tasks/text.py

from fastapi import APIRouter
from datetime import datetime
from datasets import load_dataset, Dataset
from sklearn.metrics import accuracy_score, f1_score
import random
from torch.utils.data import DataLoader

from transformers import AutoTokenizer, AutoModelForSequenceClassification, Trainer, TrainingArguments, DataCollatorWithPadding

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

import torch
import numpy as np

router = APIRouter()

DESCRIPTION = "BERT V1.1"
ROUTE = "/text"

@router.post(ROUTE, tags=["Text Task"], 
             description=DESCRIPTION)
async def evaluate_text(request: TextEvaluationRequest):
    """
    Evaluate text classification for climate disinformation detection.
    
    Current Model: BERT
    - Uses a pre-trained BERT model for sequence classification
    """
    # 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"]
    print('dataset type: ' , test_dataset.column_names)  # Debugging step
    print('dataset type: ' , test_dataset['quote'][:5])  # Debugging step

    # 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.
    #--------------------------------------------------------------------------------------------   

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model_name = ["cococli/bert-base-uncased-frugalai", 'cococli/roberta-base-frugalai', "cococli/distilbert-base-uncased-frugalai",
                    "cococli/albert-base-v2-frugalai", "cococli/bert-base-uncased-coco-frugalai",
                     "cococli/distilbert-base-uncased-coco-frugalai", "cococli/albert-base-v2-coco-frugalai","cococli/electra-small-discriminator-coco-frugalai", 
                     'cococli/roberta-base-coco-frugalai', "cococli/distilbert-base-uncased-climate-frugalai","cococli/albert-base-v2-climate-frugalai", 
                     "cococli/electra-small-discriminator-frugalai", "cococli/bert-base-uncased-climate-frugalai","cococli/roberta-base-climate-frugalai",
    ]

    
    tokenizer = AutoTokenizer.from_pretrained(model_name[0])
    model = AutoModelForSequenceClassification.from_pretrained(model[0]).to(device)
    
    # def tokenize_function(examples):
    #     return tokenizer(examples["quote"], padding=True, truncation=True, return_tensors='pt')
    # print('BEFORE TOKENIZING')
    # # Tokenize the test dataset
    # tokenized_test = test_dataset.map(tokenize_function, batched=True)
    # print('AFTER TOKENIZING')
    # print(tokenized_test.column_names)  # Debugging step
    # print(tokenized_test['input_ids'][:5])  # Debugging step

    # # Create DataLoader
    # data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
    # dataloader = DataLoader(tokenized_test, batch_size=16, shuffle=False, collate_fn=data_collator)

    print("Started prediction run")
    # tokenized_test = tokenizer(test_dataset['quote'], padding=True, truncation=True, return_tensors='pt')
    # Model inference
    model.eval()
    predictions = np.array([])
    batch_size = 32
    

    with torch.no_grad():
        for i in range(0, len(test_dataset['quote']), batch_size):
            batch_quotes = test_dataset['quote'][i:i + batch_size]
            print(f'Processing batch {i // batch_size + 1}')
            
            # Tokenize the input data for the current batch
            tokenized_inputs = tokenizer(batch_quotes, padding=True, truncation=True, return_tensors='pt').to(device)
            
            # Forward pass through the model
            p = model(**tokenized_inputs)
            output = torch.argmax(p.logits, dim=1).cpu().numpy()
            # print(p)
            predictions = np.append(predictions, output)

    print("Finished prediction run")

    # Ensure "label" column exists in dataset
    print(test_dataset.column_names)  # Debugging step

    # Extract true labels
    true_labels = test_dataset["label"]

    #--------------------------------------------------------------------------------------------
    # YOUR MODEL INFERENCE STOPS HERE
    #--------------------------------------------------------------------------------------------   
    print(predictions)
    print(true_labels)
    print('Accuracy: ', (true_labels == predictions)/len(true_labels))

    
    print('Accuracy: ', accuracy_score(true_labels, predictions))
    print('F1 SCORE: ', f1_score(true_labels, predictions))
    # 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