def _load_img_feature_pickle(self):
        for filepath in self._all_img_feature_filepaths:
            logging.info("loading %s" % filepath)
            with gfile.GFile(filepath, 'r') as f:
                filenames, features = pickle.load(f)
                self._img_feature_filenames += filenames  
                self._img_feature_data.append(features) 
        #[#(1000, 1, 1, 2048), #(1000, 1, 1, 2048)] -> [#(2000, 1, 1, 2048)]   
        self._img_feature_data = np.vstack(self._img_feature_data)
        origin_shape = self._img_feature_data.shape

为什么这里_img_feature_data 的shape是(1000, 1, 1, 2048)?

前面在提取 特征图 并 pickle时, 用的时inception倒数第二层 是一个pooling层 1x1x2048 , 训练的batch_size是 100
与现在_img_feature_data 有什么关联？ 我理不清楚了, 求老师解惑