backend/run_experiments.py

import requests
import pandas as pd
import time
import os

INPUT_CSV = "dataset_test.csv" 
OUTPUT_CSV = "augmented_dataset2.csv"
API_URL = "http://127.0.0.1:8000"

# Hyperparameters (can be modified for subsequent tests)
EDA_P = 0.15          
SBERT_THRESH = 0.70   

# Complete list of configurations: (Method, Pivot_Language)
METHODS_TO_TEST = [
    ("EDA", "none"), # Lexical perturbations (pivot ignored)
    ("BT", "en"),    # Back-Translation via English
    ("BT", "de"),    # Back-Translation via German
    ("BT", "cs"),    # Back-Translation via Czech
    ("LLM", "none")  # Generative LLM paraphrasing (pivot ignored)
]

def run_mass_augmentation():
    print(f"Starting experiment for {len(METHODS_TO_TEST)} variants. S-BERT Threshold: {SBERT_THRESH*100}%")
    
    # Check if input file exists
    if not os.path.exists(INPUT_CSV):
        print(f"CRITICAL ERROR: Input file {INPUT_CSV} not found.")
        return

    df = pd.read_csv(INPUT_CSV)
    augmented_pool = []
    
    try:
        for index, row in df.iterrows():
            original_text = row['text']
            label = row['label']
            print(f"\n[{index+1}/{len(df)}] Processing: {original_text[:40]}...")
            
            for method, pivot in METHODS_TO_TEST:
                method_name = f"{method}_{pivot}" if method == "BT" else method
                
                try:
                    # 1. GENERATION PHASE
                    payload = {
                        "text": original_text, 
                        "method": method, 
                        "pivot_lang": pivot, 
                        "eda_p": EDA_P
                    }
                    aug_response = requests.post(f"{API_URL}/augment", json=payload)
                    
                    # Check if the server responded with 200 OK status
                    if aug_response.status_code != 200:
                        error_msg = aug_response.text
                        print(f"  API Rejected ({aug_response.status_code}) for {method_name}")
                        
                        # Defensive mechanism against API rate limits (e.g., Groq Cloud)
                        if "429" in error_msg or "502" in error_msg or "rate limit" in error_msg.lower():
                            print("  API Rate limit reached. Sleeping for 60 seconds...")
                            time.sleep(60) 
                        continue 

                    aug_data = aug_response.json()
                    augmented_text = aug_data.get("augmented")
                    
                    if not augmented_text or augmented_text in ["BŁĄD", "ERROR"]:
                        continue

                    # 2. S-BERT SEMANTIC FILTRATION PHASE
                    filter_payload = {
                        "original": original_text, 
                        "augmented": augmented_text, 
                        "threshold": SBERT_THRESH
                    }
                    filt_res = requests.post(f"{API_URL}/filter", json=filter_payload).json()
                    
                    sim_score = filt_res.get("similarity", 0)
                    passed = filt_res.get("passed", False)
                    
                    # 3. MEMORY BUFFER ALLOCATION
                    if passed:
                        print(f"  {method_name.ljust(6)} ACCEPTED (Sim: {sim_score:.3f})")
                        augmented_pool.append({
                            "id": f"{row['id']}_aug_{method_name}",
                            "text": augmented_text,
                            "label": label,
                            "is_synthetic": True,
                            "source_method": method_name,
                            "similarity_score": sim_score
                        })
                    else:
                        print(f"  {method_name.ljust(6)} REJECTED (Sim: {sim_score:.3f})")
                    
                except requests.exceptions.RequestException as e:
                    print(f"  Network connection error for {method_name}: {e}")
                    
                # Safe time delay between API calls to prevent throttling
                time.sleep(1.5) if method == "LLM" else time.sleep(0.3)
                
    except KeyboardInterrupt:
        print("\n\nMANUALLY INTERRUPTED! Stopping execution, initiating emergency data save...")
        
    # === DATA PERSISTENCE BLOCK ===
    if augmented_pool:
        print("\nWriting generated samples to disk...")
        new_data_df = pd.DataFrame(augmented_pool)
        
        try:
            # If the file exists, append to it; otherwise, create a new one with baseline data
            if os.path.exists(OUTPUT_CSV):
                existing_df = pd.read_csv(OUTPUT_CSV)
                final_dataset = pd.concat([existing_df, new_data_df], ignore_index=True)
                # Remove any accidental duplicates based on ID
                final_dataset = final_dataset.drop_duplicates(subset=['id'], keep='last')
            else:
                # If no previous file exists, initialize with the baseline dataset
                baseline_df = df.copy()
                baseline_df['is_synthetic'] = False
                baseline_df['source_method'] = 'ORIGINAL'
                baseline_df['similarity_score'] = 1.0
                final_dataset = pd.concat([baseline_df, new_data_df], ignore_index=True)
            
            # Save the final dataset
            final_dataset.to_csv(OUTPUT_CSV, index=False, encoding='utf-8')
            
            print(f"Success! Saved {len(new_data_df)} new samples.")
            print(f"Total dataset volume: {len(final_dataset)} records.")
        except Exception as e:
            print(f"CRITICAL: Failed to save data. Error: {e}")
    else:
        print("\nNo new accepted samples to serialize.")

if __name__ == "__main__":
    run_mass_augmentation()