File size: 7,393 Bytes
7be91d6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
import re
import pandas as pd
import spacy
from langdetect import detect_langs
from sklearn.feature_extraction.text import CountVectorizer, TfidfVectorizer
from sklearn.decomposition import LatentDirichletAllocation
from sklearn.feature_extraction.text import ENGLISH_STOP_WORDS
from spacy.lang.fr.stop_words import STOP_WORDS as FRENCH_STOP_WORDS
from sklearn.cluster import KMeans
from sklearn.manifold import TSNE
import numpy as np
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification, AutoConfig
import streamlit as st

# Lighter model
MODEL ="cardiffnlp/twitter-xlm-roberta-base-sentiment"

# Cache model loading with fallback for quantization
@st.cache_resource
def load_model():
    device = "cuda" if torch.cuda.is_available() else "cpu"
    print(f"Using device: {device}")
    tokenizer = AutoTokenizer.from_pretrained(MODEL, use_fast=True)
    model = AutoModelForSequenceClassification.from_pretrained(MODEL).to(device)
    
    # Attempt quantization with fallback
    try:
        # Set quantization engine explicitly (fbgemm for x86, qnnpack for ARM)
        torch.backends.quantized.engine = 'fbgemm' if torch.cuda.is_available() else 'qnnpack'
        model = torch.quantization.quantize_dynamic(model, {torch.nn.Linear}, dtype=torch.qint8)
        print("Model quantized successfully.")
    except RuntimeError as e:
        print(f"Quantization failed: {e}. Using non-quantized model.")
    
    config = AutoConfig.from_pretrained(MODEL)
    return tokenizer, model, config, device

tokenizer, model, config, device = load_model()

nlp_fr = spacy.load("fr_core_news_sm")   
nlp_en = spacy.load("en_core_web_sm")
custom_stop_words = list(ENGLISH_STOP_WORDS.union(FRENCH_STOP_WORDS))

def preprocess(text):
    if text is None:
        return ""
    if not isinstance(text, str):
        try:
            text = str(text)
        except:
            return ""
    new_text = []
    for t in text.split(" "):
        t = '@user' if t.startswith('@') and len(t) > 1 else t
        t = 'http' if t.startswith('http') else t
        new_text.append(t)
    return " ".join(new_text)

def clean_message(text):
    if not isinstance(text, str):
        return ""
    text = text.lower()
    text = text.replace("<media omitted>", "").replace("this message was deleted", "").replace("null", "")
    text = re.sub(r"http\S+|www\S+|https\S+", "", text, flags=re.MULTILINE)
    text = re.sub(r"[^a-zA-ZÀ-ÿ0-9\s]", "", text)
    return text.strip()

def lemmatize_text(text, lang):
    if lang == 'fr':
        doc = nlp_fr(text)
    else:
        doc = nlp_en(text)
    return " ".join([token.lemma_ for token in doc if not token.is_punct])

def preprocess(data):
    pattern = r"^(?P<Date>\d{1,2}/\d{1,2}/\d{2,4}),\s+(?P<Time>[\d:]+(?:\S*\s?[AP]M)?)\s+-\s+(?:(?P<Sender>.*?):\s+)?(?P<Message>.*)$"
    filtered_messages, valid_dates = [], []
    
    for line in data.strip().split("\n"):
        match = re.match(pattern, line)
        if match:
            entry = match.groupdict()
            sender = entry.get("Sender")
            if sender and sender.strip().lower() != "system":
                filtered_messages.append(f"{sender.strip()}: {entry['Message']}")
                valid_dates.append(f"{entry['Date']}, {entry['Time'].replace(' ', ' ')}")

    df = pd.DataFrame({'user_message': filtered_messages, 'message_date': valid_dates})
    df['message_date'] = pd.to_datetime(df['message_date'], format='%m/%d/%y, %I:%M %p', errors='coerce')
    df.rename(columns={'message_date': 'date'}, inplace=True)

    users, messages = [], []
    msg_pattern = r"^(.*?):\s(.*)$"
    for message in df["user_message"]:
        match = re.match(msg_pattern, message)
        if match:
            users.append(match.group(1))
            messages.append(match.group(2))
        else:
            users.append("group_notification")
            messages.append(message)

    df["user"] = users
    df["message"] = messages
    df = df[df["user"] != "group_notification"].reset_index(drop=True)
    df["unfiltered_messages"] = df["message"]
    df["message"] = df["message"].apply(clean_message)
    
    # Extract time-based features
    df['year'] = pd.to_numeric(df['date'].dt.year, downcast='integer')
    df['month'] = df['date'].dt.month_name()
    df['day'] = pd.to_numeric(df['date'].dt.day, downcast='integer')
    df['hour'] = pd.to_numeric(df['date'].dt.hour, downcast='integer')
    df['day_of_week'] = df['date'].dt.day_name()
    
    # Lemmatize messages for topic modeling
    lemmatized_messages = []
    for message in df["message"]:
        try:
            lang = detect_langs(message)
            lemmatized_messages.append(lemmatize_text(message, lang))
        except:
            lemmatized_messages.append("")
    df["lemmatized_message"] = lemmatized_messages
    
    df = df[df["message"].notnull() & (df["message"] != "")].copy()
    df.drop(columns=["user_message"], inplace=True)

    # Perform topic modeling
    vectorizer = CountVectorizer(max_df=0.95, min_df=2, stop_words=custom_stop_words)
    dtm = vectorizer.fit_transform(df['lemmatized_message'])

    # Apply LDA
    lda = LatentDirichletAllocation(n_components=5, random_state=42)
    lda.fit(dtm)

    # Assign topics to messages
    topic_results = lda.transform(dtm)
    df = df.iloc[:topic_results.shape[0]].copy()
    df['topic'] = topic_results.argmax(axis=1)

    # Store topics for visualization
    topics = []
    for topic in lda.components_:
        topics.append([vectorizer.get_feature_names_out()[i] for i in topic.argsort()[-10:]])
    print("Top words for each topic-----------------------------------------------------:")
    print(topics)
    
    return df, topics

def preprocess_for_clustering(df, n_clusters=5):
    df = df[df["lemmatized_message"].notnull() & (df["lemmatized_message"].str.strip() != "")]
    df = df.reset_index(drop=True)

    vectorizer = TfidfVectorizer(max_features=5000, stop_words='english')
    tfidf_matrix = vectorizer.fit_transform(df['lemmatized_message'])
    
    if tfidf_matrix.shape[0] < 2:
        raise ValueError("Not enough messages for clustering.")

    df = df.iloc[:tfidf_matrix.shape[0]].copy()

    kmeans = KMeans(n_clusters=n_clusters, random_state=42)
    clusters = kmeans.fit_predict(tfidf_matrix)
    
    df['cluster'] = clusters
    tsne = TSNE(n_components=2, random_state=42)
    reduced_features = tsne.fit_transform(tfidf_matrix.toarray())
    
    return df, reduced_features, kmeans.cluster_centers_


def predict_sentiment_batch(texts: list, batch_size: int = 32) -> list:
    """Predict sentiment for a batch of texts"""
    if not isinstance(texts, list):
        raise TypeError(f"Expected list of texts, got {type(texts)}")
    
    processed_texts = [preprocess(text) for text in texts]
    
    predictions = []
    for i in range(0, len(processed_texts), batch_size):
        batch = processed_texts[i:i+batch_size]
        
        inputs = tokenizer(
            batch, 
            padding=True, 
            truncation=True, 
            return_tensors="pt", 
            max_length=128
        ).to(device)
        
        with torch.no_grad():
            outputs = model(**inputs)
        
        batch_preds = outputs.logits.argmax(dim=1).cpu().numpy()
        predictions.extend([config.id2label[p] for p in batch_preds])
   
    return predictions