LRCN代码复现3

心已赠人 2022-06-10 08:30 229阅读 0赞

## 训练单帧RGB模型 ##

### run\_singleFrame\_RGB.sh ###

single frame model 是在模型caffe\_imagenet\_hyb2\_wr\_rc\_solver\_sqrt\_iter\_310000的基础上finetune而来的 ，该脚本里面主要内容就是下面这句话

caffe train -solver singleFrame_solver_RGB.prototxt -weights  caffe_imagenet_hyb2_wr_rc_solver_sqrt_iter_310000

### \-solver ###

singleFrame\_solver\_RGB.prototxt 内容如下

net: "train_test_singleFrame_RGB.prototxt"
    test_iter: 75 
    test_state: { stage: 'test-on-test' }
    test_interval: 100
    base_lr: 0.001
    lr_policy: "step"
    gamma: 0.1
    stepsize: 3000
    display: 20
    max_iter: 5000
    momentum: 0.9
    weight_decay: 0.005  
    snapshot: 5000
    snapshot_prefix: "snapshots_singleFrame_RGB"
    solver_mode: GPU
    device_id: 0 
    random_seed: 1701

### net ###

现在我们知道啦，solver除了指定net,还要指定其他一些参数。这个solver里使用的net是train\_test\_singleFrame\_flow.prototxt ，以下是网络结构中的前两层

name: "singleFrame_RGB"
    layer {
      name: "data"
      type: "ImageData"
      top: "data"
      top: "label"
      include {
        phase: TRAIN      #训练阶段输入层
      }
      transform_param {
        mirror: true
        crop_size: 227
        mean_value: 103.939
        mean_value: 116.779
        mean_value: 123.68
        flow: false
      }
      image_data_param {
        source: "ucf101_singleFrame_RGB_train_split1.txt"
        root_folder: "frames/"
        batch_size: 128
        new_height: 240
        new_width: 320
      }
    }

如果想测试，把caffe后的train改成test即可，训练和测试只是data层不一样  
训练好model 的名字是什么，输出在哪定义  
如果pretrained model 和自己的网络模型不一样怎么办  
需要用那些参数，参数如何配置，比如只配置前几层的参数？

## 训练单帧flow模型 ##

run\_singleFrame\_flow.sh 中主要是一下这句话

caffe train -solver singleFrame_solver_flow.prototxt -weights caffe_imagenet_hyb2_wr_rc_solver_sqrt_iter_310000

### \-solver ###

net: "train_test_singleFrame_flow.prototxt"
    test_iter: 75 
    test_state: { stage: 'test-on-test' }
    test_interval: 100
    base_lr: 0.001
    lr_policy: "step"
    gamma: 0.1
    stepsize: 20000
    display: 20
    max_iter: 50000
    momentum: 0.9
    weight_decay: 0.005  
    snapshot: 5000
    snapshot_prefix: "snapshots_singleFrame_flow"
    solver_mode: GPU
    device_id: 1 
    random_seed: 1701

### \- net ###

layer { 
      name: "data" 
      type: "ImageData" 
      top: "data" 
      top: "label" 
      include { 
        phase: TRAIN 
      } 
      transform_param { 
        mirror: true 
        crop_size: 227 
        mean_value: 128 
        mean_value: 128 
        mean_value: 128 
        flow: true 
      } 
      image_data_param { 
        source: "ucf101_singleFrame_flow_train_split1.txt"              
        root_folder: "flow_images/"  #需要修改的地方
        batch_size: 128 
        min_height: 227 
        min_width: 227 
      } 
    }

## 训练lstm\_RGB模型 ##

将single frame RGB的训练结果当做pretrained权重，来训练lstm模型

### run\_lstm\_RGB.sh ###

caffe train -solver lstm_solver_RGB.prototxt -weights single_frame_all_layers_hyb_RGB_iter_5000.caffemodel

### \-solver ###

lstm\_solver\_RGB.prototxt

### net ###

layer { 
      name: "data" 
      type: "Python" 
      top: "data" 
      top: "label" 
      top: "clip_markers" 
      python_param { 
        module: "sequence_input_layer" 
        layer: "videoReadTrain_RGB" 
      } 
      include: { phase: TRAIN } 
    }

lstm模型的数据层是python模块，而single模型的第一层是直接输入img 。  
下面我们就来看看这个python模块

### input ###

sequence\_input\_layer.py （没有main函数）

flow_frames = 'flow_images/' 
    RGB_frames = 'frames/' 
    test_frames = 16  
    train_frames = 16 
    test_buffer = 3 
    train_buffer = 24

修改RGB图片和光流图片的路径

### \-weight ###

是singleframe模型训练好的single\_frame\_all\_layers\_hyb\_RGB\_iter\_5000.caffemodel

## 训练lstm\_flow模型 ##

run\_lstm\_flow.sh