interfaces/sentiment.py

import gradio as gr

import os
import torch
import numpy as np
from transformers import AutoModelForSequenceClassification
from transformers import AutoTokenizer
from huggingface_hub import HfApi

from label_dicts import MANIFESTO_LABEL_NAMES

HF_TOKEN = os.environ["hf_read"]

languages = [
    "Czech", "English", "French", "German", "Hungarian", "Polish", "Slovakian"
]
domains = {
    "parliamentary speech": "parlspeech",
}

SENTIMENT_LABEL_NAMES = {0: "Negative", 1: "No sentiment or Neutral sentiment", 2: "Positive"}


def build_huggingface_path(language: str):
    if language == "Czech" or language == "Slovakian":
        return "visegradmedia-emotion/Emotion_RoBERTa_pooled_V4"
    return "poltextlab/xlm-roberta-large-pooled-MORES"

def predict(text, model_id, tokenizer_id):
    device = torch.device("cpu")
    model = AutoModelForSequenceClassification.from_pretrained(model_id, low_cpu_mem_usage=True, device_map="auto", offload_folder="offload", token=HF_TOKEN)
    tokenizer = AutoTokenizer.from_pretrained(tokenizer_id)
    model.to(device)

    inputs = tokenizer(text,
                       max_length=256,
                       truncation=True,
                       padding="do_not_pad",
                       return_tensors="pt").to(device)
    model.eval()

    with torch.no_grad():
        logits = model(**inputs).logits

    probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten()
    predicted_class_id = probs.argmax()
    predicted_class_id = {4: 2, 5: 1}.get(predicted_class_id, 0)
    
    
    output_pred = SENTIMENT_LABEL_NAMES.get(predicted_class_id, predicted_class_id)
    
    
    output_info = f'<p style="text-align: center; display: block">Prediction was made using the <a href="https://huggingface.co/{model_id}">{model_id}</a> model.</p>'
    return output_pred, output_info

def predict_cap(text, language, domain):
    model_id = build_huggingface_path(language)
    tokenizer_id = "xlm-roberta-large"
    return predict(text, model_id, tokenizer_id)

demo = gr.Interface(
    fn=predict_cap,
    inputs=[gr.Textbox(lines=6, label="Input"),
            gr.Dropdown(languages, label="Language"),
            gr.Dropdown(domains.keys(), label="Domain")],
    outputs=[gr.Label(num_top_classes=3, label="Output"), gr.Markdown()])