app.py

# import tensorflow as tf
# from google.protobuf.struct_pb2 import Struct
# import struct2tensor.ops.gen_decode_proto_sparse

# def create_proto_message(text):
#     message = Struct()
#     message.fields["task_data"].string_value = text
#     return message.SerializeToString()

# class TFProtoModel:
#     def __init__(self, model_path):
#         self.model = tf.saved_model.load(model_path)
#         self.infer = self.model.signatures['serving_default']
    
#     def predict(self, text):
#         proto_data = create_proto_message(text)
#         input_tensor = tf.constant([proto_data], dtype=tf.string)
#         result = self.infer(inputs=input_tensor)
#         return result['outputs'].numpy()

# # Initialize model when the file is loaded
# model = TFProtoModel("model")

# # This is the function Hugging Face will call
# def pipeline(text):
#     return model.predict(text)


import tensorflow as tf
from google.protobuf.struct_pb2 import Struct
from transformers import Pipeline
import struct2tensor.ops.gen_decode_proto_sparse


def create_proto_message(text):
    message = Struct()
    message.fields["task_data"].string_value = text
    return message.SerializeToString()

class TFProtoModel(Pipeline):
    def __init__(self, model_path="model"):
        self.model = tf.saved_model.load(model_path)
        self.infer = self.model.signatures['serving_default']
    
    def _sanitize_parameters(self, **kwargs):
        return {}, {}, {}

    def preprocess(self, text):
        proto_data = create_proto_message(text)
        return tf.constant([proto_data], dtype=tf.string)

    def _forward(self, input_tensor):
        result = self.infer(inputs=input_tensor)
        return result['outputs'].numpy()

    def postprocess(self, model_outputs):
        return {"score": float(model_outputs[0])}

pipeline = TFProtoModel

# To specify the task
task = "text-classification"  # or another appropriate task type