caffe学习记录(六) MobileNet fine tune

记录在unbantu14.04， caffe框架下对MobileNet的自有数据集fine tune。

首先git clone一下caffe版本的mobilenet   https://github.com/shicai/MobileNet-Caffe.git

 

然后把deploy.prototxt文件修改一下

Modify deploy.prototxt and save it as your train.prototxt as follows: Remove the first 5 input/input_dim lines, and add Image Data layer in the beginning like this:

 

layer {
  name: "data"
  type: "ImageData"
  top: "data"
  top: "label"
  include {
    phase: TRAIN
  }
  transform_param {
    scale: 0.017
    mirror: true
    crop_size: 224
    mean_value: [103.94, 116.78, 123.68]
  }
  image_data_param {
    source: "your_list_train_txt"
    batch_size: 32 # your batch size
    new_height: 256
    new_width: 256
    root_folder: "your_path_to_training_data_folder"
  }
}

Remove the last prob layer, and add Loss and Accuracy layers in the end like this:

layer {
  name: "loss"
  type: "SoftmaxWithLoss"
  bottom: "fc7"
  bottom: "label"
  top: "loss"
}
layer {
  name: "top1/acc"
  type: "Accuracy"
  bottom: "fc7"
  bottom: "label"
  top: "top1/acc"
  include {
    phase: TEST
  }
}
layer {
  name: "top5/acc"
  type: "Accuracy"
  bottom: "fc7"
  bottom: "label"
  top: "top5/acc"
  include {
    phase: TEST
  }
  accuracy_param {
    top_k: 5
  }
}

然后包括了train_val.prototxt, deploy.prototxt, 

通过createDB.py来生成带label的txt文件：代码如下:
第二个方法为生成LMDB的文件

# -*- coding: UTF-8 -*-
import os
import re
import commands

def createFileList(images_path, txt_save_path):
    fw = open(txt_save_path,"w")
    images_name = os.listdir(images_path)
    for eachname in images_name:

        pattern_cat = r'(^cat.d{0,10}.jpg$)'
        pattern_dog = r'(^dog.d{0,10}.jpg$)'
        cat_name = re.search(pattern_cat, eachname)
        dog_name = re.search(pattern_dog, eachname)

        if cat_name != None:
            fw.write(cat_name.group(0) + ' n16000001n')
        if dog_name != None:
            fw.write(dog_name.group(0) + ' n16000002n')

    print "done with txt generation!"

    fw.close()

def create_db(caffe_root, images_path, txt_save_path):

    lmdb_name = 'img_train.lmdb'

    lmdb_save_path = caffe_root + 'models/MobileNet-Caffe/' + lmdb_name

    convert_imageset_path = caffe_root + 'build/tools/convert_imageset'
    cmd = """%s --shuffle --resize_height=256 --resize_width=256 %s %s %s"""
    status, output = commands.getstatusoutput(cmd % (convert_imageset_path, images_path, 
        txt_save_path, lmdb_save_path))
    print output
    if(status == 0):
        print "lmbd is done!"

if __name__ == '__main__':
  
    caffe_root = '/home/wy/ssd-caffe/caffe/'

    my_caffe_project = caffe_root + 'models/MobileNet-Caffe/'

    images_path = caffe_root + 'models/MobileNet-Caffe/val/'

    txt_name = 'label_val.txt'

    txt_save_path = my_caffe_project + txt_name

    createFileList(images_path, txt_save_path)
 
    #create_db(caffe_root, images_path, txt_save_path)

生成的txt文件如下，带labels:

对于自己的solver，由于怀疑测试集是imagenet的分支，所以，把lr调的很低，我的solver文件如下:

调参的过程很复杂，因为收敛非常快，最后总结出一个办法，要么把lr调低，要么在提前fc一层就开始fine tune。

tips:在fine tune的时候，要改layer中的名字，这样会使得caffemodel的weight参数跳过赋值。

最后训练的结果如下

有时间再把整个过程整理下，今天权当记录一下。

 最终accuracy nearly 98.5%