py faster rcnn訓練 VGG_CNN_M_1024記錄

ZF網絡已經訓練通過，參考訓練ZF的步驟修改相關文件。環境：CPU+ Ubuntu16.04

1、修改py-faster-rcnn/models/pascal_voc/VGG_CNN_M_1024/faster_rcnn_alt_opt/stage1_fast_rcnn_train.pt & stage2_fast_rcnn_train.pt

layer {
  name: 'data'
  type: 'Python'
  top: 'data'
  top: 'rois'
  top: 'labels'
  top: 'bbox_targets'
  top: 'bbox_inside_weights'
  top: 'bbox_outside_weights'
  python_param {
    module: 'roi_data_layer.layer'
    layer: 'RoIDataLayer'
    param_str: "'num_classes': 2"#original is 21 ,class_num + 1  
  }
}

layer {
  name: "cls_score"
  type: "InnerProduct"
  bottom: "fc7"
  top: "cls_score"
  param { lr_mult: 1 }
  param { lr_mult: 2 }
  inner_product_param {
    num_output: 2 # origial is 21, class_num + 1
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "bbox_pred"
  type: "InnerProduct"
  bottom: "fc7"
  top: "bbox_pred"
  param { lr_mult: 1 }
  param { lr_mult: 2 }
  inner_product_param {
    num_output: 8 # original is 84, (class_num+1)*4
    weight_filler {
      type: "gaussian"
      std: 0.001
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

2、修改py-faster-rcnn/models/pascal_voc/VGG_CNN_M_1024/faster_rcnn_alt_opt/stage1_rpn_train.pt & stage2_rpn_train.pt

layer {
  name: 'input-data'
  type: 'Python'
  top: 'data'
  top: 'im_info'
  top: 'gt_boxes'
  python_param {
    module: 'roi_data_layer.layer'
    layer: 'RoIDataLayer'
    param_str: "'num_classes': 2" # original is 21, class_num + 1
  }
}

3、py-faster-rcnn/models/pascal_voc/VGG_CNN_M_1024/faster_rcnn_alt_opt/faster_rcnn_test.pt

layer {
  name: "cls_score"
  type: "InnerProduct"
  bottom: "fc7"
  top: "cls_score"
  inner_product_param {
    num_output: 2 # original is 21, class_num + 1
  }
}
layer {
  name: "bbox_pred"
  type: "InnerProduct"
  bottom: "fc7"
  top: "bbox_pred"
  inner_product_param {
    num_output: 8 # original is 84, (class_num + 1)*4
  }
}

4、py-faster-rcnn/lib/datasets/pascal_voc.py修改（ZF時已修改）

   def __init__(self, image_set, year, devkit_path=None):
        imdb.__init__(self, 'voc_' + year + '_' + image_set)
        self._year = year
        self._image_set = image_set
        self._devkit_path = self._get_default_path() if devkit_path is None \
                            else devkit_path
        self._data_path = os.path.join(self._devkit_path, 'VOC' + self._year)
        self._classes = ('__background__', # always index 0
                         #'aeroplane', 'bicycle', 'bird', 'boat',
                         #'bottle', 'bus', 'car', 'cat', 'chair',
                         #'cow', 'diningtable', 'dog', 'horse',
                         #'motorbike', 'person', 'pottedplant',
                         #'sheep', 'sofa', 'train', 'tvmonitor'
			'leftAtrial'
			)

5、py-faster-rcnn/lib/datasets/imdb.py修改（ZF時已修改）

將append_flipped_images(self)函數的第二行widths = self._get_widths()改爲如下形式：
widths = [PIL.Image.open(self.image_path_at(i)).size[0]
	for i in xrange(num_images)]

6、修改tools/train_faster_rcnn_alt_opt.py中的迭代次數

max_iters = [8000, 4000, 8000, 4000]
修改models/pascal_voc/VGG_CNN_M_1024/faster_rcnn_alt_opt/下對應的solver文件的stepsize（學習率等超參也在這裏修改），要小於max_iters
stepsize: 3000 # original is 30000, should smaller than max_iters[1]#stage1_fast_rcnn_solver30k40k.pt
stepsize: 6000 # original is 60000, should smaller than max_iters[0]#stage1_rpn_solver60k80k.pt
stepsize: 3000 # original is 30000, should smaller than max_iters[3]#stage2_fast_rcnn_solver30k40k.pt
stepsize: 6000 # original is 60000, should smaller than max_iters[2]#stage2_rpn_solver60k80k.pt

將solver中第一行的路徑改爲絕對路徑，否則會報找不到路徑的錯誤

train_net: "/home/lys/py-faster-rcnn/models/pascal_voc/VGG_CNN_M_1024/faster_rcnn_alt_opt/stage1_fast_rcnn_train.pt"

7、刪除output以及/data/cache，py-faster-rcnn/data/VOCdevkit2007/annotation_cache（使用test_net.py測試時產生）將對應的文件放入/data/VOCdevkit2007/VOC2007/

8、訓練：

python train_faster_rcnn_alt_opt.py  --net_name VGG_CNN_M_1024 --weights /home/lys/py-faster-rcnn/data/imagenet_models/VGG_CNN_M_1024.v2.caffemodel --cfg /home/lys/py-faster-rcnn/experiments/cfgs/faster_rcnn_alt_opt.yml  --imdb voc_2007_trainval 

I0831 11:23:16.666177 22761 solver.cpp:229] Iteration 0, loss = 1.16274
I0831 11:23:16.666204 22761 solver.cpp:245]     Train net output #0: rpn_cls_loss = 0.689726 (* 1 = 0.689726 loss)
I0831 11:23:16.666209 22761 solver.cpp:245]     Train net output #1: rpn_loss_bbox = 0.473018 (* 1 = 0.473018 loss)
I0831 11:23:16.666214 22761 sgd_solver.cpp:106] Iteration 0, lr = 0.001
I0831 11:24:29.267117 22761 solver.cpp:229] Iteration 20, loss = 0.560839
I0831 11:24:29.267145 22761 solver.cpp:245]     Train net output #0: rpn_cls_loss = 0.495616 (* 1 = 0.495616 loss)
I0831 11:24:29.267151 22761 solver.cpp:245]     Train net output #1: rpn_loss_bbox = 0.0652227 (* 1 = 0.0652227 loss)
 

9、使用demo.py進行測試

需要加入NETS中加入'vgg_m':('VGG_CNN_M_1024',  'VGG_CNN_M_1024_faster_rcnn_final.caffemodel')，紅色部分應該與/models/pascal_voc/下的VGG_CNN_M_1024文件名相同。

python demo.py --net vgg_m --cpu