Function CaffeLMDB

tensorpack/dataflow/format.py:172–204

Read a Caffe-format LMDB file where each value contains a ``caffe.Datum`` protobuf. Produces datapoints of the format: [HWC image, label]. Note that Caffe LMDB format is not efficient: it stores serialized raw arrays rather than JPEG images. Args: lmdb_path, shuffle, k

(lmdb_path, shuffle=True, keys=None)

def CaffeLMDB(lmdb_path, shuffle=True, keys=None):
    """
    Read a Caffe-format LMDB file where each value contains a ``caffe.Datum`` protobuf.
    Produces datapoints of the format: [HWC image, label].

    Note that Caffe LMDB format is not efficient: it stores serialized raw
    arrays rather than JPEG images.

    Args:
        lmdb_path, shuffle, keys: same as :class:`LMDBData`.

    Example:
        .. code-block:: python

            ds = CaffeLMDB("/tmp/validation", keys='{:0>8d}')
    """

    cpb = get_caffe_pb()
    lmdb_data = LMDBData(lmdb_path, shuffle, keys)

    def decoder(k, v):
        try:
            datum = cpb.Datum()
            datum.ParseFromString(v)
            img = np.fromstring(datum.data, dtype=np.uint8)
            img = img.reshape(datum.channels, datum.height, datum.width)
        except Exception:
            log_once("Cannot read key {}".format(k), 'warn')
            return None
        return [img.transpose(1, 2, 0), datum.label]
    logger.warn("Caffe LMDB format doesn't store jpeg-compressed images, \
        it's not recommended due to its inferior performance.")
    return LMDBDataDecoder(lmdb_data, decoder)


class SVMLightData(RNGDataFlow):