I am trying to run a pre-trained model (The MuRIL model). However, it seems that it requires the tensorflow_text package.

When running the saved model with python, this is simple:

import tensorflow as tf
import tensorflow_text # important to load RegexSplitWithOffsets

preprocessor = tf.saved_model.load("./preprocess")
muril = tf.saved_model.load("./muril")

input_word = "testing word"

preprocessed = preprocessor([input_word])
output_vec = muril(preprocessed)["pooled_output"]

print(output_vec)

However, when I do something similar with the Tensorflow C API:

#include <stdio.h>
#include <tensorflow/c/c_api.h>

TF_Status* status;

TF_SessionOptions* preprocess_opts;
TF_Graph* preprocess_graph;
TF_Session* preprocess_sess;

TF_SessionOptions* muril_opts;
TF_Graph* muril_graph;
TF_Session* muril_sess;

int load_models() {

    // create a status reporter
    status = TF_NewStatus();

    // the tag is "serve"
    const char* tags[] = {"serve"};

    // load models

    // load preprocessor model
    char* preprocess_model_path = "./preprocess";
    preprocess_graph = TF_NewGraph();
    preprocess_opts = TF_NewSessionOptions();
    preprocess_sess = TF_LoadSessionFromSavedModel(preprocess_opts, NULL, preprocess_model_path, tags, 1, preprocess_graph, NULL, status);

    if (TF_GetCode(status) != TF_OK) {
        printf("Error loading preprocess model.");
        return 1;
    }

    // load main muril model
    char* muril_model_path = "./muril";
    muril_graph = TF_NewGraph();
    muril_opts = TF_NewSessionOptions();
    muril_sess = TF_LoadSessionFromSavedModel(muril_opts, NULL, muril_model_path, tags, 1, muril_graph, NULL, status);

    if (TF_GetCode(status) != TF_OK) {
        printf("Error loading muril model.");
        return 1;
    }

    return 0;
}

I get a runtime error:

fg-shape-inference{graph-version=0},tfg-prepare-attrs-export)} failed: INVALID_ARGUMENT: Unable to find OpDef for RegexSplitWithOffsets
        While importing function: __inference__wrapped_model_4156
        when importing GraphDef to MLIR module in GrapplerHook
2025-03-10 12:14:25.280346: I tensorflow/cc/saved_model/loader:220] Running initialization op on SavedModel bundle at path: ./preprocess
2025-03-10 12:14:25.300561: E tensorflow/core/grappler/optimizers/tfg_optimizer_hook:135] tfg_optimizer{any(tfg-consolidate-attrs,tfg-toposort,tfg-shape-inference{graph-version=0},tfg-prepare-attrs-export)} failed: INVALID_ARGUMENT: Unable to find OpDef for RegexSplitWithOffsets
        While importing function: __inference__wrapped_model_4156
        when importing GraphDef to MLIR module in GrapplerHook
2025-03-10 12:14:25.342181: I tensorflow/cc/saved_model/loader:466] SavedModel load for tags { serve }; Status: success: OK. Took 154554 microseconds.

I believe this is due to me not having the tensorflow_text package installed for C. I followed the instructions on the tensorflow website to install, and unpacked the tar.gz file into /usr/local/lib which created libtensorflow.so and libtensorflow.so.2 files. But there isn't any tensorflow_text.

I can't find a guide on how to install tensorflow_text for C; does anyone know how?