如何在张量流中加载并保存ssd网络的Widerface数据集？

Question

我想加载并保存ssd（single shot multibox detector）网络的Widerface标签，但是张力很大，但是wider_face_train_bbx_gt太复杂了。

如何在张量流中保存ssd网络的标签？

Answer 1

要加载数据集，我给你一个解释如何在TensorFlow中使用TensorPack（仅用于数据）来完成此操作。

首先，我们需要包含边界框的zip文件和mat文件。下面的部分主要直接从zip文件和MAT文件

class RawWiderFaceReader(RNGDataFlow): 
    """Read images directly from tar file without unpacking 
    boxes: left, top, width, height 
    """ 
    def __init__(self, matfile, zipfile): 
     super(RawWiderFaceReader, self).__init__() 
     self.matfile = matfile 
     self.zipfile = zipfile 
     self.subset = matfile.split('_')[-1].replace('.mat', '') 
     f = sio.loadmat(matfile) 
     events = [f['event_list'][i][0][0] for i in range(len(f['event_list']))] 
     raw_files = [f['file_list'][i][0] for i in range(len(f['file_list']))] 
     raw_bbx = [f['face_bbx_list'][i][0] for i in range(len(f['face_bbx_list']))] 

     col_files = [] 
     for file, bbx in zip(raw_files, raw_bbx): 
      for filee, bbxe in zip(file, bbx): 
       col_files.append((filee[0][0], bbxe[0])) 

     self.col_files2 = [] 
     for file, bbx in col_files: 
      for ev in events: 
       if file.startswith(ev.replace('--', '_')): 
        self.col_files2.append((str('WIDER_%s/images/' % self.subset + ev + 
              '/' + file + '.jpg').encode('ascii', 'ignore'), bbx)) 
        break 

    def get_data(self): 
     with ZipFile(self.zipfile, 'r') as zip_hnd: 
      for fn, bbx in self.col_files2: 
       buf = zip_hnd.read('%s' % fn) 
       yield [buf, bbx] 

它给你一台发电机get_data()返回JPEG编码的图像和包围盒读取。它的存储方式似乎很复杂，因为它是一个包含由Matlab生成的边界框的文件。 要绘制边界框，你可以使用：

def draw_rect(img, top, left, bottom, right, rgb, margin=1): 
    m = margin 
    r, g, b = rgb 
    img[top:bottom, left - m:left + m, 0] = r 
    img[top:bottom, left - m:left + m, 1] = g 
    img[top:bottom, left - m:left + m, 2] = b 

    img[top:bottom, right - m:right + m, 0] = r 
    img[top:bottom, right - m:right + m, 1] = g 
    img[top:bottom, right - m:right + m, 2] = b 

    img[top - m:top + m, left:right, 0] = r 
    img[top - m:top + m, left:right, 1] = g 
    img[top - m:top + m, left:right, 2] = b 

    img[bottom - m:bottom + m, left:right, 0] = r 
    img[bottom - m:bottom + m, left:right, 1] = g 
    img[bottom - m:bottom + m, left:right, 2] = b 

    return img 

整个脚本是在这里： https://gist.github.com/PatWie/a743d2349f388b27ed3ef783919c3882

pip install -U git+https://github.com/ppwwyyxx/tensorpack.git后，您可以通过

python data_sampler.py --zip /scratch/patwie/data/wider_face/WIDER_val.zip \ 
         --mat wider_face_split/wider_face_val.mat \ 
         --debug 

火起来要将其转换成lmdb-file可以使用其他参数。这里不需要解压缩数据。

使用数据，这就像在脚本：

from tensorpack import * 
ds = LMDBDataPoint('/scratch/wieschol/data/wider_face/WIDER_train.lmdb', shuffle=True) 
ds = RawWiderFaceReader(matfile=args.mat, zipfile=args.zip) 
ds.reset_state() 
for jpeg, bbx in ds.get_data(): 
    rgb = cv2.imdecode(np.asarray(jpeg), cv2.IMREAD_COLOR)