gepc 增强数据的加载 patch enhance

参数设置

    parser.add_argument('--patch_features', action='store_true', help='Use patch features instead of coordinates. (default: False)')
    parser.add_argument('--patch_db', action='store_true', help='Use pre-extracted patch db files. (default: False)')

• 去掉action=‘store_true’（store_true实现如果添加参数则为true,否则默认为false）
• 用default方式 添加参数的默认值，然后将patch_features设置为True

    parser.add_argument('--patch_features', default=True,type=bool,
                        help='Use patch features instead of coordinates. (default: False)')
    parser.add_argument('--patch_db', default=False,type=bool,
                        help='Use pre-extracted patch db files. (default: False)')

def init_stc_sub_args(args):
    if args.patch_features:  # No pose augmentations for patch models !!!!
        args.num_transform = 1
    if args.debug:
        args.ae_epochs = 10
        args.dcec_epochs = 25

    args.vid_path = {'train': os.path.join(args.data_dir, 'training/videos/'),
                     'test':  os.path.join(args.data_dir, 'testing/frames/')}# 视频路径参数 ！！！！！

    args.pose_path = {'train': os.path.join(args.data_dir, 'pose', '***/gepc-master1/data/pose/training/tracked_person/'),
                      'test':  os.path.join(args.data_dir, 'pose', '***/gepc-master1/data/pose/testing/tracked_person/')}

    args.ckpt_dir = create_exp_dirs(args.exp_dir)
    args.optimizer = args.ae_optimizer
    ae_args = args_rm_prefix(args, 'ae_')
    dcec_args = args_rm_prefix(args, 'dcec_')
    res_args = args_rm_prefix(args, 'res_')
    return args, ae_args, dcec_args, res_args

读取

不使用patch_db

    if args.patch_db:
        patch_suffix_str = 'ing{}x{}.lmdb'.format(patch_size, patch_size)
        patch_size = (patch_size, patch_size)
        patch_db_path = {k: os.path.join(v, k+patch_suffix_str) for k, v in args.vid_path.items()}
    else:
        patch_db_path = {k: None for k, v in args.vid_path.items()} # 如果不使用patch_db，那么就置为None

到PoseSegDataset方法

进入PoseSegDataset方法

PoseSegDataset方法调用gen_dataset方法(返回和正常调用一样的seg数据)

注意在loaddata时的不同！！！

• index由loader提供

对姿态数据进行变换

    def __getitem__(self, index):
        # Select sample and augmentation. I.e. given 5 samples and 2 transformations,
        # sample 7 is data sample 7%5=2 and transform is 7//5=1
        if self.apply_transforms:
            sample_index = index % self.num_samples
            trans_index = index // self.num_samples
            data_numpy = np.array(self.segs_data_np[sample_index])
            data_transformed = self.transform_list[trans_index](data_numpy)
        else:
            sample_index = index
            data_transformed = data_numpy = np.array(self.segs_data_np[index])
            trans_index = 0  # No transformations

        seg_metadata = self.segs_meta[sample_index]
        ret_arr = [data_transformed, trans_index]
        if self.return_metadata:# True
            ret_arr += [seg_metadata]
        if self.use_patches:  # Add patch data to loaded segments ---  True
            if self.use_patches_db:  # Use pre extracted, slice correct segment and cast ToTensor --- False
                key = ('{:02d}_{:04d}_{:02d}'.format(*seg_metadata[:3]))
                person_patches_np, _ = patches_from_db(self.patches_db, key.encode('ascii'))
                person_keys_sorted = self.person_keys[key]
                start_ofst = person_keys_sorted.index(seg_metadata[-1])
                seg_patches_np = person_patches_np[start_ofst: start_ofst+self.seg_len]
                seg_patches_tensor = seg_patches_to_tensor(seg_patches_np)
            else:  # Extract patches from individual jpeg frames
                dirn = '{:02d}_{:03d}'.format(seg_metadata[0], seg_metadata[1])
                seg_patches_tensor = self.get_set_patch_tensor(data_numpy, dirn, seg_metadata)

            if self.headless:
                seg_patches_tensor = seg_patches_tensor[:, :, :14]
            ret_arr = [seg_patches_tensor] + ret_arr[1:]  # Replace pose w/patches

        if self.return_indices:
            ret_arr += [index]
        return ret_arr

    def get_set_patch_tensor(self, data_numpy, dirn, metadata):
        return get_seg_patches(os.path.join(self.path_to_vid_dir, dirn), data_numpy, metadata,
                               lmdb_env=self.patches_db, patch_size=self.patch_size)

def get_seg_patches(img_dir, seg_pose_data, seg_meta, lmdb_env=None, patch_size=None, pre_proc_seg=None):
    """
    Collates a segments patches. Allows reuse of a previous segment with only appending missing frames
    """
    img_list = [img for img in os.listdir(img_dir) if img.endswith('.jpg')]
    fn_prefix = img_list[4].split('.')[0] # 只是用于提供计算一下前缀的长度
    if len(fn_prefix) == 3:
        fmt_str = '{:03d}.jpg'
    else:
        fmt_str = '{:04d}.jpg'

    seg_patches = []
    first_vid_frame = seg_meta[-1]
    if pre_proc_seg is not None:
        first_load_frame = pre_proc_seg.shape[0]
    else:# True
        first_load_frame = 0
    num_frames = seg_pose_data.shape[1] # 12

    for t in range(first_load_frame, num_frames):
        img_fn = fmt_str.format(first_vid_frame + t)
        img_path = os.path.join(img_dir, img_fn)
        sing_pose_data = seg_pose_data[:2, t]
        curr_frame_patches = get_single_img_patches(img_path, sing_pose_data, lmdb_env=lmdb_env,
                                                    patch_size=patch_size)
        seg_patches.append(curr_frame_patches)
    new_seg_patches = np.stack(seg_patches)
    if pre_proc_seg is not None:
        seg_patches = np.concatenate([pre_proc_seg, new_seg_patches], axis=0)
    else:
        seg_patches = new_seg_patches

    seg_patches = np.transpose(seg_patches, (4, 0, 1, 2, 3))
    seg_patches = seg_patches.astype(np.float32)
    seg_patches /= 255.0
    return seg_patches

get_single_img_patches函数！！！（\gepc-master\utils\patch_utils.py）

def get_single_img_patches(img_path图像路径, sing_pose_data（2，18）, lmdb_env=None, patch_size=None 16):
    """
    Loads a single image's patches. When an lmdb_env is provided, loads from there
    :param img_path:
    :param sing_pose_data:
    :param lmdb_env:
    :param patch_size:
    :return:
    """
    if patch_size is None:
        patch_size = np.array([32, 32])
    elif isinstance(patch_size, int):
        patch_size = (patch_size, patch_size)
    patch_size = np.array(patch_size, dtype=np.int32) // 2 # [16, 16] 到 [8, 8]

    if lmdb_env is None:
        img = Image.open(img_path)
    else:
        key = '_'.join(img_path.split('.')[0].split('/')[-2:])  # '/videos/01_003/0692.jpg' -> '01_003_0692'
        key = key.encode('ascii')
        img = img_from_db(lmdb_env, key)

    int_coords = sing_pose_data[:2].transpose().astype(np.int32)
    patch_coords = np.array([int_coords - patch_size, int_coords + patch_size])
    patch_coords = patch_coords.transpose([1, 0, 2]).reshape(-1, 4)
    patches = [img.crop(pc) for pc in patch_coords]
    patches_np = np.stack([np.asarray(i) for i in patches])
    return patches_np

    #[18,2] 
    int_coords = sing_pose_data[:2].transpose().astype(np.int32)
    #[2,18,2]   注：int_coords - patch_size的大小为[18,2] 
    patch_coords = np.array([int_coords - patch_size, int_coords + patch_size])

• 再裁剪之后，堆叠起来就变为了[18,16,16,3],即函数的最终返回

• 将12帧 patch enhance骨架 的都堆叠起来得到一个new_seg_patches(12, 18, 16, 16, 3)

    new_seg_patches = np.stack(seg_patches)
    if pre_proc_seg is not None:
        seg_patches = np.concatenate([pre_proc_seg, new_seg_patches], axis=0)
    else:# True
        seg_patches = new_seg_patches # (12, 18, 16, 16, 3)

    seg_patches = np.transpose(seg_patches, (4, 0, 1, 2, 3))# (3, 12, 18, 16, 16)
    seg_patches = seg_patches.astype(np.float32)
    seg_patches /= 255.0
    return seg_patches

• 再getitem中得到“seg_patches_tensor”

            else:  # Extract patches from individual jpeg frames
                dirn = '{:02d}_{:03d}'.format(seg_metadata[0], seg_metadata[1])
                “seg_patches_tensor” = self.get_set_patch_tensor(data_numpy, dirn, seg_metadata)
            if self.headless:# False
                seg_patches_tensor = seg_patches_tensor[:, :, :14]
            ret_arr = [seg_patches_tensor] + ret_arr[1:]  # Replace pose w/patches

        if self.return_indices:
            ret_arr += [index]
        return ret_arr

ret_arr = [seg_patches_tensor] + ret_arr[1:] # Replace pose w/patches

• 替换数组的第一个原始数据
  
  

ret_arr += [index]加上index项，得到最终的结果

获取的数据在程序中的使用(当前是batchsize=2的情况)

在调试的过程中会进行batchsize次的执行才跳到data = data_arr[0].to(args.device, non_blocking=True)句