#!/usr/bin/env python3
"""Inspect one frozen toy-transformer SAE feature over every symbol pair."""
import argparse
import csv
import hashlib
import json
import math
from datetime import datetime, timezone
from pathlib import Path

import torch
from transformer import TinyTransformer, data
from sparse_autoencoder import TopKSAE, FEATURES, capture


def sha256(path):
    return hashlib.sha256(path.read_bytes()).hexdigest()


def investigate(model, sae, feature):
    train, validation = data()
    rows = []
    for split, (tokens, _) in [('train', train), ('validation', validation)]:
        with torch.no_grad():
            residual, _ = capture(model, tokens)
            _, values = sae(residual)
        for token, value in zip(tokens, values[:, feature]):
            rows.append({'first': int(token[0]), 'second': int(token[1]),
                         'split': split, 'activation': float(value)})
    return sorted(rows, key=lambda r: (r['first'], r['second']))


def main():
    p = argparse.ArgumentParser(description='Undersøg en SAE-feature på alle symbolpar.')
    p.add_argument('--model', type=Path, required=True)
    p.add_argument('--sae-dir', type=Path, required=True)
    p.add_argument('--feature', type=int, required=True)
    p.add_argument('--symbol', type=int, required=True)
    p.add_argument('--threshold', type=float, default=0)
    p.add_argument('--output-dir', type=Path)
    args = p.parse_args()
    if not 0 <= args.feature < FEATURES or not 0 <= args.symbol < 6 or not math.isfinite(args.threshold) or not 0 <= args.threshold <= 1e6:
        p.error('Vælg feature 0–47, symbol 0–5 og en endelig tærskel fra 0 til 1.000.000.')
    if args.output_dir and args.output_dir.exists():
        p.error(f'Outputmappen findes allerede: {args.output_dir}')
    info, weights = args.sae_dir/'sae.json', args.sae_dir/'sae.pt'
    if not args.model.is_file() or not info.is_file() or not weights.is_file():
        p.error('Kontrollér stierne til model.pt og mappen med sae.json og sae.pt.')
    try:
        metadata = json.loads(info.read_text(encoding='utf-8'))
        if not isinstance(metadata, dict) or metadata.get('format') != 'learning-lab-sae-v1' or type(metadata.get('k')) is not int or not 1 <= metadata['k'] <= 16:
            raise ValueError('Ugyldig SAE-metadata.')
        model_hash, sae_hash = sha256(args.model), sha256(weights)
        if metadata.get('model_sha256') != model_hash or metadata.get('sae_sha256') != sae_hash:
            raise ValueError('Model eller SAE-vægte matcher ikke sae.json. Brug de filer, der blev gemt sammen.')
        torch.set_num_threads(1)
        model, sae = TinyTransformer(), TopKSAE(metadata['k'])
        model.load_state_dict(torch.load(args.model, map_location='cpu', weights_only=True))
        sae.load_state_dict(torch.load(weights, map_location='cpu', weights_only=True))
        model.eval(); sae.eval()
        rows = investigate(model, sae, args.feature)
        if not all(math.isfinite(r['activation']) and 0 <= r['activation'] <= 1e6 for r in rows):
            raise ValueError('Mindst én aktivering er ugyldig eller for stor.')
    except (OSError, ValueError, RuntimeError, json.JSONDecodeError) as exc:
        p.error(str(exc))
    output = args.output_dir or Path('feature-check-runs')/datetime.now(timezone.utc).strftime('%Y%m%dT%H%M%S_%fZ')
    try: output.mkdir(parents=True, exist_ok=False)
    except FileExistsError: p.error(f'Outputmappen findes allerede: {output}')
    result = {'format':'learning-lab-feature-check-v1','feature':args.feature,'symbol':args.symbol,'threshold':args.threshold,
              'k':metadata['k'],'model_sha256':model_hash,'sae_sha256':sae_hash,'torch_version':str(torch.__version__),'rows':rows}
    (output/'feature_check.json').write_text(json.dumps(result, ensure_ascii=False, indent=2, allow_nan=False), encoding='utf-8')
    with (output/'examples.csv').open('w', encoding='utf-8', newline='') as f:
        writer = csv.DictWriter(f, fieldnames=['first','second','split','activation']); writer.writeheader(); writer.writerows(rows)
    print(f'Feature {args.feature}, symbol {args.symbol}, tærskel {args.threshold:g}. Gemt i {output}: feature_check.json og examples.csv.')


if __name__ == '__main__':
    main()
