app.py

import torch
import torch.nn.functional as F
import gradio as gr
from transformers import BigBirdConfig, AutoTokenizer
from huggingface_hub import hf_hub_download

from bigbird_anayasa_classifier import BigBirdClassifier




REPO_ID = "FiratIsmailoglu/bigbird_anayasa_classifier"

# Load tokenizer and config
tokenizer = AutoTokenizer.from_pretrained(REPO_ID)
config = BigBirdConfig.from_pretrained(REPO_ID)

# Build model
model = BigBirdClassifier(config)

# Download weights from HF
weights_path = hf_hub_download(REPO_ID, filename="anayasa_bigbird_classifier_model.bin")
state_dict = torch.load(weights_path, map_location="cpu")
model.load_state_dict(state_dict)
model.eval()

def classify(text):
    enc = tokenizer(
        text,
        truncation=True,
        padding="max_length",
        max_length=3072,
        return_tensors="pt",
    )

    with torch.no_grad():
        logits = model(enc["input_ids"], enc["attention_mask"])
        probs = F.softmax(logits, dim=-1)[0].numpy()

    labels = config.id2label
    pred = int(probs.argmax())
    pred_label = labels[pred]

    probs_dict = {labels[i]: float(probs[i]) for i in range(len(probs))}

    return f"Predicted: **{pred_label}**", probs_dict


with gr.Blocks() as demo:
    gr.Markdown("# BigBird Tabanlı Metin Sınıflandırma")
    text_input = gr.Textbox(lines=10, label="Metni girin")
    out_label = gr.Markdown()
    out_probs = gr.Label()

    btn = gr.Button("Sınıflandır")
    btn.click(classify, text_input, [out_label, out_probs])

demo.launch()