File size: 5,754 Bytes
a94233d
 
 
e732be5
 
a94233d
e732be5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5aa95d8
 
e732be5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a94233d
 
 
 
 
 
 
 
 
 
e732be5
 
 
 
 
 
 
 
 
 
 
 
 
 
a94233d
 
 
 
e732be5
 
 
 
a94233d
 
 
e732be5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a94233d
e732be5
a94233d
e732be5
 
 
a94233d
 
e732be5
a94233d
e732be5
 
a94233d
 
e732be5
 
 
 
 
 
 
a94233d
e732be5
 
a94233d
e732be5
 
 
 
 
 
 
 
a94233d
 
 
 
 
 
e732be5
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
import os
import json
import torch
import numpy as np
from transformers import BertTokenizer
from ts.torch_handler.base_handler import BaseHandler
from sklearn.preprocessing import OneHotEncoder

import transformers
import torch
import torch.nn as nn
import torch.nn.functional as F

class AttentionPool(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.attention = nn.Linear(hidden_size, 1)
    
    def forward(self, last_hidden_state):
        attention_scores = self.attention(last_hidden_state).squeeze(-1)
        attention_weights = F.softmax(attention_scores, dim=1)
        pooled_output = torch.bmm(attention_weights.unsqueeze(1), last_hidden_state).squeeze(1)
        return pooled_output

class MultiSampleDropout(nn.Module):
    def __init__(self, dropout=0.5, num_samples=5):
        super().__init__()
        self.dropout = nn.Dropout(dropout)
        self.num_samples = num_samples
    
    def forward(self, x):
        return torch.mean(torch.stack([self.dropout(x) for _ in range(self.num_samples)]), dim=0)


class ImprovedBERTClass(nn.Module):
    def __init__(self, num_classes=13):
        super().__init__()
        self.bert = transformers.BertModel.from_pretrained('bert-base-uncased')
        self.attention_pool = AttentionPool(768)
        self.dropout = MultiSampleDropout()
        self.norm = nn.LayerNorm(768)
        self.classifier = nn.Linear(768, num_classes)
        
    def forward(self, input_ids, attention_mask, token_type_ids):
        bert_output = self.bert(input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids)
        pooled_output = self.attention_pool(bert_output.last_hidden_state)
        pooled_output = self.dropout(pooled_output)
        pooled_output = self.norm(pooled_output)
        logits = self.classifier(pooled_output)
        return logits


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