interfaces/cap_minor_media.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
import torch.nn.functional as F
from huggingface_hub import HfApi
from collections import defaultdict

from label_dicts import (CAP_MEDIA_NUM_DICT, CAP_MEDIA_LABEL_NAMES,
                        CAP_MIN_NUM_DICT, CAP_MIN_LABEL_NAMES)

from .utils import is_disk_full

HF_TOKEN = os.environ["hf_read"]

languages = [
    "Multilingual",
]

domains = {
    "media": "media"
}

NUM_TOP_CLASSES = 5
CAP_MEDIA_CODES = list(CAP_MEDIA_NUM_DICT.values())
CAP_MIN_CODES = list(CAP_MIN_NUM_DICT.values())

major_index_to_id = {i: code for i, code in enumerate(CAP_MEDIA_CODES)}
minor_id_to_index = {code: i for i, code in enumerate(CAP_MIN_CODES)}
minor_index_to_id = {i: code for i, code in enumerate(CAP_MIN_CODES)}

major_to_minor_map = defaultdict(list)
for code in CAP_MIN_CODES:
    major_id = int(str(code)[:-2])
    major_to_minor_map[major_id].append(code)
major_to_minor_map = dict(major_to_minor_map)


def normalize_probs(probs: dict) -> dict:
    total = sum(probs.values())
    return {k: v / total for k, v in probs.items()}


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):
    return ("poltextlab/xlm-roberta-large-pooled-cap-media", "poltextlab/xlm-roberta-large-pooled-cap-minor-v3")

def predict(text, major_model_id, minor_model_id, tokenizer_id, HF_TOKEN=None):
    device = torch.device("cpu")

    # Load major and minor models + tokenizer
    major_model = AutoModelForSequenceClassification.from_pretrained(
        major_model_id,
        low_cpu_mem_usage=True,
        device_map="auto",
        offload_folder="offload",
        token=HF_TOKEN
    ).to(device)

    minor_model = AutoModelForSequenceClassification.from_pretrained(
        minor_model_id,
        low_cpu_mem_usage=True,
        device_map="auto",
        offload_folder="offload",
        token=HF_TOKEN
    ).to(device)

    tokenizer = AutoTokenizer.from_pretrained(tokenizer_id)

    # Tokenize input
    inputs = tokenizer(text, max_length=256, truncation=True, padding="do_not_pad", return_tensors="pt").to(device)

    # Predict major topic
    major_model.eval()
    with torch.no_grad():
        major_logits = major_model(**inputs).logits
        major_probs = F.softmax(major_logits, dim=-1)
    major_probs_np = major_probs.cpu().numpy().flatten()
    top_major_index = int(np.argmax(major_probs_np))
    top_major_id = major_index_to_id[top_major_index]

    # Default: show major topic predictions
    filtered_probs = {
        i: float(major_probs_np[i])
        for i in np.argsort(major_probs_np)[::-1]
    }
    filtered_probs = normalize_probs(filtered_probs)
        
    output_pred = {
        f"[{major_index_to_id[k]}] {CAP_MEDIA_LABEL_NAMES[major_index_to_id[k]]}": v
        for k, v in sorted(filtered_probs.items(), key=lambda item: item[1], reverse=True)
    }

    # If eligible for minor prediction
    if top_major_id in major_to_minor_map:
        valid_minor_ids = major_to_minor_map[top_major_id]
        minor_model.eval()
        with torch.no_grad():
            minor_logits = minor_model(**inputs).logits
            minor_probs = F.softmax(minor_logits, dim=-1)
            
        
        print(minor_probs) # debug
        # Restrict to valid minor codes
        valid_indices = [minor_id_to_index[mid] for mid in valid_minor_ids if mid in minor_id_to_index]
        filtered_probs = {minor_index_to_id[i]: float(minor_probs[0][i]) for i in valid_indices}
        print(filtered_probs) # debug
        filtered_probs = normalize_probs(filtered_probs)
        print(filtered_probs) # debug
        
        output_pred = {
            f"[{top_major_id}] {CAP_MEDIA_LABEL_NAMES[top_major_id]} [{k}] {CAP_MIN_LABEL_NAMES[k]}": v
            for k, v in sorted(filtered_probs.items(), key=lambda item: item[1], reverse=True)
        }

    output_info = f'<p style="text-align: center; display: block">Prediction used <a href="https://huggingface.co/{major_model_id}">{major_model_id}</a> and <a href="https://huggingface.co/{minor_model_id}">{minor_model_id}</a>.</p>'

    interpretation_info = """
    ## How to Interpret These Values (Hierarchical Classification)

    This method returns either:

    - A list of **major (media) topic confidences**, or
    - A list of **minor topic confidences**.

    In the case of minor topics, the values are the confidences for minor topics **within a given major topic**, and they are **normalized to sum to 1**.

    **Note:** This method cannot return a mix of major and minor topic predictions.
    """

    return interpretation_info, output_pred, output_info

def predict_cap(tmp, method, text, language, domain):
    domain = domains[domain]
    major_model_id, minor_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, major_model_id, minor_model_id, tokenizer_id)

description = """
You can choose between two approaches for making predictions:

**1. Hierarchical Classification**  
First, the model predicts a **major topic**. Then, a second model selects the most probable **subtopic** from within that major topic's category.

**2. Flat Classification (single model)**  
A single model directly predicts the most relevant label from all available classes (both media and minor topics).
"""

demo = gr.Interface(
    title="CAP Media/Minor Topics Babel Demo",
    fn=predict_cap,
    inputs=[gr.Markdown(description),
            gr.Radio(
                choices=["Hierarchical Classification", "Flat Classification"],
                label="Prediction Mode",
                value="Hierarchical Classification"
            ),
            gr.Textbox(lines=6, label="Input"),
            gr.Dropdown(languages, label="Language"),
            gr.Dropdown(domains.keys(), label="Domain")],
    outputs=[gr.Markdown(), gr.Label(label="Output"), gr.Markdown()])