Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
| 3 |
+
import gradio as gr
|
| 4 |
+
|
| 5 |
+
MODEL_NAME = "engsoftexperimental/modelos/modelo_final_distilbert2"
|
| 6 |
+
|
| 7 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
| 8 |
+
model = AutoModelForSequenceClassification.from_pretrained(MODEL_NAME)
|
| 9 |
+
|
| 10 |
+
model.config.id2label = {0: "hate_speech", 1: "offensive", 2: "neither", 3: "spam"}
|
| 11 |
+
model.config.label2id = {"hate_speech": 0, "offensive": 1, "neither": 2, "spam": 3}
|
| 12 |
+
|
| 13 |
+
def classify_text(texto):
|
| 14 |
+
inputs = tokenizer(
|
| 15 |
+
texto,
|
| 16 |
+
return_tensors="pt",
|
| 17 |
+
truncation=True,
|
| 18 |
+
padding="max_length",
|
| 19 |
+
max_length=128
|
| 20 |
+
).to(model.device)
|
| 21 |
+
|
| 22 |
+
with torch.no_grad():
|
| 23 |
+
logits = model(**inputs).logits
|
| 24 |
+
probs = torch.softmax(logits, dim=1)[0].cpu().numpy()
|
| 25 |
+
pred = int(torch.argmax(logits, dim=1))
|
| 26 |
+
|
| 27 |
+
return {
|
| 28 |
+
"classe_prevista": model.config.id2label[pred],
|
| 29 |
+
"prob_hate_speech": float(probs[0]),
|
| 30 |
+
"prob_offensive": float(probs[1]),
|
| 31 |
+
"prob_neither": float(probs[2])
|
| 32 |
+
"spam": float(probs[3])
|
| 33 |
+
}
|
| 34 |
+
|
| 35 |
+
def gradio_predict(text):
|
| 36 |
+
result = classify_text(text)
|
| 37 |
+
return (
|
| 38 |
+
result["classe_prevista"],
|
| 39 |
+
f"{result['prob_hate_speech']:.4f}",
|
| 40 |
+
f"{result['prob_offensive']:.4f}",
|
| 41 |
+
f"{result['prob_neither']:.4f}",
|
| 42 |
+
f"{result['prob_spam']:.4f}",
|
| 43 |
+
)
|
| 44 |
+
|
| 45 |
+
demo = gr.Interface(
|
| 46 |
+
fn=gradio_predict,
|
| 47 |
+
inputs=gr.Textbox(lines=5, label="Digite ou cole um texto"),
|
| 48 |
+
outputs=[
|
| 49 |
+
gr.Label(label="Classificação Prevista"),
|
| 50 |
+
gr.Textbox(label="Probabilidade (Hate Speech)"),
|
| 51 |
+
gr.Textbox(label="Probabilidade (Offensive)"),
|
| 52 |
+
gr.Textbox(label="Probabilidade (Neither)"),
|
| 53 |
+
gr.Textbox(label="Probabilidade (Spam)")
|
| 54 |
+
],
|
| 55 |
+
title="Protótipo de Moderador de Conteúdo (HateBR + BERT)",
|
| 56 |
+
description="Ferramenta experimental utilizada na pesquisa para classificação de conteúdo em três classes: Hate Speech, Offensive, Neither e Spam."
|
| 57 |
+
)
|
| 58 |
+
|
| 59 |
+
demo.launch()
|