FASTER-RCNN+ZF training model with its own data set (Python version)

Description: This blog post assumes that you have made your own dataset, and that the dataset is formatted the same as VOC2007. Here are some pre-training changes.

(The process of doing a dataset can be seen here.)

FASTER-RCNN Source:

MATLAB version: HTTPS://GITHUB.COM/SHAOQINGREN/FASTER_RCNN

Python version: HTTPS://GITHUB.COM/RBGIRSHICK/PY-FASTER-RCNN

This article uses the Python version, which runs under Linux .

Preparatory work:

1. Configure Caffe

This does not say much, online tutorials a lot.

2. Other considerations

This is very detailed here, take a look at it carefully. Address: https://github.com/rbgirshick/py-faster-rcnn (main content below)

Here is about to translate the contents of the above URL.

(1) Installing the Cython, python-opencveasydict

Pip Install cythonpip install easydictapt-get install PYTHON-OPENCV

(2) Download py-faster-rcnn

# Make sure to clone with--recursivegit clone--recursive Https://github.com/rbgirshick/py-faster-rcnn.git

(3) Enter Py-faster-rcnn/lib

Execute make

(4) Enter PY-FASTER-RCNN\CAFFE-FAST-RCNN

Execute CP Makefile.config.example Makefile.config

Then, configure the Makefile.config file to refer to my configuration: Makefile.config file

After configuring the Makefile.config file, execute:

Make-j8 && make Pycaffe

(5) Download VOC2007 data set

Provide a Baidu cloud address: http://yun.baidu.com/s/1hr5SfK4

Unzip, then, place the dataset under Py-faster-rcnn\data and replace the VOC2007 DataSet with your dataset.

(6) Download the pre-trained model parameters under the Imagenet data set (for initialization)

Provide a Baidu cloud address: Http://pan.baidu.com/s/1hsxx8OW

Unzip, and then place the file under Py-faster-rcnn\data

Here are some pre-training changes.

1.py-faster-rcnn/models/pascal_voc/zf/faster_rcnn_alt_opt/stage1_fast_rcnn_train.pt modification

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 ': #按训练集类别改, changed to category +1  }}

Layer {  name: "Cls_score"  type: "Innerproduct"  Bottom: "fc7"  Top: "Cls_score"  param {lr_ mult:1.0}  param {lr_mult:2.0}  inner_product_param {    num_output:16 #按训练集类别改, changed to class number +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:1.0 }  param {lr_mult:2.0}  inner_product_param {    num_output:64 #按训练集类别改, the value is (category number +1),    Weight_filler {      type: "Gaussian"      std:0.001    }    bias_filler {      type: "Constant"      value:0  }}}

2.py-faster-rcnn/models/pascal_voc/zf/faster_rcnn_alt_opt/stage1_rpn_train.pt modification

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 ': + ' #按训练集类别改, The value is the number of categories +1  }}

3.py-faster-rcnn/models/pascal_voc/zf/faster_rcnn_alt_opt/stage2_fast_rcnn_train.pt modification

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 ': #按训练集类别改, changed to category +1  }}

Layer {  name: "Cls_score"  type: "Innerproduct"  Bottom: "fc7"  Top: "Cls_score"  param {lr_mult:1.0 }  param {lr_mult:2.0}  inner_product_param {    num_output:16 #按训练集类别改, this value is the number of categories +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:1.0}  param {lr_mult:2.0}  inner_product_param {    num_output:64 #按训练集类别改, this value is (category number +1)    Weight_filler {      type: ' Gaussian '      std:0.001    }    bias_filler {      type: ' Constant '      value:0    }  }}

4.py-faster-rcnn/models/pascal_voc/zf/faster_rcnn_alt_opt/stage2_rpn_train.pt modification

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 ': + ' #按训练集类别改, The value is the number of categories +1  }}

5.py-faster-rcnn/models/pascal_voc/zf/faster_rcnn_alt_opt/faster_rcnn_test.pt modification

Layer {  name: "Cls_score"  type: "Innerproduct"  Bottom: "fc7"  Top: "Cls_score"  inner_ Product_param {    num_output:16 #按训练集类别改, this value is the number of categories +1  }}

Layer {  name: "bbox_pred"  type: "Innerproduct"  Bottom: "fc7"  Top: "bbox_pred"  inner_product_ param {    num_output:64 #按训练集类别改, the value is (number of categories +1)}  }

6.py-faster-rcnn/lib/datasets/pascal_voc.py modification

Class Pascal_voc (IMDb):    def __init__ (self, image_set, year, Devkit_path=none): imdb.__init__ (Self, ' voc_ ' + yea R + image_set) self._year = year Self._image_set = Image_set Self._devkit_path = Self._get_default_pa Th () If Devkit_path is None else devkit_path Self._data_path = Os.path.join (Self._devki T_path, ' VOC ' +self._year)        self._classes = (' __background__ ', # always index 0                         ' your label 1 ',                         ' your label 2 ', ' your tag 3 ', ' Your label 4 ',                      )

The place to change is

To modify the training set folder:

Self._data_path = Os.path.join (Self._devkit_path, ' VOC ' +self._year)

Replace the annotations,imagesets and Jpegimages in the original VOC2007 with your data set directly.

To modify a label:

self._classes = (' __background__ ', # always index 0                         ' your label 1 ',                         ' your label 2 ', ' your tag 3 ', ' your tag 4 ',                      )

Modify the label on your dataset to be OK.

7.py-faster-rcnn/lib/datasets/imdb.py modify the file by modifying the append_flipped_images (self) function to:

def append_flipped_images (self):        num_images = self.num_images        widths = [PIL. Image.open (Self.image_path_at (i)). Size[0] for                  i in Xrange (num_images)] for        i in Xrange (num_images):            boxes = self.roidb[i][' boxes '].copy ()            oldx1 = boxes[:, 0].copy ()            oldx2 = boxes[:, 2].copy ()            boxes[:, 0] = wid Ths[i]-oldx2-1            print boxes[:, 0]            boxes[:, 2] = widths[i]-oldx1-1            print boxes[:, 0]            assert (boxes [:, 2] >= boxes[:, 0]). All ()            entry = {' boxes ': boxes,                     ' gt_overlaps ': self.roidb[i][' gt_overlaps '],                     ' Gt_classes ': self.roidb[i][' gt_classes ',                     ' flipped ': True}            self.roidb.append (entry)        Self._image_ index = Self._image_index * 2

！！！ To prevent mixing with the previous model, delete the output folder (or change another name) before training, and delete the files in the Py-faster-rcnn\data\cache (if any)

For settings such as the learning rate, you can set the solve file in py-faster-rcnn/models/pascal_voc/zf/faster_rcnn_alt_opt.

8. Start training

Enter PY-FASTER-RCNN, execute:

experiments/scripts/faster_rcnn_alt_opt.sh 0 ZF Pascal_voc

In this way, we start training.

9. Testing

Copy the Py-faster-rcnn\output\faster_rcnn_alt_opt\***_trainval ZF Caffemodel from the training to Py-faster-rcnn\data\faster_rcnn_ Models, and then, modify:

py-faster-rcnn\tools\demo.py, main changes:

CLASSES = (' __background__ ',           ' your label 1 ', ' Your label 2 ', ' Your label 3 ', ' your Tag 4 ')

Change to your data set label;

NETS = {' Vgg16 ': (' VGG16 ',                  ' Vgg16_faster_rcnn_final.caffemodel '),        ' ZF ': (' ZF ',                  ' zf_faster_rcnn_ Final.caffemodel ')}

The Caffemodel of the above ZF changed to your Caffemodel.

10. Results

Under the PY-FASTER-RCNN,

Perform:

./tools/demo.py--net ZF





FASTER-RCNN+ZF training model with its own data set (Python version)