interfaces/cap.py

import gradio as gr

import os
import time
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 CAP_NUM_DICT, CAP_LABEL_NAMES

from .utils import is_disk_full, release_model

HF_TOKEN = os.environ["hf_read"]

languages = [
    "English",
    "Multilingual"
]

domains = {
    "media": "media",
    "social media": "social",
    "parliamentary speech": "parlspeech",
    "legislative documents": "legislative",
    "executive speech": "execspeech",
    "executive order": "execorder",
    "party programs": "party",
    "judiciary": "judiciary",
    "budget": "budget",
    "public opinion": "publicopinion",
    "local government agenda": "localgovernment"
}


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(language: str, domain: str):
    language = language.lower()
    base_path = "xlm-roberta-large"
    lang_path = f"poltextlab/{base_path}-{language}-cap-v3"

    # some custom mapping
    if language in ["english"]:
        if domain in ["media", "legislative"]:
            return f"poltextlab/{base_path}-{language}-{domain}-cap-v4"
        elif domain in ["social"]:
            return f"poltextlab/{base_path}-{language}-{domain}-cap-v3"
        return lang_path
    
    else:
        if domain in ["social"]:
            return f"poltextlab/{base_path}-{domain}-cap-v3"
        return "poltextlab/xlm-roberta-large-pooled-cap-v3"
        

def predict(text, model_id, tokenizer_id):
    device = torch.device("cpu")

    t0 = time.perf_counter()
    jit_model_path = f"/data/jit_models/{model_id.replace('/', '_')}.pt"
    model = torch.jit.load(jit_model_path).to(device)
    model.eval()
    print(f"Model loading: {time.perf_counter() - t0:.3f}s")

    t1 = time.perf_counter()
    tokenizer = AutoTokenizer.from_pretrained(tokenizer_id)
    print(f"Tokenizer loading: {time.perf_counter() - t1:.3f}s")

    t2 = time.perf_counter()
    inputs = tokenizer(
        text,
        max_length=64,
        truncation=True,
        padding=True,
        return_tensors="pt"
    )
    inputs = {k: v.to(device) for k, v in inputs.items()}
    print(f"Tokenization: {time.perf_counter() - t2:.3f}s")

    t3 = time.perf_counter()
    with torch.no_grad():
        output = model(inputs["input_ids"], inputs["attention_mask"])
        logits = output["logits"]
    print(f"Inference: {time.perf_counter() - t3:.3f}s")

    release_model(model, model_id)

    t4 = time.perf_counter()
    probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten()
    output_pred = {f"[{CAP_NUM_DICT[i]}] {CAP_LABEL_NAMES[CAP_NUM_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>'
    print(f"Post-processing: {time.perf_counter() - t4:.3f}s")

    return output_pred, output_info
    

def predict_cap(text, language, domain):
    print(domain) # debug statement
    domain = domains[domain]
    model_id = build_huggingface_path(language, domain)
    tokenizer_id = "xlm-roberta-large"
    
    if is_disk_full():
        os.system('rm -rf /data/models*')
        os.system('rm -r ~/.cache/huggingface/hub')
        
    return predict(text, model_id, tokenizer_id)
    

demo = gr.Interface(
    title="CAP Babel Demo",
    fn=predict_cap,
    inputs=[gr.Textbox(lines=6, label="Input"),
            gr.Dropdown(languages, label="Language", value=languages[-1]),
            gr.Dropdown(domains.keys(), label="Domain", value=list(domains.keys())[0])],
    outputs=[gr.Label(num_top_classes=5, label="Output"), gr.Markdown()])