tasks/text.py

import random
from datetime import datetime

import numpy as np
import torch
from datasets import load_dataset
from fastapi import APIRouter, Query
from sklearn.metrics import accuracy_score
from torch.utils.data import DataLoader, Dataset
from transformers import AutoConfig, AutoModelForSequenceClassification, AutoTokenizer

from .utils.emissions import clean_emissions_data, get_space_info, tracker
from .utils.evaluation import TextEvaluationRequest

router = APIRouter()

MODEL_TYPE = "bert-mini"
DESCRIPTIONS = {
    "baseline": "baseline most common class",
    "bert-base": "bert base fine tuned on just training data, Nvidia T4 small",
    "bert-medium": "bert medium fine tuned on just training data, Nvidia T4 small",
    "bert-small": "bert small fine tuned on just training data, Nvidia T4 small",
    "bert-mini": "bert mini fine tuned on just training data, Nvidia T4 small",
    "bert-tiny": "bert tiny fine tuned on just training data, Nvidia T4 small",
}
ROUTE = "/text"


class TextDataset(Dataset):
    def __init__(self, texts, tokenizer, max_length=256):
        self.texts = texts
        self.encodings = tokenizer(
            texts,
            truncation=True,
            padding=True,
            max_length=max_length,
            return_tensors="pt",
        )

    def __getitem__(self, idx):
        item = {key: val[idx] for key, val in self.encodings.items()}
        return item

    def __len__(self) -> int:
        return len(self.texts)


def baseline_model(dataset_length: int):
    # Make random predictions (placeholder for actual model inference)
    # predictions = [random.randint(0, 7) for _ in range(dataset_length)]

    # My favorite baseline is the most common class.
    predictions = [0] * dataset_length

    return predictions


def bert_model(test_dataset: dict, model_type: str):
    print("Starting my code block.")
    texts = test_dataset["quote"]

    model_repo = f"Nonnormalizable/frugal-ai-text-{model_type}"
    print(f"Loading from model_repo: {model_repo}")
    config = AutoConfig.from_pretrained(model_repo)
    model = AutoModelForSequenceClassification.from_pretrained(model_repo)
    tokenizer = AutoTokenizer.from_pretrained(model_repo)

    if torch.cuda.is_available():
        device = torch.device("cuda")
    elif torch.backends.mps.is_available():
        device = torch.device("mps")
    else:
        device = torch.device("cpu")
    print("Using device:", device)
    model = model.to(device)
    dataset = TextDataset(texts, tokenizer=tokenizer)
    dataloader = DataLoader(dataset, batch_size=32, shuffle=False)
    model.eval()
    with torch.no_grad():
        print("Starting model run.")
        predictions = np.array([])
        for batch in dataloader:
            test_input_ids = batch["input_ids"].to(device)
            test_attention_mask = batch["attention_mask"].to(device)
            outputs = model(test_input_ids, test_attention_mask)
            p = torch.argmax(outputs.logits, dim=1)
            predictions = np.append(predictions, p.cpu().numpy())
        print("End of model run.")

    print("End of my code block.")
    return predictions


@router.post(ROUTE, tags=["Text Task"])
async def evaluate_text(
    request: TextEvaluationRequest,
    model_type: str = MODEL_TYPE,
    # This should be an API query parameter, but it looks like the submission repo
    # https://huggingface.co/spaces/frugal-ai-challenge/submission-portal
    # is built in a way to not accept any other endpoints or parameters.
):
    """
    Evaluate text classification for climate disinformation detection.

    Current Model: Random Baseline
    - Makes random predictions from the label space (0-7)
    - Used as a baseline for comparison
    """
    # Get space info
    username, space_url = get_space_info()

    # Define the label mapping
    LABEL_MAPPING = {
        "0_not_relevant": 0,
        "1_not_happening": 1,
        "2_not_human": 2,
        "3_not_bad": 3,
        "4_solutions_harmful_unnecessary": 4,
        "5_science_unreliable": 5,
        "6_proponents_biased": 6,
        "7_fossil_fuels_needed": 7,
    }

    # Load and prepare the dataset
    dataset = load_dataset(request.dataset_name)

    # Convert string labels to integers
    dataset = dataset.map(lambda x: {"label": LABEL_MAPPING[x["label"]]})

    # Split dataset
    train_test = dataset["train"].train_test_split(
        test_size=request.test_size, seed=request.test_seed
    )
    test_dataset = train_test["test"]

    # Start tracking emissions
    tracker.start()
    tracker.start_task("inference")

    # --------------------------------------------------------------------------------------------
    # YOUR MODEL INFERENCE CODE HERE
    # Update the code below to replace the random baseline by your model inference within the inference pass where the energy consumption and emissions are tracked.
    # --------------------------------------------------------------------------------------------

    true_labels = test_dataset["label"]
    if model_type == "baseline":
        predictions = baseline_model(len(true_labels))
    elif model_type[:5] == "bert-":
        predictions = bert_model(test_dataset, model_type)
    else:
        raise ValueError(model_type)

    # --------------------------------------------------------------------------------------------
    # YOUR MODEL INFERENCE STOPS HERE
    # --------------------------------------------------------------------------------------------

    # Stop tracking emissions
    emissions_data = tracker.stop_task()

    # Calculate accuracy
    accuracy = accuracy_score(true_labels, predictions)

    # Prepare results dictionary
    results = {
        "username": username,
        "space_url": space_url,
        "submission_timestamp": datetime.now().isoformat(),
        "model_description": DESCRIPTIONS[model_type],
        "accuracy": float(accuracy),
        "energy_consumed_wh": emissions_data.energy_consumed * 1000,
        "emissions_gco2eq": emissions_data.emissions * 1000,
        "emissions_data": clean_emissions_data(emissions_data),
        "api_route": ROUTE,
        "dataset_config": {
            "dataset_name": request.dataset_name,
            "test_size": request.test_size,
            "test_seed": request.test_seed,
        },
    }

    return results