modeling_npk.py

import pickle
import pandas as pd
from transformers import PreTrainedModel, PretrainedConfig
from huggingface_hub import hf_hub_download

class NPKConfig(PretrainedConfig):
    model_type = "npk"
    
    def __init__(self, **kwargs):
        super().__init__(**kwargs)

class NPKPredictionModel(PreTrainedModel):
    config_class = NPKConfig
    
    def __init__(self, config):
        super().__init__(config)
        self.xgb_model = None
        self.label_encoder = None
        self._load_models()
    
    def _load_models(self):
        # Load the XGBoost model and label encoder
        xgb_path = hf_hub_download(repo_id=self.config._name_or_path, filename="npk_prediction_model.pkl")
        le_path = hf_hub_download(repo_id=self.config._name_or_path, filename="label_encoder.pkl")
        
        with open(xgb_path, 'rb') as f:
            self.xgb_model = pickle.load(f)
        
        with open(le_path, 'rb') as f:
            self.label_encoder = pickle.load(f)
    
    def forward(self, inputs):
        # Preprocess inputs
        processed_inputs = {}
        for key, value in inputs.items():
            if isinstance(value, list):
                processed_inputs[key] = value[0] if value else None
            else:
                processed_inputs[key] = value

        crop_name = processed_inputs['crop_name']
        processed_inputs['crop_name'] = self.label_encoder.transform([crop_name])[0]
        
        input_df = pd.DataFrame([processed_inputs])
        
        # Make prediction
        prediction = self.xgb_model.predict(input_df)
        
        return {
            'Nitrogen Need': float(prediction[0][0]),
            'Phosphorus Need': float(prediction[0][1]),
            'Potassium Need': float(prediction[0][2])
        }
    
    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
        config = kwargs.pop("config", None)
        if config is None:
            config = NPKConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
        
        model = cls(config)
        return model