File size: 4,611 Bytes
02c45ef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
156
157
158
159
160
import gradio as gr
import torch
import torch.nn as nn
import numpy as np
import pickle
import re
import os
from nltk.tokenize.toktok import ToktokTokenizer

class CoolLSTMClassifier(nn.Module):
    def __init__(self, vocabSize, embeddingDim, dimHidden, layerAmt, num_classes=2, dropout=0.3):
        super(CoolLSTMClassifier, self).__init__()
        
        self.embedding = nn.Embedding(vocabSize, embeddingDim, padding_idx=0)
        self.embedding_dropout = nn.Dropout(0.3)
        self.dimHidden = dimHidden
        
        self.lstm = nn.LSTM(
            embeddingDim, 
            dimHidden, 
            layerAmt, 
            batch_first=True,
            bidirectional=True,
            dropout=dropout if layerAmt > 1 else 0
        )
        
        self.dropout = nn.Dropout(dropout)
        self.fc = nn.Linear(dimHidden * 2, num_classes)
    
    def forward(self, x):
        embedded = self.embedding(x)
        embedded = self.embedding_dropout(embedded)
        lstm_out, (hidden, cell) = self.lstm(embedded)
        forward_hidden = hidden[-2, :, :]
        backward_hidden = hidden[-1, :, :]
        combined = torch.cat([forward_hidden, backward_hidden], dim=1)
        combined = self.dropout(combined)
        output = self.fc(combined)
        return output

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
tokenizer = ToktokTokenizer()

vocab = None
models = None
embeddingMatrix = None

def load_resources():
    global vocab, models, embeddingMatrix
    
    if vocab is not None and models is not None:
        return
    
    print("loading vocab and models...")
    
    with open('data/processed/vocab.pkl', 'rb') as f:
        vocab = pickle.load(f)
    
    embeddingMatrix = np.load('data/processed/embedding_matrix.npy')
    
    vocabSize = len(vocab)
    embeddingDim = 300
    dimHidden = 96
    layerAmt = 1
    num_classes = 2
    dropout = 0.5
    
    models = []
    for i in range(1, 6):
        model = CoolLSTMClassifier(vocabSize, embeddingDim, dimHidden, layerAmt, num_classes, dropout)
        model.load_state_dict(torch.load(f'models/ensemble_model_{i}.pth', map_location=device))
        model.embedding.weight.data.copy_(torch.from_numpy(embeddingMatrix))
        model.embedding.weight.requires_grad = False
        model = model.to(device)
        model.eval()
        models.append(model)
    
    print("models loaded")

def cleanText(text):
    if not text:
        return ""
    text = str(text)
    text = re.sub(r'<[^>]+>', '', text)
    text = ' '.join(text.split())
    return text

def cleanTokenize(text):
    text = str(text).lower()
    text = re.sub(r'[^a-z0-9\s]', '', text)
    tokens = tokenizer.tokenize(text)
    return tokens

def predict_review(text):
    load_resources()
    
    cleaned = cleanText(text)
    tokens = cleanTokenize(cleaned)
    
    if len(tokens) == 0:
        return "invalid input", 0.0, "n/a"
    
    indices = [vocab.get(token, vocab['<UNK>']) for token in tokens]
    
    maxLen = 256
    if len(indices) > maxLen:
        indices = indices[:maxLen]
    else:
        indices = indices + [vocab['<PAD>']] * (maxLen - len(indices))
    
    inpTensor = torch.LongTensor([indices]).to(device)
    
    allOutputs = []
    with torch.no_grad():
        for model in models:
            outputs = model(inpTensor)
            probs = torch.softmax(outputs, dim=1)
            allOutputs.append(probs.cpu().numpy())
    
    avgProbs = np.mean(allOutputs, axis=0)[0]
    fakeProb = avgProbs[1]
    realProb = avgProbs[0]
    
    confidence = max(fakeProb, realProb)
    
    fakeThreshold = 0.75
    realThreshold = 0.75
    
    if fakeProb >= fakeThreshold:
        prediction = "fake"
    elif realProb >= realThreshold:
        prediction = "real"
    else:
        prediction = "uncertain"
    
    return prediction, float(confidence), f"fake: {fakeProb:.3f}, real: {realProb:.3f}"

demo = gr.Interface(
    fn=predict_review,
    inputs=gr.Textbox(
        lines=5,
        placeholder="paste review text here",
        label="review text"
    ),
    outputs=[
        gr.Textbox(label="prediction"),
        gr.Number(label="confidence"),
        gr.Textbox(label="probabilities")
    ],
    title="sentinelcheck",
    description="fake review detector using ensemble lstm models (75% threshold)",
    examples=[
        ["this product is absolutely amazing! i received it for free and it changed my life completely. five stars!"],
        ["decent quality for the price. took about a week to arrive. works as expected."]
    ]
)

if __name__ == "__main__":
    demo.launch()