Source: model_server/utils/tensor_utils.py#L0

create_tensor_proto

create_tensor_proto(array, shape=None, dtype=None, name=None)

Create a TensorProto from numpy array.

Arguments:

  • array (np.ndarray): The numpy array to convert to TensorProto
  • shape (tuple): Optional shape of the array to reshape it to. If not given, it is inferred from the numpy array. Default None
  • dtype (str): Optional dtype to convert the array to. Refer tensor.proto for supported datatypes. If not given, it is inferred from the numpy array. Default None.
  • name (str): Optional name for the TensorProto. Default None

Returns:

A TensorProto containing the given array

Note:

  • python strings will be encoded to python bytes.
  • Use dtype = "object" if the numpy array contains strings.
  • dtype "string" and "object" are treated as same.
  • "string" is converted to python "object". This is because, numpy handles variable length strings in this way

create_array_from_proto

create_array_from_proto(tensor_proto)

Retrive array from TensorProto.

Arguments:

  • tensor_proto (TensorProto): An instance of TensorProto

Returns:

numpy array

Note:

  • python strings will be encoded to python bytes.
  • Use dtype = "object" if the numpy array contains strings.
  • dtype "string" and "object" are treated as same.
  • "string" is converted to python "object". This is because, numpy handles variable length strings in this way