Inference code for trsy

#1
by SandeepBA - opened
Files changed (1) hide show
  1. trsy_inference.py +91 -0
trsy_inference.py ADDED
@@ -0,0 +1,91 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Imports
2
+ from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
3
+ import torch
4
+ from accelerate import Accelerator
5
+ from pathlib import Path
6
+ from datetime import datetime
7
+
8
+ # --- Author and Date ---
9
+ # Author: Sandeep Belgavi
10
+ # Date: July 24, 2025
11
+
12
+ # Initialize Accelerator for distributed training and mixed precision operations.
13
+ accelerator = Accelerator()
14
+
15
+ # --- Configuration ---
16
+ # Set the path to the directory where your fine-tuned model checkpoint was saved.
17
+ model_path = Path("/home/sandeep/trsy/checkpoint")
18
+
19
+ # Ensure the specified model directory exists.
20
+ if not model_path.exists():
21
+ print(f"Error: Model directory not found at '{model_path}'. Please ensure the path is correct and accessible.")
22
+ exit()
23
+
24
+ # --- Model Loading ---
25
+ print(f"Loading model and tokenizer from: {model_path}")
26
+ try:
27
+ start_time = datetime.now()
28
+ tokenizer = AutoTokenizer.from_pretrained(model_path)
29
+ model = AutoModelForSeq2SeqLM.from_pretrained(model_path)
30
+ end_time = datetime.now()
31
+ print(f"Model and tokenizer loaded successfully in {end_time - start_time}.")
32
+ except Exception as e:
33
+ print(f"Error loading model from checkpoint: {e}")
34
+ exit()
35
+
36
+ # Prepare the model for inference using Accelerator.
37
+ model = accelerator.prepare(model)
38
+
39
+ # Set the model to evaluation mode.
40
+ model.eval()
41
+
42
+ # --- Inference Function ---
43
+ def predict_poi(address: str) -> str:
44
+ """
45
+ Predicts the Point of Interest (POI) name for a given address string using the fine-tuned model.
46
+ Args:
47
+ address (str): The input address string for which to predict the POI.
48
+ Example: "Opposite SBI, BTM Layout, Bangalore"
49
+ Returns:
50
+ str: The predicted POI name as a string.
51
+ Example: "SBI"
52
+ """
53
+ inputs = tokenizer(address, return_tensors="pt", padding=True, truncation=True)
54
+ inputs = {k: v.to(accelerator.device) for k, v in inputs.items()}
55
+
56
+ with torch.no_grad():
57
+ output = model.generate(
58
+ **inputs,
59
+ max_length=32,
60
+ num_beams=4,
61
+ early_stopping=True
62
+ )
63
+
64
+ return tokenizer.decode(output[0], skip_special_tokens=True)
65
+
66
+ # --- Inference Examples ---
67
+ print("\n--- Starting Inference Examples ---")
68
+
69
+ test_address_1 = "Opposite SBI, BTM Layout, Bangalore"
70
+ predicted_poi_1 = predict_poi(test_address_1)
71
+ print(f"Input Address: '{test_address_1}'")
72
+ print(f"Predicted POI: '{predicted_poi_1}'")
73
+ print("-" * 20)
74
+
75
+ test_address_2 = "Near HSR Layout bus stand, HSR Layout, Bengaluru"
76
+ predicted_poi_2 = predict_poi(test_address_2)
77
+ print(f"Input Address: '{test_address_2}'")
78
+ print(f"Predicted POI: '{predicted_poi_2}'")
79
+ print("-" * 20)
80
+
81
+ test_address_3 = "Next to The Forum Mall, Koramangala, Bangalore"
82
+ predicted_poi_3 = predict_poi(test_address_3)
83
+ print(f"Input Address: '{test_address_3}'")
84
+ print(f"Predicted POI: '{predicted_poi_3}'")
85
+ print("-" * 20)
86
+
87
+ test_address_4 = "Indiranagar 100 Feet Road"
88
+ predicted_poi_4 = predict_poi(test_address_4)
89
+ print(f"Input Address: '{test_address_4}'")
90
+ print(f"Predicted POI: '{predicted_poi_4}'")
91
+ print("-" * 20)