analys_data.py

import json
import os
from collections import Counter

# Get list of 22 labels from standard ANSWER_VOCAB
ANSWER_VOCAB = (
    [str(d) for d in range(10)]
    + [
        "top-left", "top-center", "top-right",
        "center-left", "center", "center-right",
        "bottom-left", "bottom-center", "bottom-right",
    ]
    + ["yes", "no", "empty"]
)

def analyze_split(split_name, dataset_path):
    if not os.path.exists(dataset_path):
        print(f"[-] Skipping {split_name.upper()}: File not found {dataset_path}")
        return

    with open(dataset_path, 'r', encoding='utf-8') as f:
        data = json.load(f)
    
    total = len(data)
    print(f"\n{'='*50}")
    print(f"📊 Analyzing {split_name.upper()} split ({total:,} questions)")
    print(f"{'='*50}")
    
    if total == 0:
        print("Empty dataset!")
        return

    # Calculate counts for all 22 labels
    counts = Counter(item.get('answer') for item in data)
    
    # Print results in the original order of ANSWER_VOCAB
    print(f"{'Label (Answer)':<18} | {'Count':<10} | {'Ratio (%)'}")
    print("-" * 50)
    for ans in ANSWER_VOCAB:
        cnt = counts.get(ans, 0)
        pct = (cnt / total) * 100
        
        # Highlight yes/no case to easily see balancing effect after fix
        if ans in ['yes', 'no']:
            print(f"{ans:<18} | {cnt:<10,} | {pct:6.2f}%  <---")
        else:
            print(f"{ans:<18} | {cnt:<10,} | {pct:6.2f}%")
        
    # Calculate overview by label group
    digit_cnt = sum(counts.get(str(i), 0) for i in range(10))
    pos_cnt = sum(counts.get(pos, 0) for pos in [
        "top-left", "top-center", "top-right",
        "center-left", "center", "center-right",
        "bottom-left", "bottom-center", "bottom-right"
    ])
    yes_cnt = counts.get("yes", 0)
    no_cnt = counts.get("no", 0)
    yes_no_cnt = yes_cnt + no_cnt
    empty_cnt = counts.get("empty", 0)
    
    print("-" * 50)
    print("📈 OVERVIEW BY LABEL GROUP:")
    print(f"- Digits Group (0-9) : {digit_cnt:<7,} ({digit_cnt/total*100:5.2f}%)")
    print(f"- Position Group     : {pos_cnt:<7,} ({pos_cnt/total*100:5.2f}%)")
    print(f"- Yes/No Group       : {yes_no_cnt:<7,} ({yes_no_cnt/total*100:5.2f}%)")
    if yes_no_cnt > 0:
        print(f"    + Yes/No Ratio   : {yes_cnt/yes_no_cnt*100:.1f}% Yes / {no_cnt/yes_no_cnt*100:.1f}% No")
    print(f"- Empty Group        : {empty_cnt:<7,} ({empty_cnt/total*100:5.2f}%)")

if __name__ == "__main__":
    # Root directory containing generated datasets
    base_dir = "mnist_vqa_dataset"
    
    print("\n🚀 STARTING MNIST-VQA-V4 DATA STATISTICAL ANALYSIS")
    analyze_split("train", os.path.join(base_dir, "train", "dataset.json"))
    analyze_split("val", os.path.join(base_dir, "val", "dataset.json"))
    analyze_split("test", os.path.join(base_dir, "test", "dataset.json"))
    print(f"\n{'='*50}")
    print("Done!")