Added tensor support for strings and for getting TF_DataType #7
Conversation
| memcpy((void *) tensor_resource_alloc, (void *) &tensor, sizeof(TF_Tensor *)); | ||
| ERL_NIF_TERM new_tensor = enif_make_resource(env, tensor_resource_alloc); | ||
| enif_release_resource(tensor_resource_alloc); | ||
| return new_tensor; |
There was a problem hiding this comment.
Since we are returning {:error, ...} in failure cases, it is recommended to return {:ok, tensor} in success cases.
There was a problem hiding this comment.
Agreed. In the next PR, I'll add this. I'll extend it to read_graph too
| TF_Tensor **tensor_resource_alloc = enif_alloc_resource(tensor_resource, sizeof(TF_Tensor *)); | ||
|
|
||
| if (!(enif_is_binary(env, argv[0]))) { | ||
| return enif_make_tuple2(env, enif_make_atom(env,"error"), enif_make_atom(env,"non_binary_argument")); |
There was a problem hiding this comment.
If the error happens because of bad input, it is preferrable to just raise an error: maybe we can use enif_make_badarg? https://erlang.org/doc/man/erl_nif.html#enif_make_badarg
In a nutshell, we return tuple errors for semantic errors: for example, you couldn't load a file because it is corrupt or it doesn't exist. For language errors, such as invalid syntax or invalid arguments, we should raise instead.
There was a problem hiding this comment.
Understood. Will incorporate this in the next PR as well
@josevalim, In this PR I have added some functions that were working perfectly. However I have actually been devoting a lot of time to getting
TF_INT64andTF_FLOATtensors (especially multi-dimensional ones) to work in Elixir. I will add PRs for those soon, once I get them to work completely.In this PR, the functionalities added are:
An example for these functions is as follows: