tasks/text.py

from fastapi import APIRouter
from datetime import datetime
from datasets import load_dataset
from sklearn.metrics import accuracy_score
import random
from transformers import pipeline, AutoConfig
import os
from concurrent.futures import ThreadPoolExecutor
from typing import List, Dict, Tuple
import numpy as np
import torch

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

# Disable torch compile
os.environ["TORCH_COMPILE_DISABLE"] = "1"

router = APIRouter()

DESCRIPTION = "Random Baseline"
ROUTE = "/text"

class TextClassifier:
    def __init__(self):
        # Add retry mechanism for model initialization
        max_retries = 3
        for attempt in range(max_retries):
            try:
                self.config = AutoConfig.from_pretrained("camillebrl/ModernBERT-envclaims-overfit")
                self.label2id = self.config.label2id
                self.classifier = pipeline(
                    "text-classification",
                    "camillebrl/ModernBERT-envclaims-overfit",
                    device="cpu",
                    batch_size=16
                )
                print("Model initialized successfully")
                break
            except Exception as e:
                if attempt == max_retries - 1:
                    raise Exception(f"Failed to initialize model after {max_retries} attempts: {str(e)}")
                print(f"Attempt {attempt + 1} failed, retrying...")
                time.sleep(1)

    def process_batch(self, batch: List[str], batch_idx: int) -> Tuple[List[int], int]:
        """Process a batch of texts and return their predictions"""
        max_retries = 3
        for attempt in range(max_retries):
            try:
                print(f"Processing batch {batch_idx} with {len(batch)} items (attempt {attempt + 1})")
                # Process texts one by one in case of errors
                predictions = []
                for text in batch:
                    try:
                        pred = self.classifier(text)
                        pred_label = self.label2id[pred[0]["label"]]
                        predictions.append(pred_label)
                    except Exception as e:
                        print(f"Error processing text in batch {batch_idx}: {str(e)}")
                
                if not predictions:
                    raise Exception("No predictions generated for batch")
                
                print(f"Completed batch {batch_idx} with {len(predictions)} predictions")
                return predictions, batch_idx
            
            except Exception as e:
                if attempt == max_retries - 1:
                    print(f"Final error in batch {batch_idx}: {str(e)}")
                    return [0] * len(batch), batch_idx  # Return default predictions instead of empty list
                print(f"Error in batch {batch_idx} (attempt {attempt + 1}): {str(e)}")
                time.sleep(1)


@router.post(ROUTE, tags=["Text Task"], 
             description=DESCRIPTION)
async def evaluate_text(request: TextEvaluationRequest):
    """
    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"]
    test_dataset = dataset["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"]
    
    # Initialize the model once
    classifier = TextClassifier()

    # Prepare batches
    batch_size = 32
    quotes = test_dataset["quote"]
    num_batches = len(quotes) // batch_size + (1 if len(quotes) % batch_size != 0 else 0)
    batches = [
        quotes[i * batch_size:(i + 1) * batch_size]
        for i in range(num_batches)
    ]

    # Initialize batch_results before parallel processing
    batch_results = [[] for _ in range(num_batches)]
    
    # Process batches in parallel
    max_workers = min(os.cpu_count(), 4)  # Limit to 4 workers or CPU count
    print(f"Processing with {max_workers} workers")
    
    with ThreadPoolExecutor(max_workers=max_workers) as executor:
        # Submit all batches for processing
        future_to_batch = {
            executor.submit(
                classifier.process_batch, 
                batch, 
                idx
            ): idx for idx, batch in enumerate(batches)
        }

        # Collect results in order
        for future in future_to_batch:
            batch_idx = future_to_batch[future]
            try:
                predictions, idx = future.result()
                if predictions:  # Only store non-empty predictions
                    batch_results[idx] = predictions
                    print(f"Stored results for batch {idx} ({len(predictions)} predictions)")
            except Exception as e:
                print(f"Failed to get results for batch {batch_idx}: {e}")
                # Use default predictions instead of empty list
                batch_results[batch_idx] = [0] * len(batches[batch_idx])

    # Flatten predictions while maintaining order
    predictions = []
    for batch_preds in batch_results:
        if batch_preds is not None:
            predictions.extend(batch_preds)

    #--------------------------------------------------------------------------------------------
    # YOUR MODEL INFERENCE STOPS HERE
    #--------------------------------------------------------------------------------------------   
    
    # Stop tracking emissions
    emissions_data = tracker.stop_task()
    
    # Calculate accuracy
    accuracy = accuracy_score(true_labels, predictions)
    print("accuracy : ", accuracy)
    
    # 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
        }
    }

    print("results : ", results)
    
    return results