1. 程式人生 > >R-FCN訓練自己的模型

R-FCN訓練自己的模型

說明:

本文假設你已經做好資料集,格式和VOC2007一致,並且Linux系統已經配置好caffe所需環境(部落格裡教程很多),下面是訓練的一些修改。

py-R-FCN原始碼下載地址:

也有Matlab版本:

本文用到的是python版本。

準備工作:

(1)配置caffe環境(網上找教程)

(2)安裝cythonpython-opencveasydict

pip install cython
pip install easydict
apt-get install python-opencv

然後,我們就可以開始配置R-FCN了。

1.下載py-R-FCN

git clone https://github.com/Orpine/py-R-FCN.git

下面稱你的py-R-FCN路徑為RFCN_ROOT.

2.下載caffe

注意,該caffe版本是微軟版本
cd $RFCN_ROOT
git clone https://github.com/Microsoft/caffe.git
如果一切正常的話,python程式碼會自動新增環境變數 $RFCN_ROOT/caffe/python,否則,你需要自己新增環境變數。

3.Build Cython

cd $RFCN_ROOT/lib
make

4.Build caffe和pycaffe

cd $RFCN_ROOT/caffe
cp Makefile.config.example Makefile.config
然後修改Makefile.config。caffe必須支援python層,所以WITH_PYTHON_LAYER := 1是必須的。其他配置可參考:Makefile.config 接著:
cd $RFCN_ROOT/caffe
make -j8 && make pycaffe
如果沒有出錯,則:

5.測試Demo

經過上面的工作,我們可以測試一下是否可以正常執行。 然後將模型放在$RFCN_ROOT/data。看起來是這樣的:
$RFCN_ROOT/data/rfcn_models/resnet50_rfcn_final.caffemodel
$RFCN_ROOT/data/rfcn_models/resnet101_rfcn_final.caffemodel
執行:
cd $RFCN_ROOT
./tools/demo_rfcn.py --net ResNet-50



6.用我們的資料集訓練

(1)拷貝資料集 假設我們已經做好資料集了,格式是和VOC2007一致,將你的資料集 拷貝到$RFCN_ROOT/data下。看起來是這樣的:
$VOCdevkit0712/                           # development kit
$VOCdevkit/VOCcode/                   # VOC utility code
$VOCdevkit/VOC0712                    # image sets, annotations, etc.
# ... and several other directories ...
如果你的資料夾名字不是VOCdevkit0712和VOC0712,修改成0712就行了。 (作者是用VOC2007和VOC2012訓練的,所以資料夾名字帶0712。也可以修改程式碼,但是那樣比較麻煩一些,修改資料夾比較簡單) (2)下載預訓練模型 本文以ResNet-50為例,因此下載ResNet-50-model.caffemodel。下載地址:連結:http://pan.baidu.com/s/1slRHD0L 密碼:r3ki 然後將caffemodel放在$RFCN_ROOT/data/imagenet_models  (data下沒有該資料夾就新建一個)

(3)修改模型網路

開啟$RFCN_ROOT/models/pascal_voc/ResNet-50/rfcn_end2end  (以end2end為例) 注意:下面的cls_num指的是你資料集的類別數+1(背景)。比如我有15類,+1類背景,cls_num=16. <1>修改class-aware/train_ohem.prototxt
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': 16" #cls_num
  }
}

layer {
  name: 'roi-data'
  type: 'Python'
  bottom: 'rpn_rois'
  bottom: 'gt_boxes'
  top: 'rois'
  top: 'labels'
  top: 'bbox_targets'
  top: 'bbox_inside_weights'
  top: 'bbox_outside_weights'
  python_param {
    module: 'rpn.proposal_target_layer'
    layer: 'ProposalTargetLayer'
    param_str: "'num_classes': 16" #cls_num
  }
}

layer {
    bottom: "conv_new_1"
    top: "rfcn_cls"
    name: "rfcn_cls"
    type: "Convolution"
    convolution_param {
        num_output: 784 #cls_num*(score_maps_size^2)
        kernel_size: 1
        pad: 0
        weight_filler {
            type: "gaussian"
            std: 0.01
        }
        bias_filler {
            type: "constant"
            value: 0
        }
    }
    param {
        lr_mult: 1.0
    }
    param {
        lr_mult: 2.0
    }
}

layer {
    bottom: "conv_new_1"
    top: "rfcn_bbox"
    name: "rfcn_bbox"
    type: "Convolution"
    convolution_param {
        num_output: 3136 #4*cls_num*(score_maps_size^2)
        kernel_size: 1
        pad: 0
        weight_filler {
            type: "gaussian"
            std: 0.01
        }
        bias_filler {
            type: "constant"
            value: 0
        }
    }
    param {
        lr_mult: 1.0
    }
    param {
        lr_mult: 2.0
    }
}

layer {
    bottom: "rfcn_cls"
    bottom: "rois"
    top: "psroipooled_cls_rois"
    name: "psroipooled_cls_rois"
    type: "PSROIPooling"
    psroi_pooling_param {
        spatial_scale: 0.0625
        output_dim: 16  #cls_num
        group_size: 7
    }
}

layer {
    bottom: "rfcn_bbox"
    bottom: "rois"
    top: "psroipooled_loc_rois"
    name: "psroipooled_loc_rois"
    type: "PSROIPooling"
    psroi_pooling_param {
        spatial_scale: 0.0625
        output_dim: 64 #4*cls_num
        group_size: 7
    }
}


<2>修改class-aware/test.prototxt
layer {
    bottom: "conv_new_1"
    top: "rfcn_cls"
    name: "rfcn_cls"
    type: "Convolution"
    convolution_param {
        num_output: 784 #cls_num*(score_maps_size^2)
        kernel_size: 1
        pad: 0
        weight_filler {
            type: "gaussian"
            std: 0.01
        }
        bias_filler {
            type: "constant"
            value: 0
        }
    }
    param {
        lr_mult: 1.0
    }
    param {
        lr_mult: 2.0
    }
}

layer {
    bottom: "conv_new_1"
    top: "rfcn_bbox"
    name: "rfcn_bbox"
    type: "Convolution"
    convolution_param {
        num_output: 3136 #4*cls_num*(score_maps_size^2)
        kernel_size: 1
        pad: 0
        weight_filler {
            type: "gaussian"
            std: 0.01
        }
        bias_filler {
            type: "constant"
            value: 0
        }
    }
    param {
        lr_mult: 1.0
    }
    param {
        lr_mult: 2.0
    }
}

layer {
    bottom: "rfcn_cls"
    bottom: "rois"
    top: "psroipooled_cls_rois"
    name: "psroipooled_cls_rois"
    type: "PSROIPooling"
    psroi_pooling_param {
        spatial_scale: 0.0625
        output_dim: 16  #cls_num
        group_size: 7
    }
}

layer {
    bottom: "rfcn_bbox"
    bottom: "rois"
    top: "psroipooled_loc_rois"
    name: "psroipooled_loc_rois"
    type: "PSROIPooling"
    psroi_pooling_param {
        spatial_scale: 0.0625
        output_dim: 64  #4*cls_num
        group_size: 7
    }
}
layer {
    name: "cls_prob_reshape"
    type: "Reshape"
    bottom: "cls_prob_pre"
    top: "cls_prob"
    reshape_param {
        shape {
            dim: -1
            dim: 16  #cls_num
        }
    }
}

layer {
    name: "bbox_pred_reshape"
    type: "Reshape"
    bottom: "bbox_pred_pre"
    top: "bbox_pred"
    reshape_param {
        shape {
            dim: -1
            dim: 64  #4*cls_num
        }
    }
}

<3>修改train_agnostic.prototxt
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': 16"  #cls_num
  }
}
layer {
    bottom: "conv_new_1"
    top: "rfcn_cls"
    name: "rfcn_cls"
    type: "Convolution"
    convolution_param {
        num_output: 784 #cls_num*(score_maps_size^2)   ###
        kernel_size: 1
        pad: 0
        weight_filler {
            type: "gaussian"
            std: 0.01
        }
        bias_filler {
            type: "constant"
            value: 0
        }
    }
    param {
        lr_mult: 1.0
    }
    param {
        lr_mult: 2.0
    }
}

layer {
    bottom: "rfcn_cls"
    bottom: "rois"
    top: "psroipooled_cls_rois"
    name: "psroipooled_cls_rois"
    type: "PSROIPooling"
    psroi_pooling_param {
        spatial_scale: 0.0625
        output_dim: 16 #cls_num   ###
        group_size: 7
    }
}

<4>修改train_agnostic_ohem.prototxt

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': 16" #cls_num ###
  }
}

layer {
    bottom: "conv_new_1"
    top: "rfcn_cls"
    name: "rfcn_cls"
    type: "Convolution"
    convolution_param {
        num_output: 784 #cls_num*(score_maps_size^2)   ###
        kernel_size: 1
        pad: 0
        weight_filler {
            type: "gaussian"
            std: 0.01
        }
        bias_filler {
            type: "constant"
            value: 0
        }
    }
    param {
        lr_mult: 1.0
    }
    param {
        lr_mult: 2.0
    }
}

layer {
    bottom: "rfcn_cls"
    bottom: "rois"
    top: "psroipooled_cls_rois"
    name: "psroipooled_cls_rois"
    type: "PSROIPooling"
    psroi_pooling_param {
        spatial_scale: 0.0625
        output_dim: 16 #cls_num   ###
        group_size: 7
    }
}
<5>修改test_agnostic.prototxt
layer {
    bottom: "conv_new_1"
    top: "rfcn_cls"
    name: "rfcn_cls"
    type: "Convolution"
    convolution_param {
        num_output: 784 #cls_num*(score_maps_size^2) ###
        kernel_size: 1
        pad: 0
        weight_filler {
            type: "gaussian"
            std: 0.01
        }
        bias_filler {
            type: "constant"
            value: 0
        }
    }
    param {
        lr_mult: 1.0
    }
    param {
        lr_mult: 2.0
    }
}

layer {
    bottom: "rfcn_cls"
    bottom: "rois"
    top: "psroipooled_cls_rois"
    name: "psroipooled_cls_rois"
    type: "PSROIPooling"
    psroi_pooling_param {
        spatial_scale: 0.0625
        output_dim: 16 #cls_num   ###
        group_size: 7
    }
}

layer {
    name: "cls_prob_reshape"
    type: "Reshape"
    bottom: "cls_prob_pre"
    top: "cls_prob"
    reshape_param {
        shape {
            dim: -1
            dim: 16 #cls_num   ###
        }
    }
}

(4)修改程式碼

<1>$RFCN/lib/datasets/pascal_voc.py
class pascal_voc(imdb):
    def __init__(self, image_set, year, devkit_path=None):
        imdb.__init__(self, 'voc_' + year + '_' + image_set)
        self._year = year
        self._image_set = image_set
        self._devkit_path = self._get_default_path() if devkit_path is None \
                            else devkit_path
        self._data_path = os.path.join(self._devkit_path, 'VOC' + self._year)
        self._classes = ('__background__', # always index 0
                         '你的標籤1','你的標籤2',你的標籤3','你的標籤4'
                      )
改成你的資料集標籤。 <2>$RFCN_ROOT/lib/datasets/imdb.py 主要是assert (boxes[:, 2] >= boxes[:, 0]).all()可能出現AssertionError,具體解決辦法參考: PS: 上面將有無ohem的prototxt都改了,但是這裡訓練用的是ohem。 另外,預設的迭代次數很大,可以修改$RFCN\experiments\scripts\rfcn_end2end_ohem.sh:
case $DATASET in
  pascal_voc)
    TRAIN_IMDB="voc_0712_trainval"
    TEST_IMDB="voc_0712_test"
    PT_DIR="pascal_voc"
    ITERS=110000

修改ITERS為你想要的迭代次數即可。

(5)開始訓練

cd $RFCN_ROOT
./experiments/scripts/rfcn_end2end_ohem.sh 0 ResNet-50 pascal_voc
正常的話,就開始迭代了:

$RFCN_ROOT/experiments/scripts裡還有一些其他的訓練方法,也可以測試一下(經過上面的修改,無ohem的end2end訓練也改好了,其他訓練方法修改的過程差不多)。

(6)結果

將訓練得到的模型($RFCN_ROOT/output/rfcn_end2end_ohem/voc_0712_trainval裡最後的caffemodel)拷貝到$RFCN_ROOT/data/rfcn_models下,然後開啟$RFCN_ROOT/tools/demo_rfcn.py,將CLASSES修改成你的標籤,NETS修改成你的model,im_names修改成你的測試圖片(放在data/demo下),最後:
cd $RFCN_ROOT
./tools/demo_rfcn.py --net ResNet-50