ISGDModelDemo / main.py
KagChi's picture
chore: update matrix
ebd3a5d verified
Raw History Blame Contribute Delete
10.7 kB
import io
import os
from fastapi import FastAPI, HTTPException, Request
from fastapi.staticfiles import StaticFiles
from fastapi.templating import Jinja2Templates
from fastapi.middleware.gzip import GZipMiddleware
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
import pickle
import torch
import torch.nn.functional as F
import csv
from transformers import AutoModelForSequenceClassification, AutoTokenizer
from datasets import load_dataset
import emoji
from wordcloud import WordCloud
from fastapi.responses import StreamingResponse
import matplotlib.pyplot as plt
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.model_selection import train_test_split
from sklearn.naive_bayes import MultinomialNB
from sklearn.svm import SVC
from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay
app = FastAPI()
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
# Global cache
hf_dataset_cache = None
@app.on_event("startup")
async def load_hf_dataset_once():
global hf_dataset_cache
try:
hf_dataset_cache = load_dataset("KagChi/indonesian-gambling-words", split="test")
print("Hugging Face dataset loaded and cached.")
except Exception as e:
print(f"Failed to load HF dataset on startup: {e}")
hf_dataset_cache = None
app.mount("/static",StaticFiles(directory="static"),name="static")
app.add_middleware(GZipMiddleware)
templates = Jinja2Templates(directory="templates")
original_torch_load = torch.load
def torch_load_cpu(*args, **kwargs):
if 'map_location' not in kwargs:
kwargs['map_location'] = torch.device('cpu')
return original_torch_load(*args, **kwargs)
torch.load = torch_load_cpu
pkl_model_path = "models/model.pkl"
pkl_tokenizer_path = "models/tokenizer.pkl"
use_pickle = os.path.exists(pkl_model_path) and os.path.exists(pkl_tokenizer_path)
if use_pickle:
print("Loading model from pickle files...")
with open(pkl_model_path, "rb") as f:
model = pickle.load(f)
with open(pkl_tokenizer_path, "rb") as f:
tokenizer = pickle.load(f)
else:
print("Loading model from Hugging Face...")
model_name = "KagChi/indonesian-sentiment-gambling-classification"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
model.eval()
# Pydantic models for request/response
class PredictionInput(BaseModel):
text: str # Text input for the model
class PredictionResponse(BaseModel):
prediction: int # Predicted class (-1: unknown, 0: not gambling, 1: gambling)
prediction_text: str # Human readable prediction
probabilities: dict # Probabilities for each class
confidence: float # Probability for the predicted class
status: str # "judi", "bukan judi", or "ragu-ragu"
highlighted_tokens: list[str]
def get_important_tokens(text: str, tokenizer, model, target_class=1, top_k=3):
model.eval()
# Tokenize dengan offset mapping untuk kata asli
encoding = tokenizer(
text,
return_tensors="pt",
return_offsets_mapping=True,
padding=True,
truncation=True,
max_length=128
)
input_ids = encoding["input_ids"].to(model.device)
attention_mask = encoding["attention_mask"].to(model.device)
offset_mapping = encoding["offset_mapping"][0].tolist()
# Dapatkan embeddings
embedding_layer = model.get_input_embeddings()
input_embeddings = embedding_layer(input_ids)
input_embeddings.retain_grad()
# Forward pass
outputs = model(inputs_embeds=input_embeddings, attention_mask=attention_mask)
probs = F.softmax(outputs.logits, dim=-1)
target_score = probs[0][target_class]
# Backward pass
model.zero_grad()
target_score.backward()
# Importance score per token
grads = input_embeddings.grad.abs().sum(dim=-1).squeeze().tolist()
# Mapping: offset range → importance score
token_scores = []
for i, (start, end) in enumerate(offset_mapping):
if start == end: # Special tokens
continue
token_scores.append((start, end, grads[i]))
# Gabungkan berdasarkan overlap character range
from collections import defaultdict
char_score_map = defaultdict(float)
for start, end, score in token_scores:
for i in range(start, end):
char_score_map[i] += score
# Bagi teks ke kata
import re
word_spans = [(m.start(), m.end(), m.group()) for m in re.finditer(r'\S+', text)]
word_scores = []
for start, end, word in word_spans:
score = sum(char_score_map[i] for i in range(start, end))
word_scores.append((word, score))
# Ambil kata dengan skor tertinggi
word_scores.sort(key=lambda x: x[1], reverse=True)
highlighted_words = [w for w, s in word_scores[:top_k]]
return highlighted_words
@app.get("/")
def read_root(request: Request):
return templates.TemplateResponse(
request=request, name="index.html", context={"id": id}
)
@app.get("/visualize")
async def visualize(request: Request):
return templates.TemplateResponse("visualize.html", {"request": request})
@app.get("/wordcloud")
async def generate_wordcloud():
global hf_dataset_cache
if hf_dataset_cache:
# Filter: only samples with text AND label == 1 (Gambling)
filtered = [
sample["text"]
for sample in hf_dataset_cache
if sample.get("text") and sample.get("label") == 1 and sample["text"].strip()
]
text_corpus = " ".join(filtered)
else:
text_corpus = ""
# fallback if empty
if not text_corpus.strip():
text_corpus = "judi bola togel slot kasino poker"
wordcloud = WordCloud(width=800, height=400, background_color='white').generate(text_corpus)
buf = io.BytesIO()
plt.figure(figsize=(10, 5))
plt.imshow(wordcloud, interpolation='bilinear')
plt.axis("off")
plt.tight_layout()
plt.savefig(buf, format='png')
buf.seek(0)
return StreamingResponse(buf, media_type="image/png")
@app.get("/confusion-matrix/{model_type}")
async def confusion_matrix_image(model_type: str):
global hf_dataset_cache
if hf_dataset_cache is None:
raise HTTPException(status_code=500, detail="Dataset not loaded")
# Include all labeled samples (both 0 and 1)
filtered_samples = [
sample for sample in hf_dataset_cache
if sample.get("text") and sample["text"].strip()
]
if not filtered_samples:
raise HTTPException(status_code=400, detail="No valid samples found.")
texts = [sample["text"] for sample in filtered_samples]
labels = [sample["label"] for sample in filtered_samples]
vectorizer = CountVectorizer()
X = vectorizer.fit_transform(texts)
y = labels
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
if model_type == "naive-bayes":
clf = MultinomialNB()
elif model_type == "svm":
clf = SVC()
else:
raise HTTPException(status_code=400, detail="Invalid model type (use 'naive-bayes' or 'svm')")
clf.fit(X_train, y_train)
y_pred = clf.predict(X_test)
cm = confusion_matrix(y_test, y_pred, labels=[0, 1])
disp = ConfusionMatrixDisplay(confusion_matrix=cm, display_labels=["Bukan Judi", "Judi"])
buf = io.BytesIO()
disp.plot(cmap='Blues')
plt.tight_layout()
plt.savefig(buf, format="png")
plt.close()
buf.seek(0)
return StreamingResponse(buf, media_type="image/png")
@app.post("/predict", response_model=PredictionResponse)
async def predict(input_data: PredictionInput):
try:
threshold_low = 0.4
threshold_high = 0.7
text = str(input_data.text) if input_data.text else ""
text = emoji.replace_emoji(text, replace='').strip()
# Tokenize input text
inputs = tokenizer(
text,
return_tensors="pt",
padding=True,
truncation=True,
max_length=128
)
inputs = {key: val.to(device) for key, val in inputs.items()}
# Make prediction
with torch.no_grad():
outputs = model(**inputs)
probs = F.softmax(outputs.logits, dim=-1)
prob_0 = probs[0][0].item()
prob_1 = probs[0][1].item()
# Interpret prediction
if prob_1 >= threshold_high:
prediction = 1
prediction_text = f'judi ({prob_1:.2f})'
status = "judi"
confidence = prob_1
elif prob_1 <= threshold_low:
prediction = 0
prediction_text = f'bukan judi ({prob_0:.2f})'
status = "bukan judi"
confidence = prob_0
else:
prediction = -1
prediction_text = f'ragu-ragu ({prob_1:.2f})'
status = "ragu-ragu"
confidence = max(prob_0, prob_1)
highlighted_tokens = get_important_tokens(text, tokenizer, model, target_class=1)
return PredictionResponse(
prediction=prediction,
prediction_text=prediction_text,
probabilities={"prob_bukan_judi": prob_0, "prob_judi": prob_1},
confidence=confidence,
status=status,
highlighted_tokens=highlighted_tokens
)
except Exception as e:
raise HTTPException(status_code=400, detail=f"Prediction error: {str(e)}")
@app.get("/predict")
async def predict_get(request: Request):
"""GET endpoint to render prediction form"""
return templates.TemplateResponse(
request=request, name="predict.html", context={}
)
@app.get("/samples")
async def get_samples():
file_path = "static/samples.csv"
text_field = "text"
# Try loading from local CSV first
try:
if os.path.exists(file_path):
with open(file_path, newline='', encoding='utf-8') as csvfile:
reader = csv.DictReader(csvfile)
samples = [row["text"] for row in reader if row.get("text")]
return samples
except Exception as e:
# Log CSV read failure and fallback
print(f"Failed to read CSV: {e}")
# Fallback to Hugging Face dataset
global hf_dataset_cache
if hf_dataset_cache is not None:
samples = [example[text_field] for example in hf_dataset_cache if text_field in example]
return samples
else:
raise HTTPException(status_code=500, detail="Hugging Face dataset is not available.")