上一篇已经把 detectron装起来了

现在就把voc数据转为coco数据，因为detectron只支持coco。
原谅我不懂caffe2.
我有一批数据，是人脸检测的。

我要用这个框架就必须进行如下几个步骤。

1. 数据转换
2. 下载pkl预训练文件
3. 设置yaml，这是配置文件
4. 修改文件
5. 推送数据

1. 数据转换

使用voc_to_coco.py把voc装换成coco

"""
Created on Tue Jun 12 10:24:36 2018
将voc格式转json格式用于caffe2的detectron的训练
在detectron中voc_2007_train.json和voc_2007_val.json中categories的顺序必须保持一致
因此最好事先确定category的顺序，书写在category_set中
@author: yantianwang
"""
 
import xml.etree.ElementTree as ET
import os
import json
import collections
 
coco = dict()
coco['images'] = []
coco['type'] = 'instances'
coco['annotations'] = []
coco['categories'] = []
 
#category_set = dict()
image_set = set()
image_id = 2019000001  #train:2018xxx; val:2019xxx; test:2020xxx
category_item_id = 1
annotation_id = 1
# category_set = ['small-vehicle', 'large-vehicle', 'plane', 'harbor', 'ship',
#                 'tennis-court', 'soccer-ball-field', 'ground-track-field',
#                 'baseball-diamond', 'swimming-pool', 'roundabout', 'basketball-court',
#                 'storage-tank', 'bridge', 'helicopter']

category_set = ['face', ]
'''
def addCatItem(name):
    global category_item_id
    category_item = dict()
    category_item['supercategory'] = 'none'
    category_item_id += 1
    category_item['id'] = category_item_id
    category_item['name'] = name
    coco['categories'].append(category_item)
    category_set[name] = category_item_id
    return category_item_id
'''
 
def addCatItem(name):
    '''
    增加json格式中的categories部分
    '''
    global category_item_id
    category_item = collections.OrderedDict()
    category_item['supercategory'] = 'none'
    category_item['id'] = category_item_id
    category_item['name'] = name
    coco['categories'].append(category_item)
    category_item_id += 1
 
def addImgItem(file_name, size):
    global image_id
    if file_name is None:
        raise Exception('Could not find filename tag in xml file.')
    if size['width'] is None:
        raise Exception('Could not find width tag in xml file.')
    if size['height'] is None:
        raise Exception('Could not find height tag in xml file.')
    #image_item = dict()    #按照一定的顺序，这里采用collections.OrderedDict()
    image_item = collections.OrderedDict()
    jpg_name = os.path.splitext(file_name)[0]+'.png'
    image_item['file_name'] = jpg_name  
    image_item['width'] = size['width']   
    image_item['height'] = size['height']
    image_item['id'] = image_id
    coco['images'].append(image_item)
    image_set.add(jpg_name)    
    image_id = image_id+1
    return image_id
 
 
def addAnnoItem(object_name, image_id, category_id, bbox):
    global annotation_id
    #annotation_item = dict()
    annotation_item = collections.OrderedDict()
    annotation_item['segmentation'] = []
    seg = []
    # bbox[] is x,y,w,h
    # left_top
    seg.append(bbox[0])
    seg.append(bbox[1])
    # left_bottom
    seg.append(bbox[0])
    seg.append(bbox[1] + bbox[3])
    # right_bottom
    seg.append(bbox[0] + bbox[2])
    seg.append(bbox[1] + bbox[3])
    # right_top
    seg.append(bbox[0] + bbox[2])
    seg.append(bbox[1])
    annotation_item['segmentation'].append(seg)
    annotation_item['area'] = bbox[2] * bbox[3]
    annotation_item['iscrowd'] = 0
    annotation_item['image_id'] = image_id
    annotation_item['bbox'] = bbox
    annotation_item['category_id'] = category_id
    annotation_item['id'] = annotation_id
    annotation_item['ignore'] = 0 
    annotation_id += 1 
    coco['annotations'].append(annotation_item)
 
 
def parseXmlFiles(xml_path):
    xmllist = os.listdir(xml_path)
    xmllist.sort()
    for f in xmllist:
        if not f.endswith('.xml'):
            continue
 
        bndbox = dict()
        size = dict()
        current_image_id = None
        current_category_id = None
        file_name = None
        size['width'] = None
        size['height'] = None
        size['depth'] = None
 
        xml_file = os.path.join(xml_path, f)
        print(xml_file)
 
        tree = ET.parse(xml_file)
        root = tree.getroot() #抓根结点元素
 
        if root.tag != 'annotation': #根节点标签
            raise Exception('pascal voc xml root element should be annotation, rather than {}'.format(root.tag))
 
        # elem is <folder>, <filename>, <size>, <object>
        for elem in root:
            current_parent = elem.tag
            current_sub = None
            object_name = None
 
            #elem.tag, elem.attrib，elem.text
            if elem.tag == 'folder':
                continue
 
            if elem.tag == 'filename':
                file_name = elem.text
                if file_name in category_set:
                    raise Exception('file_name duplicated')
 
            # add img item only after parse <size> tag
            elif current_image_id is None and file_name is not None and size['width'] is not None:
                if file_name not in image_set:
                    current_image_id = addImgItem(file_name, size)#图片信息
                    print('add image with {} and {}'.format(file_name, size))
                else:
                    raise Exception('duplicated image: {}'.format(file_name))
                    # subelem is <width>, <height>, <depth>, <name>, <bndbox>
            for subelem in elem:
                bndbox['xmin'] = None
                bndbox['xmax'] = None
                bndbox['ymin'] = None
                bndbox['ymax'] = None
 
                current_sub = subelem.tag
                if current_parent == 'object' and subelem.tag == 'name':
                    object_name = subelem.text
                    #if object_name not in category_set:
                    #    current_category_id = addCatItem(object_name)
                    #else:
                    #current_category_id = category_set[object_name]
                    current_category_id = category_set.index(object_name)+1 #index默认从0开始,但是json文件是从1开始，所以+1
                elif current_parent == 'size':
                    if size[subelem.tag] is not None:
                        raise Exception('xml structure broken at size tag.')
                    size[subelem.tag] = int(subelem.text)
 
                # option is <xmin>, <ymin>, <xmax>, <ymax>, when subelem is <bndbox>
                for option in subelem:
                    if current_sub == 'bndbox':
                        if bndbox[option.tag] is not None:
                            raise Exception('xml structure corrupted at bndbox tag.')
                        bndbox[option.tag] = int(option.text)
 
                # only after parse the <object> tag
                if bndbox['xmin'] is not None:
                    if object_name is None:
                        raise Exception('xml structure broken at bndbox tag')
                    if current_image_id is None:
                        raise Exception('xml structure broken at bndbox tag')
                    if current_category_id is None:
                        raise Exception('xml structure broken at bndbox tag')
                    bbox = []
                    # x
                    bbox.append(bndbox['xmin'])
                    # y
                    bbox.append(bndbox['ymin'])
                    # w
                    bbox.append(bndbox['xmax'] - bndbox['xmin'])
                    # h
                    bbox.append(bndbox['ymax'] - bndbox['ymin'])
                    print(
                    'add annotation with {},{},{},{}'.format(object_name, current_image_id-1, current_category_id, bbox))
                    addAnnoItem(object_name, current_image_id-1, current_category_id, bbox)
    #categories部分
    for categoryname in category_set:
        addCatItem(categoryname) 
 
 
if __name__ == '__main__':
    xml_path = '/home/yantianwang/clone/VOC2007/val_Annotations'
    json_file = '/home/yantianwang/clone/VOC2007/Annotations/voc_2007_val.json'
    parseXmlFiles(xml_path)
json.dump(coco, open(json_file, 'w'))

唯一需要修改的的地方就是category_set=['face', ]

2. 下载pkl预训练文件

我需要ResNeXt-101-64x4d，从预训练参数开始

这里model zoo

我下的是X-101-64x4d.pkl：

3. 设置yaml，这是参数

从docker中的config文件中cp一份yaml文件出来。。

如下：

我使用e2e_faster_rcnn_X-101-64x4d-FPN_2x.yaml。。。

如下e2e_faster_rcnn_X-101-64x4d-FPN_2x.yaml：

MODEL:
  TYPE: generalized_rcnn
  CONV_BODY: FPN.add_fpn_ResNet101_conv5_body
  NUM_CLASSES: 2
  FASTER_RCNN: True
NUM_GPUS: 1
SOLVER:
  WEIGHT_DECAY: 0.0001
  LR_POLICY: steps_with_decay
  # 2x schedule (note TRAIN.IMS_PER_BATCH: 1)
  BASE_LR: 0.001
  GAMMA: 0.1
  MAX_ITER: 360000
  STEPS: [0, 240000, 320000]
FPN:
  FPN_ON: True
  MULTILEVEL_ROIS: True
  MULTILEVEL_RPN: True
  RPN_ANCHOR_START_SIZE: 8
RESNETS:
  STRIDE_1X1: False  # default True for MSRA; False for C2 or Torch models
  TRANS_FUNC: bottleneck_transformation
  NUM_GROUPS: 64
  WIDTH_PER_GROUP: 4
FAST_RCNN:
  ROI_BOX_HEAD: fast_rcnn_heads.add_roi_2mlp_head
  ROI_XFORM_METHOD: RoIAlign
  ROI_XFORM_RESOLUTION: 7
  ROI_XFORM_SAMPLING_RATIO: 2
TRAIN:
  # md5sum of weights pkl file: aa14062280226e48f569ef1c7212e7c7
  WEIGHTS: /detectron/work/data/pkl/X-101-64x4d.pkl
  DATASETS: ('train_data', )
  SCALES: (550,650,750,850,950)
  MAX_SIZE: 1600
  IMS_PER_BATCH: 1
  BATCH_SIZE_PER_IM: 512
  RPN_PRE_NMS_TOP_N: 2000  # Per FPN level
TEST:
  DATASETS: ('val_data',)
  SCALE: 750
  MAX_SIZE: 1600
  NMS: 0.5
  RPN_PRE_NMS_TOP_N: 1000  # Per FPN level
  RPN_POST_NMS_TOP_N: 1000
  DETECTIONS_PER_IM: 5000
  BBOX_AUG:
    ENABLED: True
    SCORE_HEUR: UNION
    COORD_HEUR: UNION
    H_FLIP: True
    SCALES: (550, 650,750, 850,950)
    MAX_SIZE: 1600
    SCALE_H_FLIP: True
    SCALE_SIZE_DEP: False
    AREA_TH_LO: 64  # 8^2
    AREA_TH_HI: 32400  # 180^2
  BBOX_VOTE:
    ENABLED: True
    VOTE_TH: 0.8
  SOFT_NMS:
    ENABLED: True
OUTPUT_DIR: .

4. 修改文件
   训练之前还需要修改data_cotalog.py
   在docker中有一个文件需要修改，但是我先在本地修改再推上去。。。

就是detectron/datasets/data_cotalog.py

在文件最后加上train_data和test_data就是对应yaml文件的数据地址keywords。

    'voc_2012_train': {
        _IM_DIR:
            _DATA_DIR + '/VOC2012/JPEGImages',
        _ANN_FN:
            _DATA_DIR + '/VOC2012/annotations/voc_2012_train.json',
        _DEVKIT_DIR:
            _DATA_DIR + '/VOC2012/VOCdevkit2012'
    },
    'voc_2012_val': {
        _IM_DIR:
            _DATA_DIR + '/VOC2012/JPEGImages',
        _ANN_FN:
            _DATA_DIR + '/VOC2012/annotations/voc_2012_val.json',
        _DEVKIT_DIR:
            _DATA_DIR + '/VOC2012/VOCdevkit2012'
    },
    'train_data': {
        _IM_DIR:
            _DATA_DIR + '/train_data/JPEGImages',
        _ANN_FN:
            _DATA_DIR + '/train_data/my_train.json',
    },
    'val_data': {
        _IM_DIR:
            _DATA_DIR + '/test_data/JPEGImages',
        _ANN_FN:
            _DATA_DIR + '/test_data/my_val.json',
    }

5. 推送数据

把本地数据推到docker中去

在打开nvidia-docker，把本地文件与docker进行映射。

如下：

sudo nvidia-docker run -it  -v $PWD:/detectron/work housebw/detectron:latest /bin/bash

这时我的当前目录的文件在，detectron/work 中。

推送数据

cp -r work/val_data detectron/datasets/data
cp -r work/train_data detectron/datasets/data

我的本地数据test_data和train_data已经到data中了。

6 训练：

注意我的预训练数据就在work/data/pkl。因为我已经进行映射了

执行如下进行训练

python  tools/train_net.py --cfg work/data/our_e2e_faster_rcnn_X-101-64x4d-FPN_2x.yaml OUTPUT_DIR work/data/output

训练过程信息

7 推断

修改dummp_datasets.py和infer_simple.py

在Detectron的lib/datasets/dummy_datasets.py文件中仿照get_coco_dataset()函数添加一个get_face_dataset()函数

def get_VOC_dataset():
    """A dummy VOC dataset"""
    ds = AttrDict()
    classes = ['__background__', 'face']
    ds.classes = {i:name for i, name in enumerate(classes)}
    return ds

然后在infer_simple.py文件中将main()函数里面的
dummy_coco_dataset = dummy_datasets.get_coco_dataset()
改成dummy_coco_dataset = dummy_datasets.get_voc_dataset()

python tools/test_net.py --cfg work/data/our_e2e_faster_rcnn_X-101-64x4d-FPN_2x.yaml TEST.WEIGHTS work/data/output/train/train_data/generalized_rcnn/model_final.pkl  
  work/testdata NUM_GPUS 1

8 评估

python2 tools/test_net.py \
    --cfg work/data/our_e2e_faster_rcnn_X-101-64x4d-FPN_2x.yaml  \
    TEST.WEIGHTS /work/data/output/train/train_data/generalizer_rnn/model_final.pkl \
    NUM_GPUS 1

多GPU运行

看这里
https://github.com/facebookresearch/Detectron/issues/551

未完待续....................

参考：

将voc格式的数据转为json格式用于detectron的训练
使用自己的数据集（voc2007格式）训练Detectron
Detectron研读和实践三
Detectron系列文章