interfaces/illframes.py

import gradio as gr

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

from label_dicts import ILLFRAMES_MIGRATION_LABEL_NAMES, ILLFRAMES_COVID_LABEL_NAMES

HF_TOKEN = os.environ["hf_read"]

languages = [
    "English"
]

domains = {
    "Covid": "covid",
    "Migration": "migration"
}

def check_huggingface_path(checkpoint_path: str):
    try:
        hf_api = HfApi(token=HF_TOKEN)
        hf_api.model_info(checkpoint_path, token=HF_TOKEN)
        return True
    except:
        return False

def build_huggingface_path(domain: str):
    return f"poltextlab/xlm-roberta-large-english-ILLFRAMES-{domain}"

def predict(text, model_id, tokenizer_id, label_names):
    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)

    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()

    NUMS_DICT = {i: key for i, key in enumerate(sorted(label_names.keys()))}

    output_pred = {f"[{NUMS_DICT[i]}] {label_names[NUMS_DICT[i]]}": probs[i] for i in np.argsort(probs)[::-1]}
    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_illframes(text, domain):
    domain = domains[domain]
    model_id = build_huggingface_path(domain)
    tokenizer_id = "xlm-roberta-large"

    if domain == "migration":
        label_names = ILLFRAMES_MIGRATION_LABEL_NAMES
    else:
        label_names = ILLFRAMES_COVID_LABEL_NAMES

    return predict(text, model_id, tokenizer_id, label_names)

demo = gr.Interface(
    fn=predict_illframes,
    inputs=[gr.Textbox(lines=6, label="Input"),
            gr.Dropdown(languages, label="Language"),
            gr.Dropdown(domains.keys(), label="Domain")],
    outputs=[gr.Label(num_top_classes=5, label="Output"), gr.Markdown()])