|
import gradio as gr |
|
from transformers import AutoModelForSequenceClassification, AutoTokenizer |
|
import torch |
|
|
|
|
|
model_name = "fohake/cert" |
|
tokenizer = AutoTokenizer.from_pretrained(model_name) |
|
model = AutoModelForSequenceClassification.from_pretrained(model_name) |
|
|
|
|
|
def predict(text): |
|
inputs = tokenizer(text, return_tensors="pt") |
|
with torch.no_grad(): |
|
outputs = model(**inputs) |
|
logits = outputs.logits |
|
probabilities = torch.nn.functional.softmax(logits, dim=-1) |
|
predicted_class = torch.argmax(probabilities, dim=-1).item() |
|
confidence = probabilities[0][predicted_class].item() |
|
return {"class": predicted_class, "confidence": confidence} |
|
|
|
|
|
iface = gr.Interface( |
|
fn=predict, |
|
inputs=gr.inputs.Textbox(lines=2, placeholder="Enter text here..."), |
|
outputs="json", |
|
title="Text Classification with CERT", |
|
description="Enter a piece of text to classify it using the CERT model." |
|
) |
|
|
|
if __name__ == "__main__": |
|
iface.launch() |
|
|