pycaffe對lmdb的寫入與讀取

輕量數據庫，刪和改就沒啥必要了。

1，將圖片數據寫入lmdb。

def image_dataset_to_lmdb(image_folder_path,lmdb_path):
    # create db
    db = lmdb.open(lmdb_path, map_size=int(1e12))
    # start transaction
    with db.begin(write=True) as db_txn:
        # 枚舉圖片地址
        for parent, dirnames, filenames in os.walk(image_folder_path):
            # 限制目錄范圍，子目錄跳過
            if (parent == image_folder_path):
                for index, filename in enumerate(filenames, start=0):
                    image_path = image_folder_path + filename

                    # read image into numpy.ndarray
                    image_array = numpy.array(Image.open(image_path))

                    # image type recognition
                    if type(image_array[0][0]) is numpy.uint8:
                        print(image_path + ' => image type : grayscale, this function can not handle this case, skip.')
                        continue
                    elif type(image_array[0][0]) is numpy.ndarray:
                        print(image_path + ' => image type : color, go.')

                    # RGB format to BGR format
                    image_array = image_array[:, :, ::-1]
                    # channel split by transpose matrix operation
                    image_array = image_array.transpose((2, 0, 1))
                    # convert numpy ndarray to caffe datum format
                    im_dat = caffe.io.array_to_datum(image_array)
                    # write to db
                    # 這里可以為每張圖片設置y值
                    im_dat.label = 1
                    # '{:0>10d}'.format(index) => key, im_dat.SerializeToString() => value
                    db_txn.put('{:0>10d}'.format(index), im_dat.SerializeToString())
    db.close()

2，讀取數據庫

def read_lmdb(lmdb_path):

    #open lmdb
    lmdb_env = lmdb.open(lmdb_path)
    #begin transaction
    lmdb_txn = lmdb_env.begin()
    #get cursor
    lmdb_cursor = lmdb_txn.cursor()
    #get data object
    datum = caffe.proto.caffe_pb2.Datum()

    for key, value in lmdb_cursor:
        #parse back to datum
        datum.ParseFromString(value)
        #get y value
        label = datum.label
        print('label = ' + str(label))
        data_array = caffe.io.datum_to_array(datum)
        print('data is numpy.ndarray :')
        for data in data_array:
            print(str(data) + '\n')

　　lmdb_env.close()