app.py

import jmespath
import asyncio
import json
from urllib.parse import urlencode
from typing import List, Dict
from httpx import AsyncClient, Response
from loguru import logger as log
import torch
import torch.nn as nn
from transformers import AutoModel
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import gradio as gr

client = AsyncClient(
    # enable http2
    http2=True,
    headers={
        "Accept-Language": "en-US,en;q=0.9",
        "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/96.0.4664.110 Safari/537.36",
        "Accept": "text/html,application/xhtml+xml,application/xml;q=0.9,image/webp,image/apng,*/*;q=0.8",
        "Accept-Encoding": "gzip, deflate, br",
        "content-type": "application/json"
    },
)

def parse_comments(response: Response) -> Dict:
    try:
        data = json.loads(response.text)
    except json.JSONDecodeError:
        return {"comments": [], "total_comments": 0}

    comments_data = data.get("comments", [])
    total_comments = data.get("total", 0)

    if not comments_data:
        return {"comments": [], "total_comments": total_comments}

    parsed_comments = []
    for comment in comments_data:
        result = jmespath.search(
            """{
            text: text
            }""",
            comment
        )
        parsed_comments.append(result)
    return {"comments": parsed_comments, "total_comments": total_comments}

async def scrape_comments(post_id: int, comments_count: int = 20, max_comments: int = None) -> List[Dict]:
    
    def form_api_url(cursor: int):
        base_url = "https://www.tiktok.com/api/comment/list/?"
        params = {
            "aweme_id": post_id,
            'count': comments_count,
            'cursor': cursor # the index to start from      
        }
        return base_url + urlencode(params)
    
    first_page = await client.get(form_api_url(0))
    data = parse_comments(first_page)
    comments_data = data["comments"]
    total_comments = data["total_comments"]

    if not comments_data:
        return []
    if max_comments and max_comments < total_comments:
        total_comments = max_comments

    _other_pages = [
        client.get(form_api_url(cursor=cursor))
        for cursor in range(comments_count, total_comments + comments_count, comments_count)
    ]

    for response in asyncio.as_completed(_other_pages):
        response = await response
        new_comments = parse_comments(response)["comments"]
        comments_data.extend(new_comments)
        
        # If we have reached or exceeded the maximum number of comments to scrape, stop the process
        if max_comments and len(comments_data) >= max_comments:
            comments_data = comments_data[:max_comments]
            break
    return comments_data

class SentimentClassifier(nn.Module):
    def __init__(self, n_classes):
        super(SentimentClassifier, self).__init__()
        self.bert = AutoModel.from_pretrained("vinai/phobert-base")
        self.drop = nn.Dropout(p=0.3)
        self.fc = nn.Linear(self.bert.config.hidden_size, n_classes)
        nn.init.normal_(self.fc.weight, std=0.02)
        nn.init.normal_(self.fc.bias, 0)

    def forward(self, input_ids, attention_mask):
        last_hidden_state, output = self.bert(
            input_ids=input_ids,
            attention_mask=attention_mask,
            return_dict=False # Dropout will errors if without this
        )

        x = self.drop(output)
        x = self.fc(x)
        return x

def infer(text, tokenizer, max_len=120):
    encoded_review = tokenizer.encode_plus(
        text,
        max_length=max_len,
        truncation=True,
        add_special_tokens=True,
        padding='max_length',
        return_attention_mask=True,
        return_token_type_ids=False,
        return_tensors='pt',
    )

    input_ids = encoded_review['input_ids'].to(device)
    attention_mask = encoded_review['attention_mask'].to(device)

    output = model(input_ids, attention_mask)
    _, y_pred = torch.max(output, dim=1)

    return class_names[y_pred]

async def predict_comments(video_id):
    comments = await scrape_comments(
        post_id=int(video_id),
        max_comments=2000,
        comments_count=20
    )
    predictions = []
    for comment in comments:
        text = comment['text']
        probs = infer(text, tokenizer)
        predictions.append({'comment': text, 'predictions': probs})
    
    # Tính toán tỷ lệ phần trăm của mỗi nhãn
    total_comments = len(predictions)
    label_counts = [0, 0, 0]  # Assuming there are 3 labels
    comment_off = []
    comment_hate = []
    for prediction in predictions:
        probs = prediction['predictions']
        if probs == 'CLEAN':
            label_counts[0] += 1
        elif probs == 'OFFENSIVE':
            label_counts[1] += 1
            comment_off.append(prediction['comment'])
        else :
            label_counts[2] += 1
            comment_hate.append(prediction['comment'])

    label_percentages = [count / total_comments * 100 for count in label_counts]
    results = {
        'total_comments': total_comments,
        'label_percentages': {
            'CLEAN': label_percentages[0],
            'OFFENSIVE': label_percentages[1],
            'HATE': label_percentages[2],
            'CMT OFFENSIVE': comment_off,
            'CMT HATE': comment_hate,
        }
    }

    return results

device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')

model = SentimentClassifier(n_classes=3)
model.to(device)
model.load_state_dict(torch.load('phobert_fold1.pth', map_location=torch.device('cpu')))

tokenizer = AutoTokenizer.from_pretrained("vinai/phobert-base")

class_names = ['CLEAN', 'OFFENSIVE', 'HATE']


iface = gr.Interface(
    fn=predict_comments,
    inputs="text",
    outputs="json"
)

iface.launch()