tasks/text.py

from fastapi import APIRouter
from datetime import datetime
from datasets import load_dataset
from sklearn.metrics import accuracy_score
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification, AutoConfig
from torch.utils.data import DataLoader
from transformers import DataCollatorWithPadding

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

router = APIRouter()

DESCRIPTION = "Climate Guard Toxic Agent is a ModernBERT for Climate Disinformation Detection"
ROUTE = "/text"

@router.post(ROUTE, tags=["Text Task"], 
             description=DESCRIPTION)
async def evaluate_text(request: TextEvaluationRequest):
    """
    Evaluate text classification for climate disinformation detection using ModernBERT.
    """
    # 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"]]})

    # Get test dataset
    test_dataset = dataset["test"]
    
    # Start tracking emissions
    tracker.start()
    tracker.start_task("inference")

    #--------------------------------------------------------------------------------------------
    # MODEL INFERENCE CODE
    #--------------------------------------------------------------------------------------------   
    
    try:
        # Set device
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        
        # Model and tokenizer paths
        model_name = "answerdotai/ModernBERT-base"
        
        # Load config and modify it
        config = AutoConfig.from_pretrained(model_name)
        config.num_labels = 8
        config.id2label = {
            "0": "0_not_relevant",
            "1": "1_not_happening",
            "2": "2_not_human",
            "3": "3_not_bad",
            "4": "4_solutions_harmful_unnecessary",
            "5": "5_science_is_unreliable",
            "6": "6_proponents_biased",
            "7": "7_fossil_fuels_needed"
        }
        config.label2id = {v: int(k) for k, v in config.id2label.items()}
        config.problem_type = "single_label_classification"
        
        # Load tokenizer
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        
        # Load model with modified config
        model = AutoModelForSequenceClassification.from_pretrained(
            "Tonic/climate-guard-toxic-agent",
            config=config,
            torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
            trust_remote_code=True
        ).to(device)
        
        # Set model to evaluation mode
        model.eval()
        
        # Preprocess function
        def preprocess_function(examples):
            return tokenizer(
                examples["quote"],
                padding=False,
                truncation=True,
                max_length=512,
                return_tensors=None
            )
        
        # Tokenize dataset
        tokenized_test = test_dataset.map(
            preprocess_function,
            batched=True,
            remove_columns=test_dataset.column_names
        )
        
        # Set format for pytorch
        tokenized_test.set_format("torch")
        
        # Create DataLoader
        data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
        test_loader = DataLoader(
            tokenized_test,
            batch_size=16,
            collate_fn=data_collator,
            shuffle=False
        )
        
        # Get predictions
        predictions = []
        with torch.no_grad():
            for batch in test_loader:
                # Move batch to device
                batch = {k: v.to(device) for k, v in batch.items()}
                
                # Get model outputs
                outputs = model(**batch)
                preds = torch.argmax(outputs.logits, dim=-1)
                
                # Add batch predictions to list
                predictions.extend(preds.cpu().numpy().tolist())
                
        # Clean up GPU memory
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
            
    except Exception as e:
        print(f"Error during model inference: {str(e)}")
        raise

    #--------------------------------------------------------------------------------------------
    # MODEL INFERENCE ENDS HERE
    #--------------------------------------------------------------------------------------------   

    # Stop tracking emissions
    emissions_data = tracker.stop_task()
    
    # Calculate accuracy
    accuracy = accuracy_score(test_dataset["label"], predictions)
    
    # Prepare results dictionary
    results = {
        "username": username,
        "space_url": space_url,
        "submission_timestamp": datetime.now().isoformat(),
        "model_description": DESCRIPTION,
        "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