# writefile facility_predict.py import os import random import json import numpy as np import torch import heapq import pandas as pd from tqdm import tqdm from transformers import AutoModelForSequenceClassification, AutoTokenizer from torch.utils.data import TensorDataset, DataLoader class Preprocess: def __init__(self, tokenizer_vocab_path, tokenizer_max_len): self.stopwords = ["i", "was", "transferred", "from", "to", "nilienda", "kituo", "cha", "lakini", "saa", "hii", "niko", "at", "nilienda", "nikahudumiwa", "pole", "deliver", "na", "ni", "baada", "ya", "kutumwa", "kutoka", "nilienda", "ndipo", "nikapewa", "hiyo", "lindam ama", "nikawa", "mgonjwa", "nikatibiwa", "in", "had", "a", "visit", "gynaecologist", "ndio", "karibu", "mimi", "niko", "sehemu", "hospitali", "serikali", "delivered", "katika", "kaunti", "kujifungua", "katika", "huko", "nilipoenda", "kwa", "bado", "naedelea", "sija", "maliza", "mwisho", "nilianza", "kliniki", "yangu", "nilianzia", "nilijifungua"] self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_vocab_path, use_auth_token='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') self.max_len = tokenizer_max_len def clean_text(self, text): text = text.lower() self.text_single = ' '.join(word for word in text.split() if word not in self.stopwords) return self.text_single def encode_fn(self): """ Using tokenizer to preprocess the text example of text_single:'Nairobi Hospital' """ tokenizer = self.tokenizer(self.text_single, padding=True, truncation=True, max_length=self.max_len, return_tensors='pt' ) input_ids = tokenizer['input_ids'] attention_mask = tokenizer['attention_mask'] return input_ids, attention_mask def process_tokenizer(self, data): """ Preprocess text and prepare dataloader for a single new sentence """ self.clean_text(data) input_ids, attention_mask = self.encode_fn() data = TensorDataset(input_ids, attention_mask) return data class Facility_Model: def __init__(self, facility_model_path: any, max_len: int): self.max_len = max_len self.softmax = torch.nn.Softmax(dim=1) self.gpu = False self.model = AutoModelForSequenceClassification.from_pretrained(facility_model_path, use_auth_token='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') self.model.eval() # set pytorch model for inference mode if torch.cuda.device_count() > 1: self.model = torch.nn.DataParallel(self.model) if self.gpu: seed = 42 random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) torch.cuda.manual_seed_all(seed) torch.backends.cudnn.deterministic = True self.device = torch.device('cuda') else: self.device = 'cpu' self.model = self.model.to(self.device) def predict_single(self, model, pred_data): """ Model inference for new single sentence """ pred_dataloader = DataLoader(pred_data, batch_size=10, shuffle=False) for i, batch in enumerate(pred_dataloader): with torch.no_grad(): outputs = model(input_ids=batch[0].to(self.device), attention_mask=batch[1].to(self.device) ) loss, logits = outputs.loss, outputs.logits probability = self.softmax(logits) probability_list = probability.detach().cpu().numpy() return probability_list def output_intent_probability(self, pred: any) -> dict: """ convert the model output into a dictionary with all intents and its probability """ output_dict = {} # transform the relation table(between label and intent) path_table = pd.read_csv('dhis_label_relation_14357.csv') label_intent_dict = path_table[["label", "corresponding_label"]].set_index("corresponding_label").to_dict()[ 'label'] # transform the output into dictionary(between intent and probability) for intent in range(pred.shape[1]): output_dict[label_intent_dict[intent]] = pred[0][intent] return output_dict def inference(self, prepared_data): """ Make predictions on one new sentence and output a JSON format variable """ temp = [] prob_distribution = self.predict_single(self.model, prepared_data) prediction_results = self.output_intent_probability(prob_distribution.astype(float)) # Filter out predictions containing "dental" or "optical" keywords filtered_results = {intent: prob for intent, prob in prediction_results.items() if "dental" not in intent.lower() and "optical" not in intent.lower() and "eye" not in intent.lower()} sorted_pred_intent_results = sorted(filtered_results.items(), key=lambda x: x[1], reverse=True) sorted_pred_intent_results_dict = dict(sorted_pred_intent_results) # Return the top result top_results = dict(list(sorted_pred_intent_results)[:1]) # temp.append(top_results) # final_preds = json.dumps(temp) final_preds = ', '.join(top_results.keys()) final_preds = final_preds.replace("'", "") return final_preds jacaranda_hugging_face_model = "Jacaranda/dhis_14000_600k_Test_Model" obj_Facility_Model = Facility_Model(facility_model_path=jacaranda_hugging_face_model, max_len=128 ) processor = Preprocess(tokenizer_vocab_path=jacaranda_hugging_face_model, tokenizer_max_len=128 ) def predict_batch_from_csv(input_file, output_file): # Load batch data from CSV batch_data = pd.read_csv(input_file) # Initialize predictions list predictions = [] # Iterate over rows with tqdm for progress tracking # Iterate over rows with tqdm for progress tracking for _, row in tqdm(batch_data.iterrows(), total=len(batch_data)): text = row['facility_name'] # Replace 'facility_name' with the actual column name containing the text data if pd.isnull(text): cleaned_text = "" else: cleaned_text = processor.clean_text(text) prepared_data = processor.process_tokenizer(cleaned_text) if cleaned_text == "": prediction = "" # Set prediction as empty string else: prediction = obj_Facility_Model.inference(prepared_data) predictions.append(prediction) # Create DataFrame for predictions output_data = pd.DataFrame({'prediction': predictions}) # Merge with input DataFrame pred_output_df = pd.concat([batch_data.reset_index(drop=True), output_data], axis=1) # Save predictions to CSV pred_output_df.to_csv(output_file, index=False)