: 2.0
    90.       }
    91.       sample_constraint {
    92.         min_jaccard_overlap: 0.5
    93.       }
    94.       max_sample: 1
    95.       max_trials: 50
    96.     }
    97.     batch_sampler {
    98.       sampler {
    99.         min_scale: 0.3
    100.         max_scale: 1.0
    101.         min_aspect_ratio: 0.5
    102.         max_aspect_ratio: 2.0
    103.       }
    104.       sample_constraint {
    105.         min_jaccard_overlap: 0.7
    106.       }
    107.       max_sample: 1
    108.       max_trials: 50
    109.     }
    110.     batch_sampler {
    111.       sampler {
    112.         min_scale: 0.3
    113.         max_scale: 1.0
    114.         min_aspect_ratio: 0.5
    115.         max_aspect_ratio: 2.0
    116.       }
    117.       sample_constraint {
    118.         min_jaccard_overlap: 0.9
    119.       }
    120.       max_sample: 1
    121.       max_trials: 50
    122.     }
    123.     batch_sampler {
    124.       sampler {
    125.         min_scale: 0.3
    126.         max_scale: 1.0
    127.         min_aspect_ratio: 0.5
    128.         max_aspect_ratio: 2.0
    129.       }
    130.       sample_constraint {
    131.         max_jaccard_overlap: 1.0
    132.       }
    133.       max_sample: 1
    134.       max_trials: 50
    135.     }
    136.     label_map_file: "./labelmap_bsd3class.prototxt"
    137.   }
    138. }
    140. layer {
    141.   name: "data"
    142.   type: "AnnotatedData"
    143.   top: "data"
    144.   top: "label"
    145.   include {
    146.     phase: TEST
    147.   }
    148.   transform_param {
    149.     mean_value: 0
    150.     mean_value: 0
    151.     mean_value: 0
    152.     force_color: false
    153.     resize_param {
    154.       prob: 1
    155.       resize_mode: WARP
    156.       height: 160 #160
    157.       width:  320 #320
    158.       interp_mode: LINEAR
    159.     }
    160.   }
    161.   data_param {
    162.     source: "/home/caffe/examples/bsd_ti_5l_320/bsd_left/lmdb/bsd_left_test_lmdb"
    163.     batch_size: 12
    164.     backend: LMDB
    165.   }
    166.   annotated_data_param {
    167.     batch_sampler {
    168.     }
    169.     label_map_file: "./labelmap_bsd3class.prototxt"
    170.   }
    171. }
    173. layer {
    174.   name: "data/bias"
    175.   type: "Bias"
    176.   bottom: "data"
    177.   top: "data/bias"
    178.   param {
    179.     lr_mult: 0
    180.     decay_mult: 0
    181.   }
    182.   bias_param {
    183.     filler {
    184.       type: "constant"
    185.       value: -128
    186.     }
    187.   }
    188. }
    190. layer {
    191.   name: "conv1a"
    192.   type: "Convolution"
    193.   bottom: "data/bias"
    194.   top: "conv1a"
    195.   param {
    196.     lr_mult: 1
    197.     decay_mult: 1
    198.   }
    199.   param {
    200.     lr_mult: 2
    201.     decay_mult: 0
    202.   }
    203.   convolution_param {
    204.     num_output: 32
    205.     bias_term: true
    206.     pad: 2
    207.     kernel_size: 5
    208.     group: 1
    209.     stride: 2
    210.     weight_filler {
    211.       type: "msra"
    212.     }
    213.     bias_filler {
    214.       type: "constant"
    215.       value: 0
    216.     }
    217.     dilation: 1
    218.   }
    219. }
    220. layer {
    221.   name: "conv1a/bn"
    222.   type: "BatchNorm"
    223.   bottom: "conv1a"
    224.   top: "conv1a"
    225.   batch_norm_param {
    226.     moving_average_fraction: 0.99
    227.     eps: 0.0001
    228.     #scale_bias: true
    229.   }
    230. }
    231. layer {
    232.   name: "conv1a/relu"
    233.   type: "ReLU"
    234.   bottom: "conv1a"
    235.   top: "conv1a"
    236. }
    237. layer {
    238.   name: "conv1b"
    239.   type: "Convolution"
    240.   bottom: "conv1a"
    241.   top: "conv1b"
    242.   param {
    243.     lr_mult: 1
    244.     decay_mult: 1
    245.   }
    246.   param {
    247.     lr_mult: 2
    248.     decay_mult: 0
    249.   }
    250.   convolution_param {
    251.     num_output: 32
    252.     bias_term: true
    253.     pad: 1
    254.     kernel_size: 3
    255.     group: 4
    256.     stride: 1
    257.     weight_filler {
    258.       type: "msra"
    259.     }
    260.     bias_filler {
    261.       type: "constant"
    262.       value: 0
    263.     }
    264.     dilation: 1
    265.   }
    266. }
    267. layer {
    268.   name: "conv1b/bn"
    269.   type: "BatchNorm"
    270.   bottom: "conv1b"
    271.   top: "conv1b"
    272.   batch_norm_param {
    273.     moving_average_fraction: 0.99
    274.     eps: 0.0001
    275.     #scale_bias: true
    276.   }
    277. }
    278. layer {
    279.   name: "conv1b/relu"
    280.   type: "ReLU"
    281.   bottom: "conv1b"
    282.   top: "conv1b"
    283. }
    284. layer {
    285.   name: "pool1"
    286.   type: "Pooling"
    287.   bottom: "conv1b"
    288.   top: "pool1"
    289.   pooling_param {
    290.     pool: MAX
    291.     kernel_size: 2
    292.     stride: 2
    293.   }
    294. }
    295. layer {
    296.   name: "res2a_branch2a"
    297.   type: "Convolution"
    298.   bottom: "pool1"
    299.   top: "res2a_branch2a"
    300.   param {
    301.     lr_mult: 1
    302.     decay_mult: 1
    303.   }
    304.   param {
    305.     lr_mult: 2
    306.     decay_mult: 0
    307.   }
    308.   convolution_param {
    309.     num_output: 64
    310.     bias_term: true
    311.     pad: 1
    312.     kernel_size: 3
    313.     group: 1
    314.     stride: 1
    315.     weight_filler {
    316.       type: "msra"
    317.     }
    318.     bias_filler {
    319.       type: "constant"
    320.       value: 0
    321.     }
    322.     dilation: 1
    323.   }
    324. }
    325. layer {
    326.   name: "res2a_branch2a/bn"
    327.   type: "BatchNorm"
    328.   bottom: "res2a_branch2a"
    329.   top: "res2a_branch2a"
    330.   batch_norm_param {
    331.     moving_average_fraction: 0.99
    332.     eps: 0.0001
    333.     #scale_bias: true
    334.   }
    335. }
    336. layer {
    337.   name: "res2a_branch2a/relu"
    338.   type: "ReLU"
    339.   bottom: "res2a_branch2a"
    340.   top: "res2a_branch2a"
    341. }
    342. layer {
    343.   name: "res2a_branch2b"
    344.   type: "Convolution"
    345.   bottom: "res2a_branch2a"
    346.   top: "res2a_branch2b"
    347.   param {
    348.     lr_mult: 1
    349.     decay_mult: 1
    350.   }
    351.   param {
    352.     lr_mult: 2
    353.     decay_mult: 0
    354.   }
    355.   convolution_param {
    356.     num_output: 64
    357.     bias_term: true
    358.     pad: 1
    359.     kernel_size: 3
    360.     group: 4
    361.     stride: 1
    362.     weight_filler {
    363.       type: "msra"
    364.     }
    365.     bias_filler {
    366.       type: "constant"
    367.       value: 0
    368.     }
    369.     dilation: 1
    370.   }
    371. }
    372. layer {
    373.   name: "res2a_branch2b/bn"
    374.   type: "BatchNorm"
    375.   bottom: "res2a_branch2b"
    376.   top: "res2a_branch2b"
    377.   batch_norm_param {
    378.     moving_average_fraction: 0.99
    379.     eps: 0.0001
    380.     #scale_bias: true
    381.   }
    382. }
    383. layer {
    384.   name: "res2a_branch2b/relu"
    385.   type: "ReLU"
    386.   bottom: "res2a_branch2b"
    387.   top: "res2a_branch2b"
    388. }
    389. layer {
    390.   name: "pool2"
    391.   type: "Pooling"
    392.   bottom: "res2a_branch2b"
    393.   top: "pool2"
    394.   pooling_param {
    395.     pool: MAX
    396.     kernel_size: 2
    397.     stride: 2
    398.   }
    399. }
    400. layer {
    401.   name: "res3a_branch2a"
    402.   type: "Convolution"
    403.   bottom: "pool2"
    404.   top: "res3a_branch2a"
    405.   param {
    406.     lr_mult: 1
    407.     decay_mult: 1
    408.   }
    409.   param {
    410.     lr_mult: 2
    411.     decay_mult: 0
    412.   }
    413.   convolution_param {
    414.     num_output: 128
    415.     bias_term: true
    416.     pad: 1
    417.     kernel_size: 3
    418.     group: 1
    419.     stride: 1
    420.     weight_filler {
    421.       type: "msra"
    422.     }
    423.     bias_filler {
    424.       type: "constant"
    425.       value: 0
    426.     }
    427.     dilation: 1
    428.   }
    429. }
    430. layer {
    431.   name: "res3a_branch2a/bn"
    432.   type: "BatchNorm"
    433.   bottom: "res3a_branch2a"
    434.   top: "res3a_branch2a"
    435.   batch_norm_param {
    436.     moving_average_fraction: 0.99
    437.     eps: 0.0001
    438.     #scale_bias: true
    439.   }
    440. }
    441. layer {
    442.   name: "res3a_branch2a/relu"
    443.   type: "ReLU"
    444.   bottom: "res3a_branch2a"
    445.   top: "res3a_branch2a"
    446. }
    447. layer {
    448.   name: "res3a_branch2b"
    449.   type: "Convolution"
    450.   bottom: "res3a_branch2a"
    451.   top: "res3a_branch2b"
    452.   param {
    453.     lr_mult: 1
    454.     decay_mult: 1
    455.   }
    456.   param {
    457.     lr_mult: 2
    458.     decay_mult: 0
    459.   }
    460.   convolution_param {
    461.     num_output: 128
    462.     bias_term: true
    463.     pad: 1
    464.     kernel_size: 3
    465.     group: 4
    466.     stride: 1
    467.     weight_filler {
    468.       type: "msra"
    469.     }
    470.     bias_filler {
    471.       type: "constant"
    472.       value: 0
    473.     }
    474.     dilation: 1
    475.   }
    476. }
    477. layer {
    478.   name: "res3a_branch2b/bn"
    479.   type: "BatchNorm"
    480.   bottom: "res3a_branch2b"
    481.   top: "res3a_branch2b"
    482.   batch_norm_param {
    483.     moving_average_fraction: 0.99
    484.     eps: 0.0001
    485.     #scale_bias: true
    486.   }
    487. }
    488. layer {
    489.   name: "res3a_branch2b/relu"
    490.   type: "ReLU"
    491.   bottom: "res3a_branch2b"
    492.   top: "res3a_branch2b"
    493. }
    494. layer {
    495.   name: "pool3"
    496.   type: "Pooling"
    497.   bottom: "res3a_branch2b"
    498.   top: "pool3"
    499.   pooling_param {
    500.     pool: MAX
    501.     kernel_size: 2
    502.     stride: 2
    503.   }
    504. }
    505. layer {
    506.   name: "res4a_branch2a"
    507.   type: "Convolution"
    508.   bottom: "pool3"
    509.   top: "res4a_branch2a"
    510.   param {
    511.     lr_mult: 1
    512.     decay_mult: 1
    513.   }
    514.   param {
    515.     lr_mult: 2
    516.     decay_mult: 0
    517.   }
    518.   convolution_param {
    519.     num_output: 256
    520.     bias_term: true
    521.     pad: 1
    522.     kernel_size: 3
    523.     group: 1
    524.     stride: 1
    525.     weight_filler {
    526.       type: "msra"
    527.     }
    528.     bias_filler {
    529.       type: "constant"
    530.       value: 0
    531.     }
    532.     dilation: 1
    533.   }
    534. }
    535. layer {
    536.   name: "res4a_branch2a/bn"
    537.   type: "BatchNorm"
    538.   bottom: "res4a_branch2a"
    539.   top: "res4a_branch2a"
    540.   batch_norm_param {
    541.     moving_average_fraction: 0.99
    542.     eps: 0.0001
    543.     #scale_bias: true
    544.   }
    545. }
    546. layer {
    547.   name: "res4a_branch2a/relu"
    548.   type: "ReLU"
    549.   bottom: "res4a_branch2a"
    550.   top: "res4a_branch2a"
    551. }
    552. layer {
    553.   name: "res4a_branch2b"
    554.   type: "Convolution"
    555.   bottom: "res4a_branch2a"
    556.   top: "res4a_branch2b"
    557.   param {
    558.     lr_mult: 1
    559.     decay_mult: 1
    560.   }
    561.   param {
    562.     lr_mult: 2
    563.     decay_mult: 0
    564.   }
    565.   convolution_param {
    566.     num_output: 256
    567.     bias_term: true
    568.     pad: 1
    569.     kernel_size: 3
    570.     group: 4
    571.     stride: 1
    572.     weight_filler {
    573.       type: "msra"
    574.     }
    575.     bias_filler {
    576.       type: "constant"
    577.       value: 0
    578.     }
    579.     dilation: 1
    580.   }
    581. }
    582. layer {
    583.   name: "res4a_branch2b/bn"
    584.   type: "BatchNorm"
    585.   bottom: "res4a_branch2b"
    586.   top: "res4a_branch2b"
    587.   batch_norm_param {
    588.     moving_average_fraction: 0.99
    589.     eps: 0.0001
    590.     #scale_bias: true
    591.   }
    592. }
    593. layer {
    594.   name: "res4a_branch2b/relu"
    595.   type: "ReLU"
    596.   bottom: "res4a_branch2b"
    597.   top: "res4a_branch2b"
    598. }
    599. layer {
    600.   name: "pool4"
    601.   type: "Pooling"
    602.   bottom: "res4a_branch2b"
    603.   top: "pool4"
    604.   pooling_param {
    605.     pool: MAX
    606.     kernel_size: 2
    607.     stride: 2
    608.   }
    609. }
    610. layer {
    611.   name: "res5a_branch2a"
    612.   type: "Convolution"
    613.   bottom: "pool4"
    614.   top: "res5a_branch2a"
    615.   param {
    616.     lr_mult: 1
    617.     decay_mult: 1
    618.   }
    619.   param {
    620.     lr_mult: 2
    621.     decay_mult: 0
    622.   }
    623.   convolution_param {
    624.     num_output: 512
    625.     bias_term: true
    626.     pad: 1
    627.     kernel_size: 3
    628.     group: 1
    629.     stride: 1
    630.     weight_filler {
    631.       type: "msra"
    632.     }
    633.     bias_filler {
    634.       type: "constant"
    635.       value: 0
    636.     }
    637.     dilation: 1
    638.   }
    639. }
    640. layer {
    641.   name: "res5a_branch2a/bn"
    642.   type: "BatchNorm"
    643.   bottom: "res5a_branch2a"
    644.   top: "res5a_branch2a"
    645.   batch_norm_param {
    646.     moving_average_fraction: 0.99
    647.     eps: 0.0001
    648.     #scale_bias: true
    649.   }
    650. }
    651. layer {
    652.   name: "res5a_branch2a/relu"
    653.   type: "ReLU"
    654.   bottom: "res5a_branch2a"
    655.   top: "res5a_branch2a"
    656. }
    657. layer {
    658.   name: "res5a_branch2b"
    659.   type: "Convolution"
    660.   bottom: "res5a_branch2a"
    661.   top: "res5a_branch2b"
    662.   param {
    663.     lr_mult: 1
    664.     decay_mult: 1
    665.   }
    666.   param {
    667.     lr_mult: 2
    668.     decay_mult: 0
    669.   }
    670.   convolution_param {
    671.     num_output: 512
    672.     bias_term: true
    673.     pad: 1
    674.     kernel_size: 3
    675.     group: 4
    676.     stride: 1
    677.     weight_filler {
    678.       type: "msra"
    679.     }
    680.     bias_filler {
    681.       type: "constant"
    682.       value: 0
    683.     }
    684.     dilation: 1
    685.   }
    686. }
    687. layer {
    688.   name: "res5a_branch2b/bn"
    689.   type: "BatchNorm"
    690.   bottom: "res5a_branch2b"
    691.   top: "res5a_branch2b"
    692.   batch_norm_param {
    693.     moving_average_fraction: 0.99
    694.     eps: 0.0001
    695.     #scale_bias: true
    696.   }
    697. }
    698. layer {
    699.   name: "res5a_branch2b/relu"
    700.   type: "ReLU"
    701.   bottom: "res5a_branch2b"
    702.   top: "res5a_branch2b"
    703. }
    704. layer {
    705.   name: "pool6"
    706.   type: "Pooling"
    707.   bottom: "res5a_branch2b"
    708.   top: "pool6"
    709.   pooling_param {
    710.     pool: MAX
    711.     kernel_size: 2
    712.     stride: 2
    713.     pad: 0
    714.   }
    715. }
    716. layer {
    717.   name: "pool7"
    718.   type: "Pooling"
    719.   bottom: "pool6"
    720.   top: "pool7"
    721.   pooling_param {
    722.     pool: MAX
    723.     kernel_size: 2
    724.     stride: 2
    725.     pad: 0
    726.   }
    727. }
    728. layer {
    729.   name: "pool8"
    730.   type: "Pooling"
    731.   bottom: "pool7"
    732.   top: "pool8"
    733.   pooling_param {
    734.     pool: MAX
    735.     kernel_size: 2
    736.     stride: 2
    737.     pad: 0
    738.   }
    739. }
    740. layer {
    741.   name: "ctx_output1"
    742.   type: "Convolution"
    743.   bottom: "res4a_branch2b"
    744.   top: "ctx_output1"
    745.   param {
    746.     lr_mult: 1
    747.     decay_mult: 1
    748.   }
    749.   param {
    750.     lr_mult: 2
    751.     decay_mult: 0
    752.   }
    753.   convolution_param {
    754.     num_output: 256
    755.     bias_term: true
    756.     pad: 0
    757.     kernel_size: 1
    758.     group: 1
    759.     stride: 1
    760.     weight_filler {
    761.       type: "msra"
    762.     }
    763.     bias_filler {
    764.       type: "constant"
    765.       value: 0
    766.     }
    767.     dilation: 1
    768.   }
    769. }
    770. layer {
    771.   name: "ctx_output1/bn"
    772.   type: "BatchNorm"
    773.   bottom: "ctx_output1"
    774.   top: "ctx_output1"
    775.   batch_norm_param {
    776.     moving_average_fraction: 0.99
    777.     eps: 0.0001
    778.     #scale_bias: true
    779.   }
    780. }
    781. layer {
    782.   name: "ctx_output1/relu"
    783.   type: "ReLU"
    784.   bottom: "ctx_output1"
    785.   top: "ctx_output1"
    786. }
    787. layer {
    788.   name: "ctx_output2"
    789.   type: "Convolution"
    790.   bottom: "res5a_branch2b"
    791.   top: "ctx_output2"
    792.   param {
    793.     lr_mult: 1
    794.     decay_mult: 1
    795.   }
    796.   param {
    797.     lr_mult: 2
    798.     decay_mult: 0
    799.   }
    800.   convolution_param {
    801.     num_output: 256
    802.     bias_term: true
    803.     pad: 0
    804.     kernel_size: 1
    805.     group: 1
    806.     stride: 1
    807.     weight_filler {
    808.       type: "msra"
    809.     }
    810.     bias_filler {
    811.       type: "constant"
    812.       value: 0
    813.     }
    814.     dilation: 1
    815.   }
    816. }
    817. layer {
    818.   name: "ctx_output2/bn"
    819.   type: "BatchNorm"
    820.   bottom: "ctx_output2"
    821.   top: "ctx_output2"
    822.   batch_norm_param {
    823.     moving_average_fraction: 0.99
    824.     eps: 0.0001
    825.     #scale_bias: true
    826.   }
    827. }
    828. layer {
    829.   name: "ctx_output2/relu"
    830.   type: "ReLU"
    831.   bottom: "ctx_output2"
    832.   top: "ctx_output2"
    833. }
    834. layer {
    835.   name: "ctx_output3"
    836.   type: "Convolution"
    837.   bottom: "pool6"
    838.   top: "ctx_output3"
    839.   param {
    840.     lr_mult: 1
    841.     decay_mult: 1
    842.   }
    843.   param {
    844.     lr_mult: 2
    845.     decay_mult: 0
    846.   }
    847.   convolution_param {
    848.     num_output: 256
    849.     bias_term: true
    850.     pad: 0
    851.     kernel_size: 1
    852.     group: 1
    853.     stride: 1
    854.     weight_filler {
    855.       type: "msra"
    856.     }
    857.     bias_filler {
    858.       type: "constant"
    859.       value: 0
    860.     }
    861.     dilation: 1
    862.   }
    863. }
    864. layer {
    865.   name: "ctx_output3/bn"
    866.   type: "BatchNorm"
    867.   bottom: "ctx_output3"
    868.   top: "ctx_output3"
    869.   batch_norm_param {
    870.     moving_average_fraction: 0.99
    871.     eps: 0.0001
    872.     #scale_bias: true
    873.   }
    874. }
    875. layer {
    876.   name: "ctx_output3/relu"
    877.   type: "ReLU"
    878.   bottom: "ctx_output3"
    879.   top: "ctx_output3"
    880. }
    881. layer {
    882.   name: "ctx_output4"
    883.   type: "Convolution"
    884.   bottom: "pool7"
    885.   top: "ctx_output4"
    886.   param {
    887.     lr_mult: 1
    888.     decay_mult: 1
    889.   }
    890.   param {
    891.     lr_mult: 2
    892.     decay_mult: 0
    893.   }
    894.   convolution_param {
    895.     num_output: 256
    896.     bias_term: true
    897.     pad: 0
    898.     kernel_size: 1
    899.     group: 1
    900.     stride: 1
    901.     weight_filler {
    902.       type: "msra"
    903.     }
    904.     bias_filler {
    905.       type: "constant"
    906.       value: 0
    907.     }
    908.     dilation: 1
    909.   }
    910. }
    911. layer {
    912.   name: "ctx_output4/bn"
    913.   type: "BatchNorm"
    914.   bottom: "ctx_output4"
    915.   top: "ctx_output4"
    916.   batch_norm_param {
    917.     moving_average_fraction: 0.99
    918.     eps: 0.0001
    919.     #scale_bias: true
    920.   }
    921. }
    922. layer {
    923.   name: "ctx_output4/relu"
    924.   type: "ReLU"
    925.   bottom: "ctx_output4"
    926.   top: "ctx_output4"
    927. }
    928. layer {
    929.   name: "ctx_output5"
    930.   type: "Convolution"
    931.   bottom: "pool8"
    932.   top: "ctx_output5"
    933.   param {
    934.     lr_mult: 1
    935.     decay_mult: 1
    936.   }
    937.   param {
    938.     lr_mult: 2
    939.     decay_mult: 0
    940.   }
    941.   convolution_param {
    942.     num_output: 256
    943.     bias_term: true
    944.     pad: 0
    945.     kernel_size: 1
    946.     group: 1
    947.     stride: 1
    948.     weight_filler {
    949.       type: "msra"
    950.     }
    951.     bias_filler {
    952.       type: "constant"
    953.       value: 0
    954.     }
    955.     dilation: 1
    956.   }
    957. }
    958. layer {
    959.   name: "ctx_output5/bn"
    960.   type: "BatchNorm"
    961.   bottom: "ctx_output5"
    962.   top: "ctx_output5"
    963.   batch_norm_param {
    964.     moving_average_fraction: 0.99
    965.     eps: 0.0001
    966.     #scale_bias: true
    967.   }
    968. }
    969. layer {
    970.   name: "ctx_output5/relu"
    971.   type: "ReLU"
    972.   bottom: "ctx_output5"
    973.   top: "ctx_output5"
    974. }
    975. layer {
    976.   name: "ctx_output1/relu_mbox_loc"
    977.   type: "Convolution"
    978.   bottom: "ctx_output1"
    979.   top: "ctx_output1/relu_mbox_loc"
    980.   param {
    981.     lr_mult: 1
    982.     decay_mult: 1
    983.   }
    984.   param {
    985.     lr_mult: 2
    986.     decay_mult: 0
    987.   }
    988.   convolution_param {
    989.     num_output: 16
    990.     bias_term: true
    991.     pad: 0
    992.     kernel_size: 1
    993.     group: 1
    994.     stride: 1
    995.     weight_filler {
    996.       type: "msra"
    997.     }
    998.     bias_filler {
    999.       type: "constant"
    1000.       value: 0
    1001.     }
    1002.     dilation: 1
    1003.   }
    1004. }
    1005. layer {
    1006.   name: "ctx_output1/relu_mbox_loc_perm"
    1007.   type: "Permute"
    1008.   bottom: "ctx_output1/relu_mbox_loc"
    1009.   top: "ctx_output1/relu_mbox_loc_perm"
    1010.   permute_param {
    1011.     order: 0
    1012.     order: 2
    1013.     order: 3
    1014.     order: 1
    1015.   }
    1016. }
    1017. layer {
    1018.   name: "ctx_output1/relu_mbox_loc_flat"
    1019.   type: "Flatten"
    1020.   bottom: "ctx_output1/relu_mbox_loc_perm"
    1021.   top: "ctx_output1/relu_mbox_loc_flat"
    1022.   flatten_param {
    1023.     axis: 1
    1024.   }
    1025. }
    1026. layer {
    1027.   name: "ctx_output1/relu_mbox_conf"
    1028.   type: "Convolution"
    1029.   bottom: "ctx_output1"
    1030.   top: "ctx_output1/relu_mbox_conf"
    1031.   param {
    1032.     lr_mult: 1
    1033.     decay_mult: 1
    1034.   }
    1035.   param {
    1036.     lr_mult: 2
    1037.     decay_mult: 0
    1038.   }
    1039.   convolution_param {
    1040.     num_output: 16#84
    1041.     bias_term: true
    1042.     pad: 0
    1043.     kernel_size: 1
    1044.     group: 1
    1045.     stride: 1
    1046.     weight_filler {
    1047.       type: "msra"
    1048.     }
    1049.     bias_filler {
    1050.       type: "constant"
    1051.       value: 0
    1052.     }
    1053.     dilation: 1
    1054.   }
    1055. }
    1056. layer {
    1057.   name: "ctx_output1/relu_mbox_conf_perm"
    1058.   type: "Permute"
    1059.   bottom: "ctx_output1/relu_mbox_conf"
    1060.   top: "ctx_output1/relu_mbox_conf_perm"
    1061.   permute_param {
    1062.     order: 0
    1063.     order: 2
    1064.     order: 3
    1065.     order: 1
    1066.   }
    1067. }
    1068. layer {
    1069.   name: "ctx_output1/relu_mbox_conf_flat"
    1070.   type: "Flatten"
    1071.   bottom: "ctx_output1/relu_mbox_conf_perm"
    1072.   top: "ctx_output1/relu_mbox_conf_flat"
    1073.   flatten_param {
    1074.     axis: 1
    1075.   }
    1076. }
    1077. layer {
    1078.   name: "ctx_output1/relu_mbox_priorbox"
    1079.   type: "PriorBox"
    1080.   bottom: "ctx_output1"
    1081.   bottom: "data"
    1082.   top: "ctx_output1/relu_mbox_priorbox"
    1083.   prior_box_param {
    1084.     min_size: 24 #14.72
    1085.     max_size: 50.24 #36.8
    1086.     aspect_ratio: 2
    1087.     flip: true
    1088.     clip: false
    1089.     variance: 0.1
    1090.     variance: 0.1
    1091.     variance: 0.2
    1092.     variance: 0.2
    1093.     offset: 0.5
    1094.   }
    1095. }
    1096. layer {
    1097.   name: "ctx_output2/relu_mbox_loc"
    1098.   type: "Convolution"
    1099.   bottom: "ctx_output2"
    1100.   top: "ctx_output2/relu_mbox_loc"
    1101.   param {
    1102.     lr_mult: 1
    1103.     decay_mult: 1
    1104.   }
    1105.   param {
    1106.     lr_mult: 2
    1107.     decay_mult: 0
    1108.   }
    1109.   convolution_param {
    1110.     num_output: 24
    1111.     bias_term: true
    1112.     pad: 0
    1113.     kernel_size: 1
    1114.     group: 1
    1115.     stride: 1
    1116.     weight_filler {
    1117.       type: "msra"
    1118.     }
    1119.     bias_filler {
    1120.       type: "constant"
    1121.       value: 0
    1122.     }
    1123.     dilation: 1
    1124.   }
    1125. }
    1126. layer {
    1127.   name: "ctx_output2/relu_mbox_loc_perm"
    1128.   type: "Permute"
    1129.   bottom: "ctx_output2/relu_mbox_loc"
    1130.   top: "ctx_output2/relu_mbox_loc_perm"
    1131.   permute_param {
    1132.     order: 0
    1133.     order: 2
    1134.     order: 3
    1135.     order: 1
    1136.   }
    1137. }
    1138. layer {
    1139.   name: "ctx_output2/relu_mbox_loc_flat"
    1140.   type: "Flatten"
    1141.   bottom: "ctx_output2/relu_mbox_loc_perm"
    1142.   top: "ctx_output2/relu_mbox_loc_flat"
    1143.   flatten_param {
    1144.     axis: 1
    1145.   }
    1146. }
    1147. layer {
    1148.   name: "ctx_output2/relu_mbox_conf"
    1149.   type: "Convolution"
    1150.   bottom: "ctx_output2"
    1151.   top: "ctx_output2/relu_mbox_conf"
    1152.   param {
    1153.     lr_mult: 1
    1154.     decay_mult: 1
    1155.   }
    1156.   param {
    1157.     lr_mult: 2
    1158.     decay_mult: 0
    1159.   }
    1160.   convolution_param {
    1161.     num_output: 24#126
    1162.     bias_term: true
    1163.     pad: 0
    1164.     kernel_size: 1
    1165.     group: 1
    1166.     stride: 1
    1167.     weight_filler {
    1168.       type: "msra"
    1169.     }
    1170.     bias_filler {
    1171.       type: "constant"
    1172.       value: 0
    1173.     }
    1174.     dilation: 1
    1175.   }
    1176. }
    1177. layer {
    1178.   name: "ctx_output2/relu_mbox_conf_perm"
    1179.   type: "Permute"
    1180.   bottom: "ctx_output2/relu_mbox_conf"
    1181.   top: "ctx_output2/relu_mbox_conf_perm"
    1182.   permute_param {
    1183.     order: 0
    1184.     order: 2
    1185.     order: 3
    1186.     order: 1
    1187.   }
    1188. }
    1189. layer {
    1190.   name: "ctx_output2/relu_mbox_conf_flat"
    1191.   type: "Flatten"
    1192.   bottom: "ctx_output2/relu_mbox_conf_perm"
    1193.   top: "ctx_output2/relu_mbox_conf_flat"
    1194.   flatten_param {
    1195.     axis: 1
    1196.   }
    1197. }
    1198. layer {
    1199.   name: "ctx_output2/relu_mbox_priorbox"
    1200.   type: "PriorBox"
    1201.   bottom: "ctx_output2"
    1202.   bottom: "data"
    1203.   top: "ctx_output2/relu_mbox_priorbox"
    1204.   prior_box_param {
    1205.     min_size: 50.24 #36.8
    1206.     max_size: 76.48 #110.4
    1207.     aspect_ratio: 2
    1208.     aspect_ratio: 3
    1209.     flip: true
    1210.     clip: false
    1211.     variance: 0.1
    1212.     variance: 0.1
    1213.     variance: 0.2
    1214.     variance: 0.2
    1215.     offset: 0.5
    1216.   }
    1217. }
    1218. layer {
    1219.   name: "ctx_output3/relu_mbox_loc"
    1220.   type: "Convolution"
    1221.   bottom: "ctx_output3"
    1222.   top: "ctx_output3/relu_mbox_loc"
    1223.   param {
    1224.     lr_mult: 1
    1225.     decay_mult: 1
    1226.   }
    1227.   param {
    1228.     lr_mult: 2
    1229.     decay_mult: 0
    1230.   }
    1231.   convolution_param {
    1232.     num_output: 24
    1233.     bias_term: true
    1234.     pad: 0
    1235.     kernel_size: 1
    1236.     group: 1
    1237.     stride: 1
    1238.     weight_filler {
    1239.       type: "msra"
    1240.     }
    1241.     bias_filler {
    1242.       type: "constant"
    1243.       value: 0
    1244.     }
    1245.     dilation: 1
    1246.   }
    1247. }
    1248. layer {
    1249.   name: "ctx_output3/relu_mbox_loc_perm"
    1250.   type: "Permute"
    1251.   bottom: "ctx_output3/relu_mbox_loc"
    1252.   top: "ctx_output3/relu_mbox_loc_perm"
    1253.   permute_param {
    1254.     order: 0
    1255.     order: 2
    1256.     order: 3
    1257.     order: 1
    1258.   }
    1259. }
    1260. layer {
    1261.   name: "ctx_output3/relu_mbox_loc_flat"
    1262.   type: "Flatten"
    1263.   bottom: "ctx_output3/relu_mbox_loc_perm"
    1264.   top: "ctx_output3/relu_mbox_loc_flat"
    1265.   flatten_param {
    1266.     axis: 1
    1267.   }
    1268. }
    1269. layer {
    1270.   name: "ctx_output3/relu_mbox_conf"
    1271.   type: "Convolution"
    1272.   bottom: "ctx_output3"
    1273.   top: "ctx_output3/relu_mbox_conf"
    1274.   param {
    1275.     lr_mult: 1
    1276.     decay_mult: 1
    1277.   }
    1278.   param {
    1279.     lr_mult: 2
    1280.     decay_mult: 0
    1281.   }
    1282.   convolution_param {
    1283.     num_output: 24#126
    1284.     bias_term: true
    1285.     pad: 0
    1286.     kernel_size: 1
    1287.     group: 1
    1288.     stride: 1
    1289.     weight_filler {
    1290.       type: "msra"
    1291.     }
    1292.     bias_filler {
    1293.       type: "constant"
    1294.       value: 0
    1295.     }
    1296.     dilation: 1
    1297.   }
    1298. }
    1299. layer {
    1300.   name: "ctx_output3/relu_mbox_conf_perm"
    1301.   type: "Permute"
    1302.   bottom: "ctx_output3/relu_mbox_conf"
    1303.   top: "ctx_output3/relu_mbox_conf_perm"
    1304.   permute_param {
    1305.     order: 0
    1306.     order: 2
    1307.     order: 3
    1308.     order: 1
    1309.   }
    1310. }
    1311. layer {
    1312.   name: "ctx_output3/relu_mbox_conf_flat"
    1313.   type: "Flatten"
    1314.   bottom: "ctx_output3/relu_mbox_conf_perm"
    1315.   top: "ctx_output3/relu_mbox_conf_flat"
    1316.   flatten_param {
    1317.     axis: 1
    1318.   }
    1319. }
    1320. layer {
    1321.   name: "ctx_output3/relu_mbox_priorbox"
    1322.   type: "PriorBox"
    1323.   bottom: "ctx_output3"
    1324.   bottom: "data"
    1325.   top: "ctx_output3/relu_mbox_priorbox"
    1326.   prior_box_param {
    1327.     min_size: 76.48 #110.4
    1328.     max_size: 102.72 #184.0
    1329.     aspect_ratio: 2
    1330.     aspect_ratio: 3
    1331.     flip: true
    1332.     clip: false
    1333.     variance: 0.1
    1334.     variance: 0.1
    1335.     variance: 0.2
    1336.     variance: 0.2
    1337.     offset: 0.5
    1338.   }
    1339. }
    1340. layer {
    1341.   name: "ctx_output4/relu_mbox_loc"
    1342.   type: "Convolution"
    1343.   bottom: "ctx_output4"
    1344.   top: "ctx_output4/relu_mbox_loc"
    1345.   param {
    1346.     lr_mult: 1
    1347.     decay_mult: 1
    1348.   }
    1349.   param {
    1350.     lr_mult: 2
    1351.     decay_mult: 0
    1352.   }
    1353.   convolution_param {
    1354.     num_output: 24
    1355.     bias_term: true
    1356.     pad: 0
    1357.     kernel_size: 1
    1358.     group: 1
    1359.     stride: 1
    1360.     weight_filler {
    1361.       type: "msra"
    1362.     }
    1363.     bias_filler {
    1364.       type: "constant"
    1365.       value: 0
    1366.     }
    1367.     dilation: 1
    1368.   }
    1369. }
    1370. layer {
    1371.   name: "ctx_output4/relu_mbox_loc_perm"
    1372.   type: "Permute"
    1373.   bottom: "ctx_output4/relu_mbox_loc"
    1374.   top: "ctx_output4/relu_mbox_loc_perm"
    1375.   permute_param {
    1376.     order: 0
    1377.     order: 2
    1378.     order: 3
    1379.     order: 1
    1380.   }
    1381. }
    1382. layer {
    1383.   name: "ctx_output4/relu_mbox_loc_flat"
    1384.   type: "Flatten"
    1385.   bottom: "ctx_output4/relu_mbox_loc_perm"
    1386.   top: "ctx_output4/relu_mbox_loc_flat"
    1387.   flatten_param {
    1388.     axis: 1
    1389.   }
    1390. }
    1391. layer {
    1392.   name: "ctx_output4/relu_mbox_conf"
    1393.   type: "Convolution"
    1394.   bottom: "ctx_output4"
    1395.   top: "ctx_output4/relu_mbox_conf"
    1396.   param {
    1397.     lr_mult: 1
    1398.     decay_mult: 1
    1399.   }
    1400.   param {
    1401.     lr_mult: 2
    1402.     decay_mult: 0
    1403.   }
    1404.   convolution_param {
    1405.     num_output: 24#126
    1406.     bias_term: true
    1407.     pad: 0
    1408.     kernel_size: 1
    1409.     group: 1
    1410.     stride: 1
    1411.     weight_filler {
    1412.       type: "msra"
    1413.     }
    1414.     bias_filler {
    1415.       type: "constant"
    1416.       value: 0
    1417.     }
    1418.     dilation: 1
    1419.   }
    1420. }
    1421. layer {
    1422.   name: "ctx_output4/relu_mbox_conf_perm"
    1423.   type: "Permute"
    1424.   bottom: "ctx_output4/relu_mbox_conf"
    1425.   top: "ctx_output4/relu_mbox_conf_perm"
    1426.   permute_param {
    1427.     order: 0
    1428.     order: 2
    1429.     order: 3
    1430.     order: 1
    1431.   }
    1432. }
    1433. layer {
    1434.   name: "ctx_output4/relu_mbox_conf_flat"
    1435.   type: "Flatten"
    1436.   bottom: "ctx_output4/relu_mbox_conf_perm"
    1437.   top: "ctx_output4/relu_mbox_conf_flat"
    1438.   flatten_param {
    1439.     axis: 1
    1440.   }
    1441. }
    1442. layer {
    1443.   name: "ctx_output4/relu_mbox_priorbox"
    1444.   type: "PriorBox"
    1445.   bottom: "ctx_output4"
    1446.   bottom: "data"
    1447.   top: "ctx_output4/relu_mbox_priorbox"
    1448.   prior_box_param {
    1449.     min_size: 102.72 #184.0
    1450.     max_size: 128.96 #257.6
    1451.     aspect_ratio: 2
    1452.     aspect_ratio: 3
    1453.     flip: true
    1454.     clip: false
    1455.     variance: 0.1
    1456.     variance: 0.1
    1457.     variance: 0.2
    1458.     variance: 0.2
    1459.     offset: 0.5
    1460.   }
    1461. }
    1462. layer {
    1463.   name: "ctx_output5/relu_mbox_loc"
    1464.   type: "Convolution"
    1465.   bottom: "ctx_output5"
    1466.   top: "ctx_output5/relu_mbox_loc"
    1467.   param {
    1468.     lr_mult: 1
    1469.     decay_mult: 1
    1470.   }
    1471.   param {
    1472.     lr_mult: 2
    1473.     decay_mult: 0
    1474.   }
    1475.   convolution_param {
    1476.     num_output: 16
    1477.     bias_term: true
    1478.     pad: 0
    1479.     kernel_size: 1
    1480.     group: 1
    1481.     stride: 1
    1482.     weight_filler {
    1483.       type: "msra"
    1484.     }
    1485.     bias_filler {
    1486.       type: "constant"
    1487.       value: 0
    1488.     }
    1489.     dilation: 1
    1490.   }
    1491. }
    1492. layer {
    1493.   name: "ctx_output5/relu_mbox_loc_perm"
    1494.   type: "Permute"
    1495.   bottom: "ctx_output5/relu_mbox_loc"
    1496.   top: "ctx_output5/relu_mbox_loc_perm"
    1497.   permute_param {
    1498.     order: 0
    1499.     order: 2
    1500.     order: 3
    1501.     order: 1
    1502.   }
    1503. }
    1504. layer {
    1505.   name: "ctx_output5/relu_mbox_loc_flat"
    1506.   type: "Flatten"
    1507.   bottom: "ctx_output5/relu_mbox_loc_perm"
    1508.   top: "ctx_output5/relu_mbox_loc_flat"
    1509.   flatten_param {
    1510.     axis: 1
    1511.   }
    1512. }
    1513. layer {
    1514.   name: "ctx_output5/relu_mbox_conf"
    1515.   type: "Convolution"
    1516.   bottom: "ctx_output5"
    1517.   top: "ctx_output5/relu_mbox_conf"
    1518.   param {
    1519.     lr_mult: 1
    1520.     decay_mult: 1
    1521.   }
    1522.   param {
    1523.     lr_mult: 2
    1524.     decay_mult: 0
    1525.   }
    1526.   convolution_param {
    1527.     num_output: 16#84
    1528.     bias_term: true
    1529.     pad: 0
    1530.     kernel_size: 1
    1531.     group: 1
    1532.     stride: 1
    1533.     weight_filler {
    1534.       type: "msra"
    1535.     }
    1536.     bias_filler {
    1537.       type: "constant"
    1538.       value: 0
    1539.     }
    1540.     dilation: 1
    1541.   }
    1542. }
    1543. layer {
    1544.   name: "ctx_output5/relu_mbox_conf_perm"
    1545.   type: "Permute"
    1546.   bottom: "ctx_output5/relu_mbox_conf"
    1547.   top: "ctx_output5/relu_mbox_conf_perm"
    1548.   permute_param {
    1549.     order: 0
    1550.     order: 2
    1551.     order: 3
    1552.     order: 1
    1553.   }
    1554. }
    1555. layer {
    1556.   name: "ctx_output5/relu_mbox_conf_flat"
    1557.   type: "Flatten"
    1558.   bottom: "ctx_output5/relu_mbox_conf_perm"
    1559.   top: "ctx_output5/relu_mbox_conf_flat"
    1560.   flatten_param {
    1561.     axis: 1
    1562.   }
    1563. }
    1564. layer {
    1565.   name: "ctx_output5/relu_mbox_priorbox"
    1566.   type: "PriorBox"
    1567.   bottom: "ctx_output5"
    1568.   bottom: "data"
    1569.   top: "ctx_output5/relu_mbox_priorbox"
    1570.   prior_box_param {
    1571.     min_size: 128.96 #257.6
    1572.     max_size: 160    #331.2
    1573.     aspect_ratio: 2
    1574.     flip: true
    1575.     clip: false
    1576.     variance: 0.1
    1577.     variance: 0.1
    1578.     variance: 0.2
    1579.     variance: 0.2
    1580.     offset: 0.5
    1581.   }
    1582. }
    1583. layer {
    1584.   name: "mbox_loc"
    1585.   type: "Concat"
    1586.   bottom: "ctx_output1/relu_mbox_loc_flat"
    1587.   bottom: "ctx_output2/relu_mbox_loc_flat"
    1588.   bottom: "ctx_output3/relu_mbox_loc_flat"
    1589.   bottom: "ctx_output4/relu_mbox_loc_flat"
    1590.   bottom: "ctx_output5/relu_mbox_loc_flat"
    1591.   top: "mbox_loc"
    1592.   concat_param {
    1593.     axis: 1
    1594.   }
    1595. }
    1596. layer {
    1597.   name: "mbox_conf"
    1598.   type: "Concat"
    1599.   bottom: "ctx_output1/relu_mbox_conf_flat"
    1600.   bottom: "ctx_output2/relu_mbox_conf_flat"
    1601.   bottom: "ctx_output3/relu_mbox_conf_flat"
    1602.   bottom: "ctx_output4/relu_mbox_conf_flat"
    1603.   bottom: "ctx_output5/relu_mbox_conf_flat"
    1604.   top: "mbox_conf"
    1605.   concat_param {
    1606.     axis: 1
    1607.   }
    1608. }
    1609. layer {
    1610.   name: "mbox_priorbox"
    1611.   type: "Concat"
    1612.   bottom: "ctx_output1/relu_mbox_priorbox"
    1613.   bottom: "ctx_output2/relu_mbox_priorbox"
    1614.   bottom: "ctx_output3/relu_mbox_priorbox"
    1615.   bottom: "ctx_output4/relu_mbox_priorbox"
    1616.   bottom: "ctx_output5/relu_mbox_priorbox"
    1617.   top: "mbox_priorbox"
    1618.   concat_param {
    1619.     axis: 2
    1620.   }
    1621. }
    1622. layer {
    1623.   name: "mbox_loss"
    1624.   type: "MultiBoxLoss"
    1625.   bottom: "mbox_loc"
    1626.   bottom: "mbox_conf"
    1627.   bottom: "mbox_priorbox"
    1628.   bottom: "label"
    1629.   top: "mbox_loss"
    1630.   include {
    1631.     phase: TRAIN
    1632.   }
    1633.   propagate_down: true
    1634.   propagate_down: true
    1635.   propagate_down: false
    1636.   propagate_down: false
    1637.   loss_param {
    1638.     normalization: VALID
    1639.   }
    1640.   multibox_loss_param {
    1641.     loc_loss_type: SMOOTH_L1
    1642.     conf_loss_type: SOFTMAX
    1643.     loc_weight: 1.0
    1644.     num_classes: 4  
    1645.     share_location: true
    1646.     match_type: PER_PREDICTION
    1647.     overlap_threshold: 0.5
    1648.     use_prior_for_matching: true
    1649.     background_label_id: 0
    1650.     use_difficult_gt: true
    1651.     neg_pos_ratio: 3.0
    1652.     neg_overlap: 0.5
    1653.     code_type: CENTER_SIZE
    1654.     ignore_cross_boundary_bbox: false
    1655.     mining_type: MAX_NEGATIVE
    1656.   }
    1657. }
    1658. layer {
    1659.   name: "mbox_conf_reshape"
    1660.   type: "Reshape"
    1661.   bottom: "mbox_conf"
    1662.   top: "mbox_conf_reshape"
    1663.   reshape_param {
    1664.     shape {
    1665.       dim: 0
    1666.       dim: -1
    1667.       dim: 4
    1668.     }
    1669.   }
    1670. }
    1671. layer {
    1672.   name: "mbox_conf_softmax"
    1673.   type: "Softmax"
    1674.   bottom: "mbox_conf_reshape"
    1675.   top: "mbox_conf_softmax"
    1676.   softmax_param {
    1677.     axis: 2
    1678.   }
    1679. }
    1680. layer {
    1681.   name: "mbox_conf_flatten"
    1682.   type: "Flatten"
    1683.   bottom: "mbox_conf_softmax"
    1684.   top: "mbox_conf_flatten"
    1685.   flatten_param {
    1686.     axis: 1
    1687.   }
    1688. }
    1689. layer {
    1690.   name: "detection_out"
    1691.   type: "DetectionOutput"
    1692.   bottom: "mbox_loc"
    1693.   bottom: "mbox_conf_flatten"
    1694.   bottom: "mbox_priorbox"
    1695.   top: "detection_out"
    1696.   include {
    1697.     phase: TEST
    1698.   }
    1699.   detection_output_param {
    1700.     num_classes: 4
    1701.     share_location: true
    1702.     background_label_id: 0
    1703.     nms_param {
    1704.       nms_threshold: 0.45
    1705.       top_k: 400
    1706.     }
    1707.     code_type: CENTER_SIZE
    1708.     keep_top_k: 200
    1709.     confidence_threshold: 0.01
    1710.   }
    1711. }
    1712. layer {
    1713.   name: "detection_eval"
    1714.   type: "DetectionEvaluate"
    1715.   bottom: "detection_out"
    1716.   bottom: "label"
    1717.   top: "detection_eval"
    1718.   include {
    1719.     phase: TEST
    1720.   }
    1721.   detection_evaluate_param {
    1722.     num_classes: 4
    1723.     background_label_id: 0
    1724.     overlap_threshold: 0.5
    1725.     evaluate_difficult_gt: false
    1726.     #name_size_file: "/home/caffe/examples/bsd_ti_5l_320/bsd_r_mini/VOC2007/test_name_size.txt"
    1727.   }
    1728. }

    3) 训练超参数

    net: "ti_bsd_5little_200_400_4pr.prototxt" ##########
test_iter: 10 ##########
test_interval: 100 ##########
base_lr: 0.01 #0.0005
display: 100
max_iter: 100000 ##########
lr_policy: "multistep"
gamma: 0.1
power: 1.0
momentum: 0.9
weight_decay: 0.0001
snapshot: 10000
snapshot_prefix: "snapshot/4pr" ###########
solver_mode: GPU
debug_info: false
snapshot_after_train: true
test_initialization: false
average_loss: 10
stepvalue: 60000
stepvalue: 90000
stepvalue: 300000
iter_size: 2
type: "SGD"
eval_type: "detection"
ap_version: "11point"
show_per_class_result: true

    4）模型训练
    sh train.sh

    #!/bin/sh
../../build/tools/caffe train -solver="solver_train.prototxt" -gpu=0 -snapshot="./snapshot/4pr_iter_100000.solverstate" -gpu=0  
2>&1 | tee ./bsd.log

    5）模型预测 ./build/tools/caffe.bin test -model ./test.prototxt -weights ./lenet_iter_1000.caffemodel -iterations 100 (bachsize*iter=number of test images)
    deploy文件

    name: "ssdJacintoNetV2"
input: "data"
input_shape{
      dim=1
      dim=3
      dim: 160 #160
      dim: 320  #320
    }

layer {
  name: "data/bias"
  type: "Bias"
  bottom: "data"
  top: "data/bias"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  bias_param {
    filler {
      type: "constant"
      value: -128
    }
  }
}

layer {
  name: "conv1a"
  type: "Convolution"
  bottom: "data/bias"
  top: "conv1a"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 32
    bias_term: true
    pad: 2
    kernel_size: 5
    group: 1
    stride: 2
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "conv1a/bn"
  type: "BatchNorm"
  bottom: "conv1a"
  top: "conv1a"
  batch_norm_param {
    moving_average_fraction: 0.99
    eps: 0.0001
    #scale_bias: true
  }
}
layer {
  name: "conv1a/relu"
  type: "ReLU"
  bottom: "conv1a"
  top: "conv1a"
}
layer {
  name: "conv1b"
  type: "Convolution"
  bottom: "conv1a"
  top: "conv1b"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 32
    bias_term: true
    pad: 1
    kernel_size: 3
    group: 4
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "conv1b/bn"
  type: "BatchNorm"
  bottom: "conv1b"
  top: "conv1b"
  batch_norm_param {
    moving_average_fraction: 0.99
    eps: 0.0001
    #scale_bias: true
  }
}
layer {
  name: "conv1b/relu"
  type: "ReLU"
  bottom: "conv1b"
  top: "conv1b"
}
layer {
  name: "pool1"
  type: "Pooling"
  bottom: "conv1b"
  top: "pool1"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "res2a_branch2a"
  type: "Convolution"
  bottom: "pool1"
  top: "res2a_branch2a"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    bias_term: true
    pad: 1
    kernel_size: 3
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "res2a_branch2a/bn"
  type: "BatchNorm"
  bottom: "res2a_branch2a"
  top: "res2a_branch2a"
  batch_norm_param {
    moving_average_fraction: 0.99
    eps: 0.0001
    #scale_bias: true
  }
}
layer {
  name: "res2a_branch2a/relu"
  type: "ReLU"
  bottom: "res2a_branch2a"
  top: "res2a_branch2a"
}
layer {
  name: "res2a_branch2b"
  type: "Convolution"
  bottom: "res2a_branch2a"
  top: "res2a_branch2b"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    bias_term: true
    pad: 1
    kernel_size: 3
    group: 4
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "res2a_branch2b/bn"
  type: "BatchNorm"
  bottom: "res2a_branch2b"
  top: "res2a_branch2b"
  batch_norm_param {
    moving_average_fraction: 0.99
    eps: 0.0001
    #scale_bias: true
  }
}
layer {
  name: "res2a_branch2b/relu"
  type: "ReLU"
  bottom: "res2a_branch2b"
  top: "res2a_branch2b"
}
layer {
  name: "pool2"
  type: "Pooling"
  bottom: "res2a_branch2b"
  top: "pool2"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "res3a_branch2a"
  type: "Convolution"
  bottom: "pool2"
  top: "res3a_branch2a"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 128
    bias_term: true
    pad: 1
    kernel_size: 3
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "res3a_branch2a/bn"
  type: "BatchNorm"
  bottom: "res3a_branch2a"
  top: "res3a_branch2a"
  batch_norm_param {
    moving_average_fraction: 0.99
    eps: 0.0001
    #scale_bias: true
  }
}
layer {
  name: "res3a_branch2a/relu"
  type: "ReLU"
  bottom: "res3a_branch2a"
  top: "res3a_branch2a"
}
layer {
  name: "res3a_branch2b"
  type: "Convolution"
  bottom: "res3a_branch2a"
  top: "res3a_branch2b"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 128
    bias_term: true
    pad: 1
    kernel_size: 3
    group: 4
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "res3a_branch2b/bn"
  type: "BatchNorm"
  bottom: "res3a_branch2b"
  top: "res3a_branch2b"
  batch_norm_param {
    moving_average_fraction: 0.99
    eps: 0.0001
    #scale_bias: true
  }
}
layer {
  name: "res3a_branch2b/relu"
  type: "ReLU"
  bottom: "res3a_branch2b"
  top: "res3a_branch2b"
}
layer {
  name: "pool3"
  type: "Pooling"
  bottom: "res3a_branch2b"
  top: "pool3"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "res4a_branch2a"
  type: "Convolution"
  bottom: "pool3"
  top: "res4a_branch2a"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 256
    bias_term: true
    pad: 1
    kernel_size: 3
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "res4a_branch2a/bn"
  type: "BatchNorm"
  bottom: "res4a_branch2a"
  top: "res4a_branch2a"
  batch_norm_param {
    moving_average_fraction: 0.99
    eps: 0.0001
    #scale_bias: true
  }
}
layer {
  name: "res4a_branch2a/relu"
  type: "ReLU"
  bottom: "res4a_branch2a"
  top: "res4a_branch2a"
}
layer {
  name: "res4a_branch2b"
  type: "Convolution"
  bottom: "res4a_branch2a"
  top: "res4a_branch2b"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 256
    bias_term: true
    pad: 1
    kernel_size: 3
    group: 4
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "res4a_branch2b/bn"
  type: "BatchNorm"
  bottom: "res4a_branch2b"
  top: "res4a_branch2b"
  batch_norm_param {
    moving_average_fraction: 0.99
    eps: 0.0001
    #scale_bias: true
  }
}
layer {
  name: "res4a_branch2b/relu"
  type: "ReLU"
  bottom: "res4a_branch2b"
  top: "res4a_branch2b"
}
layer {
  name: "pool4"
  type: "Pooling"
  bottom: "res4a_branch2b"
  top: "pool4"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "res5a_branch2a"
  type: "Convolution"
  bottom: "pool4"
  top: "res5a_branch2a"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 512
    bias_term: true
    pad: 1
    kernel_size: 3
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "res5a_branch2a/bn"
  type: "BatchNorm"
  bottom: "res5a_branch2a"
  top: "res5a_branch2a"
  batch_norm_param {
    moving_average_fraction: 0.99
    eps: 0.0001
    #scale_bias: true
  }
}
layer {
  name: "res5a_branch2a/relu"
  type: "ReLU"
  bottom: "res5a_branch2a"
  top: "res5a_branch2a"
}
layer {
  name: "res5a_branch2b"
  type: "Convolution"
  bottom: "res5a_branch2a"
  top: "res5a_branch2b"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 512
    bias_term: true
    pad: 1
    kernel_size: 3
    group: 4
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "res5a_branch2b/bn"
  type: "BatchNorm"
  bottom: "res5a_branch2b"
  top: "res5a_branch2b"
  batch_norm_param {
    moving_average_fraction: 0.99
    eps: 0.0001
    #scale_bias: true
  }
}
layer {
  name: "res5a_branch2b/relu"
  type: "ReLU"
  bottom: "res5a_branch2b"
  top: "res5a_branch2b"
}
layer {
  name: "pool6"
  type: "Pooling"
  bottom: "res5a_branch2b"
  top: "pool6"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
    pad: 0
  }
}
layer {
  name: "pool7"
  type: "Pooling"
  bottom: "pool6"
  top: "pool7"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
    pad: 0
  }
}
layer {
  name: "pool8"
  type: "Pooling"
  bottom: "pool7"
  top: "pool8"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
    pad: 0
  }
}
layer {
  name: "ctx_output1"
  type: "Convolution"
  bottom: "res4a_branch2b"
  top: "ctx_output1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 256
    bias_term: true
    pad: 0
    kernel_size: 1
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "ctx_output1/bn"
  type: "BatchNorm"
  bottom: "ctx_output1"
  top: "ctx_output1"
  batch_norm_param {
    moving_average_fraction: 0.99
    eps: 0.0001
    #scale_bias: true
  }
}
layer {
  name: "ctx_output1/relu"
  type: "ReLU"
  bottom: "ctx_output1"
  top: "ctx_output1"
}
layer {
  name: "ctx_output2"
  type: "Convolution"
  bottom: "res5a_branch2b"
  top: "ctx_output2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 256
    bias_term: true
    pad: 0
    kernel_size: 1
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "ctx_output2/bn"
  type: "BatchNorm"
  bottom: "ctx_output2"
  top: "ctx_output2"
  batch_norm_param {
    moving_average_fraction: 0.99
    eps: 0.0001
    #scale_bias: true
  }
}
layer {
  name: "ctx_output2/relu"
  type: "ReLU"
  bottom: "ctx_output2"
  top: "ctx_output2"
}
layer {
  name: "ctx_output3"
  type: "Convolution"
  bottom: "pool6"
  top: "ctx_output3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 256
    bias_term: true
    pad: 0
    kernel_size: 1
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "ctx_output3/bn"
  type: "BatchNorm"
  bottom: "ctx_output3"
  top: "ctx_output3"
  batch_norm_param {
    moving_average_fraction: 0.99
    eps: 0.0001
    #scale_bias: true
  }
}
layer {
  name: "ctx_output3/relu"
  type: "ReLU"
  bottom: "ctx_output3"
  top: "ctx_output3"
}
layer {
  name: "ctx_output4"
  type: "Convolution"
  bottom: "pool7"
  top: "ctx_output4"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 256
    bias_term: true
    pad: 0
    kernel_size: 1
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "ctx_output4/bn"
  type: "BatchNorm"
  bottom: "ctx_output4"
  top: "ctx_output4"
  batch_norm_param {
    moving_average_fraction: 0.99
    eps: 0.0001
    #scale_bias: true
  }
}
layer {
  name: "ctx_output4/relu"
  type: "ReLU"
  bottom: "ctx_output4"
  top: "ctx_output4"
}
layer {
  name: "ctx_output5"
  type: "Convolution"
  bottom: "pool8"
  top: "ctx_output5"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 256
    bias_term: true
    pad: 0
    kernel_size: 1
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "ctx_output5/bn"
  type: "BatchNorm"
  bottom: "ctx_output5"
  top: "ctx_output5"
  batch_norm_param {
    moving_average_fraction: 0.99
    eps: 0.0001
    #scale_bias: true
  }
}
layer {
  name: "ctx_output5/relu"
  type: "ReLU"
  bottom: "ctx_output5"
  top: "ctx_output5"
}
layer {
  name: "ctx_output1/relu_mbox_loc"
  type: "Convolution"
  bottom: "ctx_output1"
  top: "ctx_output1/relu_mbox_loc"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 16
    bias_term: true
    pad: 0
    kernel_size: 1
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "ctx_output1/relu_mbox_loc_perm"
  type: "Permute"
  bottom: "ctx_output1/relu_mbox_loc"
  top: "ctx_output1/relu_mbox_loc_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "ctx_output1/relu_mbox_loc_flat"
  type: "Flatten"
  bottom: "ctx_output1/relu_mbox_loc_perm"
  top: "ctx_output1/relu_mbox_loc_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "ctx_output1/relu_mbox_conf"
  type: "Convolution"
  bottom: "ctx_output1"
  top: "ctx_output1/relu_mbox_conf"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 16#84
    bias_term: true
    pad: 0
    kernel_size: 1
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "ctx_output1/relu_mbox_conf_perm"
  type: "Permute"
  bottom: "ctx_output1/relu_mbox_conf"
  top: "ctx_output1/relu_mbox_conf_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "ctx_output1/relu_mbox_conf_flat"
  type: "Flatten"
  bottom: "ctx_output1/relu_mbox_conf_perm"
  top: "ctx_output1/relu_mbox_conf_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "ctx_output1/relu_mbox_priorbox"
  type: "PriorBox"
  bottom: "ctx_output1"
  bottom: "data"
  top: "ctx_output1/relu_mbox_priorbox"
  prior_box_param {
    min_size: 24 #14.72
    max_size: 50.24 #36.8
    aspect_ratio: 2
    flip: true
    clip: false
    variance: 0.1
    variance: 0.1
    variance: 0.2
    variance: 0.2
    offset: 0.5
  }
}
layer {
  name: "ctx_output2/relu_mbox_loc"
  type: "Convolution"
  bottom: "ctx_output2"
  top: "ctx_output2/relu_mbox_loc"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 24
    bias_term: true
    pad: 0
    kernel_size: 1
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "ctx_output2/relu_mbox_loc_perm"
  type: "Permute"
  bottom: "ctx_output2/relu_mbox_loc"
  top: "ctx_output2/relu_mbox_loc_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "ctx_output2/relu_mbox_loc_flat"
  type: "Flatten"
  bottom: "ctx_output2/relu_mbox_loc_perm"
  top: "ctx_output2/relu_mbox_loc_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "ctx_output2/relu_mbox_conf"
  type: "Convolution"
  bottom: "ctx_output2"
  top: "ctx_output2/relu_mbox_conf"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 24#126
    bias_term: true
    pad: 0
    kernel_size: 1
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "ctx_output2/relu_mbox_conf_perm"
  type: "Permute"
  bottom: "ctx_output2/relu_mbox_conf"
  top: "ctx_output2/relu_mbox_conf_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "ctx_output2/relu_mbox_conf_flat"
  type: "Flatten"
  bottom: "ctx_output2/relu_mbox_conf_perm"
  top: "ctx_output2/relu_mbox_conf_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "ctx_output2/relu_mbox_priorbox"
  type: "PriorBox"
  bottom: "ctx_output2"
  bottom: "data"
  top: "ctx_output2/relu_mbox_priorbox"
  prior_box_param {
    min_size: 50.24 #36.8
    max_size: 76.48 #110.4
    aspect_ratio: 2
    aspect_ratio: 3
    flip: true
    clip: false
    variance: 0.1
    variance: 0.1
    variance: 0.2
    variance: 0.2
    offset: 0.5
  }
}
layer {
  name: "ctx_output3/relu_mbox_loc"
  type: "Convolution"
  bottom: "ctx_output3"
  top: "ctx_output3/relu_mbox_loc"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 24
    bias_term: true
    pad: 0
    kernel_size: 1
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "ctx_output3/relu_mbox_loc_perm"
  type: "Permute"
  bottom: "ctx_output3/relu_mbox_loc"
  top: "ctx_output3/relu_mbox_loc_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "ctx_output3/relu_mbox_loc_flat"
  type: "Flatten"
  bottom: "ctx_output3/relu_mbox_loc_perm"
  top: "ctx_output3/relu_mbox_loc_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "ctx_output3/relu_mbox_conf"
  type: "Convolution"
  bottom: "ctx_output3"
  top: "ctx_output3/relu_mbox_conf"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 24#126
    bias_term: true
    pad: 0
    kernel_size: 1
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "ctx_output3/relu_mbox_conf_perm"
  type: "Permute"
  bottom: "ctx_output3/relu_mbox_conf"
  top: "ctx_output3/relu_mbox_conf_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "ctx_output3/relu_mbox_conf_flat"
  type: "Flatten"
  bottom: "ctx_output3/relu_mbox_conf_perm"
  top: "ctx_output3/relu_mbox_conf_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "ctx_output3/relu_mbox_priorbox"
  type: "PriorBox"
  bottom: "ctx_output3"
  bottom: "data"
  top: "ctx_output3/relu_mbox_priorbox"
  prior_box_param {
    min_size: 76.48 #110.4
    max_size: 102.72 #184.0
    aspect_ratio: 2
    aspect_ratio: 3
    flip: true
    clip: false
    variance: 0.1
    variance: 0.1
    variance: 0.2
    variance: 0.2
    offset: 0.5
  }
}
layer {
  name: "ctx_output4/relu_mbox_loc"
  type: "Convolution"
  bottom: "ctx_output4"
  top: "ctx_output4/relu_mbox_loc"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 24
    bias_term: true
    pad: 0
    kernel_size: 1
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "ctx_output4/relu_mbox_loc_perm"
  type: "Permute"
  bottom: "ctx_output4/relu_mbox_loc"
  top: "ctx_output4/relu_mbox_loc_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "ctx_output4/relu_mbox_loc_flat"
  type: "Flatten"
  bottom: "ctx_output4/relu_mbox_loc_perm"
  top: "ctx_output4/relu_mbox_loc_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "ctx_output4/relu_mbox_conf"
  type: "Convolution"
  bottom: "ctx_output4"
  top: "ctx_output4/relu_mbox_conf"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 24#126
    bias_term: true
    pad: 0
    kernel_size: 1
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "ctx_output4/relu_mbox_conf_perm"
  type: "Permute"
  bottom: "ctx_output4/relu_mbox_conf"
  top: "ctx_output4/relu_mbox_conf_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "ctx_output4/relu_mbox_conf_flat"
  type: "Flatten"
  bottom: "ctx_output4/relu_mbox_conf_perm"
  top: "ctx_output4/relu_mbox_conf_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "ctx_output4/relu_mbox_priorbox"
  type: "PriorBox"
  bottom: "ctx_output4"
  bottom: "data"
  top: "ctx_output4/relu_mbox_priorbox"
  prior_box_param {
    min_size: 102.72 #184.0
    max_size: 128.96 #257.6
    aspect_ratio: 2
    aspect_ratio: 3
    flip: true
    clip: false
    variance: 0.1
    variance: 0.1
    variance: 0.2
    variance: 0.2
    offset: 0.5
  }
}
layer {
  name: "ctx_output5/relu_mbox_loc"
  type: "Convolution"
  bottom: "ctx_output5"
  top: "ctx_output5/relu_mbox_loc"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 16
    bias_term: true
    pad: 0
    kernel_size: 1
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "ctx_output5/relu_mbox_loc_perm"
  type: "Permute"
  bottom: "ctx_output5/relu_mbox_loc"
  top: "ctx_output5/relu_mbox_loc_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "ctx_output5/relu_mbox_loc_flat"
  type: "Flatten"
  bottom: "ctx_output5/relu_mbox_loc_perm"
  top: "ctx_output5/relu_mbox_loc_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "ctx_output5/relu_mbox_conf"
  type: "Convolution"
  bottom: "ctx_output5"
  top: "ctx_output5/relu_mbox_conf"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 16#84
    bias_term: true
    pad: 0
    kernel_size: 1
    group: 1
    stride: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    dilation: 1
  }
}
layer {
  name: "ctx_output5/relu_mbox_conf_perm"
  type: "Permute"
  bottom: "ctx_output5/relu_mbox_conf"
  top: "ctx_output5/relu_mbox_conf_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "ctx_output5/relu_mbox_conf_flat"
  type: "Flatten"
  bottom: "ctx_output5/relu_mbox_conf_perm"
  top: "ctx_output5/relu_mbox_conf_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "ctx_output5/relu_mbox_priorbox"
  type: "PriorBox"
  bottom: "ctx_output5"
  bottom: "data"
  top: "ctx_output5/relu_mbox_priorbox"
  prior_box_param {
    min_size: 128.96 #257.6
    max_size: 160    #331.2
    aspect_ratio: 2
    flip: true
    clip: false
    variance: 0.1
    variance: 0.1
    variance: 0.2
    variance: 0.2
    offset: 0.5
  }
}
layer {
  name: "mbox_loc"
  type: "Concat"
  bottom: "ctx_output1/relu_mbox_loc_flat"
  bottom: "ctx_output2/relu_mbox_loc_flat"
  bottom: "ctx_output3/relu_mbox_loc_flat"
  bottom: "ctx_output4/relu_mbox_loc_flat"
  bottom: "ctx_output5/relu_mbox_loc_flat"
  top: "mbox_loc"
  concat_param {
    axis: 1
  }
}
layer {
  name: "mbox_conf"
  type: "Concat"
  bottom: "ctx_output1/relu_mbox_conf_flat"
  bottom: "ctx_output2/relu_mbox_conf_flat"
  bottom: "ctx_output3/relu_mbox_conf_flat"
  bottom: "ctx_output4/relu_mbox_conf_flat"
  bottom: "ctx_output5/relu_mbox_conf_flat"
  top: "mbox_conf"
  concat_param {
    axis: 1
  }
}
layer {
  name: "mbox_priorbox"
  type: "Concat"
  bottom: "ctx_output1/relu_mbox_priorbox"
  bottom: "ctx_output2/relu_mbox_priorbox"
  bottom: "ctx_output3/relu_mbox_priorbox"
  bottom: "ctx_output4/relu_mbox_priorbox"
  bottom: "ctx_output5/relu_mbox_priorbox"
  top: "mbox_priorbox"
  concat_param {
    axis: 2
  }
}

layer {
  name: "mbox_conf_reshape"
  type: "Reshape"
  bottom: "mbox_conf"
  top: "mbox_conf_reshape"
  reshape_param {
    shape {
      dim: 0
      dim: -1
      dim: 4
    }
  }
}
layer {
  name: "mbox_conf_softmax"
  type: "Softmax"
  bottom: "mbox_conf_reshape"
  top: "mbox_conf_softmax"
  softmax_param {
    axis: 2
  }
}
layer {
  name: "mbox_conf_flatten"
  type: "Flatten"
  bottom: "mbox_conf_softmax"
  top: "mbox_conf_flatten"
  flatten_param {
    axis: 1
  }
}

    import numpy as np  
import sys,os  
import cv2
caffe_root = '/home/caffe/'
sys.path.insert(0, caffe_root + 'python')  
import caffe  
import time

net_file = 'ti_bsd_5little_200_400_4pr_deploy.prototxt'
caffe_model = './snapshot/400*200_r/4pr_iter_130000.caffemodel' 
test_dir = "./testjpg_r_1524/"
image_outdir = "./testjpg_r_4pr_result/"
if not os.path.exists(caffe_model):
    print("MobileNetSSD_deploy.caffemodel does not exist,")
    print("use merge_bn.py to generate it.")
    exit()
net = caffe.Net(net_file,caffe_model,caffe.TEST)  
caffe.set_mode_gpu()
caffe.set_device(0)
CLASSES = ('background',
           'car', 'person', 'bicycle')
def preprocess(src):
    img = cv2.resize(src, (400,200))  
    #img = img - 127.5
    #img = img * 0.007843
    return img
def postprocess(img, out):   
    h = img.shape[0]
    w = img.shape[1]
    box = out['detection_out'][0,0,:,3:7] * np.array([w, h, w, h])
    cls = out['detection_out'][0,0,:,1]
    conf = out['detection_out'][0,0,:,2]
    return (box.astype(np.int32), conf, cls)

def detect(imgfile):
    file_path,file_name = os.path.split(imgfile)
    #print file_path,file_name
    start = time.time()
    origimg = cv2.imread(imgfile)
    img = preprocess(origimg)
    img = img.astype(np.float32)
    img = img.transpose((2, 0, 1))
    end = time.time()
    print ('Read image took {:.3f}s').format(  end - start)
    net.blobs['data'].data[...] = img
    start = time.time()
    out = net.forward()  
    box, conf, cls = postprocess(origimg, out)
    end = time.time()
    print ('Detection took {:.3f}s').format(  end - start)
    for i in range(len(box)):
       if (conf[i] > 0.3)&((str(CLASSES[int(cls[i])]) == "car")):  
         p1 = (box[i][0], box[i][1])
         p2 = (box[i][2], box[i][3])         
         cv2.rectangle(origimg, p1, p2, (0,255,0))
         p3 = (max(p1[0], 15), max(p1[1], 15))
         title = "%s:%.2f" % (CLASSES[int(cls[i])], conf[i])
         cv2.putText(origimg, title, p3, cv2.FONT_ITALIC, 0.3, (0, 255, 0), 1)        
       if (conf[i] > 0.3)&(str(CLASSES[int(cls[i])]) == "person"):
           p1 = (box[i][0], box[i][1])
           p2 = (box[i][2], box[i][3])         
           cv2.rectangle(origimg, p1, p2, (0,255,255))
           p3 = (max(p1[0], 15), max(p1[1], 15))
           title = "%s:%.2f" % (CLASSES[int(cls[i])], conf[i])
           cv2.putText(origimg, title, p3, cv2.FONT_ITALIC, 0.3, (0, 255, 0), 1)
       if (conf[i] > 0.3)&(str(CLASSES[int(cls[i])]) == "bicycle"):
           p1 = (box[i][0], box[i][1])
           p2 = (box[i][2], box[i][3])         
           cv2.rectangle(origimg, p1, p2, (0,255,255))
           p3 = (max(p1[0], 15), max(p1[1], 15))
           title = "%s:%.2f" % (CLASSES[int(cls[i])], conf[i])
           cv2.putText(origimg, title, p3, cv2.FONT_ITALIC, 0.3, (0, 255, 0), 1)
    #cv2.imshow("SSD", origimg)
    #k = cv2.waitKey(1) & 0xff
    cv2.imwrite(image_outdir + '/'+ file_name, origimg)    
        #Exit if ESC pressed
    #if k == 27 : return False
    return True

for f in sorted(os.listdir(test_dir)):
    if detect(test_dir + "/" + f) == False:
       break

    caffe代码
    c++编写的深度学习框架，关注3个目录 include/src/ tools。
    Blob 是 Caffe 中处理和传递实际数据的数据封装包，并且在 CPU 与 GPU 之间具有同步处理能力。
    Layer 是 Caffe 模型的本质内容和执行计算的基本单元。Layer 可以进行很多运算，如：convolve，pool，normalize，load data，softmax losses。
    Net 是由一系列层组成的有向无环（DAG）计算图，Caffe 保留了计算图中所有的中间值以确保前向和反向迭代的准确性。一个典型的 Net 开始于 data layer——从磁盘中加载数据，终止于 loss layer——计算如分类和重构这些任务的目标函数。
    Net::Forward()和 Net::Backward()方法实现网络的前传和后传，Solver 优化一个模型，首先通过调用前传来获得输出和损失，然后调用反传产生模型的梯度，将梯度与权值更新后相结合来最小化损失。Solver、网络和层之间的分工使得 Caffe可以模块化并且开源。
    Caffe 模型的学习被分为两个部分：由 Solver 进行优化、更新参数，由 Net 计算出 loss 和 gradient。在Caffe中，损失是通过网络的前向计算得到。
    Solver 通过协调 Net 的前向推断计算和反向梯度计算（forward inference and backward gradients）， 来对参数进行更新，从而达到减小 loss 的目的。Caffe 模型的学习被分为两个部分：由 Solver 进行优化、更新参数，由 Net 计算出 loss 和 gradient。
    为了创建一个 caffe 模型，我们需要在一个 protocol buffer(prototxt)文件中定义模型的结构。
    在 caffe 中，层和相应的参数都定义在 caffe.proto 文件里Caffe 有命令行、Python 和 MATLAB 三种接口，来实现日常使用、研究代码的交互以及实现快速原型。
    Caffe 中数据流以 Blobs 进行传输。数据层将输入转换为 blob 加载数据，将 blob 转换为其他格式保存输出。
    模型可视化工具
    http://ethereon.github.io/netscope/#/editor
    shift+enter