"""Recompute Learn the Machine's exported safety-model scores; no third-party packages."""
import argparse
import json
import math
from pathlib import Path


def finite(value):
    return isinstance(value, (int, float)) and not isinstance(value, bool) and math.isfinite(value)


def verify(data):
    if not isinstance(data, dict) or data.get('format') != 'learning-lab-safety-evaluation-v1':
        raise ValueError('Ukendt filformat.')
    model = data['model']['parameters']
    w1, b1, w2, b2 = model['w1'], model['b1'], model['w2'], model['b2']
    if not (len(w1) == len(b1) == len(w2) == 4 and all(len(row) == 2 for row in w1)):
        raise ValueError('Modellens dimensioner er ugyldige.')
    if not all(finite(x) for row in w1 for x in row) or not all(finite(x) for x in [*b1, *w2, b2]):
        raise ValueError('Ugyldige modelparametre.')
    threshold, cutoff = data['evaluation']['threshold'], data['evaluation']['group_cutoff']
    if not finite(threshold) or not 0 <= threshold <= 1 or not finite(cutoff) or not 0 <= cutoff <= 2:
        raise ValueError('Ugyldig grænse.')
    pairs = data['pairs']
    if not isinstance(pairs, list) or not pairs or len(pairs) != data['evaluation']['pairs']:
        raise ValueError('Ugyldigt antal inputpar.')
    maximum = 0.0
    counts = {'near': [0, 0, 0], 'far': [0, 0, 0]}
    for i, row in enumerate(pairs):
        x1, x2, shifted, label = row['x1'], row['x2_known'], row['x2_shifted'], row['label']
        if row['index'] != i or not all(finite(x) for x in (x1, x2, shifted)) or shifted != -x2 or type(label) is not int or label not in (0, 1):
            raise ValueError('Ugyldigt eller forkert parret input.')
        group = 'near' if abs(x1) < cutoff else 'far'
        if row['group'] != group:
            raise ValueError('Gruppe stemmer ikke med grænsen.')
        correct = []
        for name, x in (('known', x2), ('shifted', shifted)):
            hidden = [math.tanh(a*x1 + b*x + c) for (a, b), c in zip(w1, b1)]
            z = b2 + sum(a*b for a, b in zip(w2, hidden))
            score = 1/(1+math.exp(-z)) if z >= 0 else math.exp(z)/(1+math.exp(z))
            saved = row['score_'+name]
            if not finite(saved) or not 0 <= saved <= 1:
                raise ValueError('Ugyldig gemt score.')
            maximum = max(maximum, abs(score-saved))
            prediction = int(score >= threshold)
            if prediction != row['prediction_'+name]:
                raise ValueError('En forudsigelse stemmer ikke med den genberegnede score.')
            correct.append(prediction == label)
        bucket = counts[group]
        bucket[0] += 1
        bucket[1] += int(correct[0] and not correct[1])
        bucket[2] += int(not correct[0] and correct[1])
    if maximum > 1e-9:
        raise ValueError(f'Scores afviger med op til {maximum:.3g}.')
    return maximum, counts


def main():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument('json_file', type=Path, help='Den eksporterede JSON-fil')
    args = parser.parse_args()
    try:
        with args.json_file.open(encoding='utf-8') as f:
            data = json.load(f)
        maximum, counts = verify(data)
    except (OSError, ValueError, KeyError, TypeError, OverflowError, json.JSONDecodeError) as error:
        parser.exit(1, f'Kunne ikke kontrollere filen: {error}\n')
    print(f"Kontrolleret {len(data['pairs'])} par; største scoreforskel {maximum:.3g}.")
    for group, (count, lost, gained) in counts.items():
        delta = f'{100*(gained-lost)/count:+.1f} procentpoint' if count else 'ikke defineret'
        print(f'{group}: {count} par, korrekt→forkert {lost}, forkert→korrekt {gained}, ændring {delta}.')
    print('Dette verificerer beregningerne i filen, ikke modellens sikkerhed eller dataenes repræsentativitet.')


if __name__ == '__main__':
    main()
