handler.py

import os
import json
import torch
import numpy as np
from transformers import BertTokenizer
from ts.torch_handler.base_handler import BaseHandler
from model import ImprovedBERTClass
from sklearn.preprocessing import OneHotEncoder

class UICardMappingHandler(BaseHandler):
    def __init__(self):
        super().__init__()
        self.initialized = False

    def initialize(self, context):
        self.manifest = context.manifest
        properties = context.system_properties
        model_dir = properties.get("model_dir")
        
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        
        # Load config
        with open(os.path.join(model_dir, 'config.json'), 'r') as f:
            self.config = json.load(f)
        
        # Initialize encoder and labels
        self.labels = ['Videos', 'Unit Conversion', 'Translation', 'Shopping Product Comparison', 'Restaurants', 'Product', 'Information', 'Images', 'Gift', 'General Comparison', 'Flights', 'Answer', 'Aircraft Seat Map']
        labels_np = np.array(self.labels).reshape(-1, 1)
        self.encoder = OneHotEncoder(sparse_output=False)
        self.encoder.fit(labels_np)
        
        # Load model
        self.model = ImprovedBERTClass()
        self.model.load_state_dict(torch.load(os.path.join(model_dir, 'model.pth'), map_location=self.device))
        self.model.to(self.device)
        self.model.eval()
        
        # Load tokenizer
        self.tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
        
        self.initialized = True

    def preprocess(self, data):
        text = data[0].get("body").get("text", "")
        k = data[0].get("body").get("k", 3)
        
        inputs = self.tokenizer.encode_plus(
            text,
            add_special_tokens=True,
            max_length=64,
            padding='max_length',
            return_tensors='pt',
            truncation=True
        )
        
        return {
            "ids": inputs['input_ids'].to(self.device, dtype=torch.long),
            "mask": inputs['attention_mask'].to(self.device, dtype=torch.long),
            "token_type_ids": inputs['token_type_ids'].to(self.device, dtype=torch.long),
            "k": k
        }

    def inference(self, data):
        with torch.no_grad():
            outputs = self.model(data["ids"], data["mask"], data["token_type_ids"])
        probabilities = torch.sigmoid(outputs)
        return probabilities.cpu().detach().numpy().flatten(), data["k"]

    def postprocess(self, inference_output):
        probabilities, k = inference_output
        
        # Get top k predictions
        top_k_indices = np.argsort(probabilities)[-k:][::-1]
        top_k_probs = probabilities[top_k_indices]
        
        # Create one-hot encodings for top k indices
        top_k_one_hot = np.zeros((k, len(probabilities)))
        for i, idx in enumerate(top_k_indices):
            top_k_one_hot[i, idx] = 1
        
        # Decode the top k predictions
        top_k_cards = [self.decode_vector(one_hot.reshape(1, -1)) for one_hot in top_k_one_hot]
        
        # Create a list of tuples (card, probability) for top k predictions
        top_k_predictions = list(zip(top_k_cards, top_k_probs.tolist()))
        
        # Determine the most likely card
        predicted_labels = (probabilities > 0.5).astype(int)
        if sum(predicted_labels) == 0:
            most_likely_card = "Answer"
        else:
            most_likely_card = self.decode_vector(predicted_labels.reshape(1, -1))
        
        # Prepare the response
        result = {
            "most_likely_card": most_likely_card,
            "top_k_predictions": top_k_predictions
        }
        
        return [result]

    def decode_vector(self, vector):
        original_label = self.encoder.inverse_transform(vector)
        return original_label[0][0]  # Returns the label as a string

    def handle(self, data, context):
        self.context = context
        data = self.preprocess(data)
        data = self.inference(data)
        data = self.postprocess(data)
        return data