SignX/models/evalu_enhanced.py

# coding: utf-8
"""
Enhanced evaluation module with multiple metrics
"""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import os
import sys
sys.path.insert(0, os.path.dirname(__file__))

from utils import util, metric
from models import evalu


def eval_metrics_multi(trans, target_file, indices=None, remove_bpe=False):
    """
    Evaluate with multiple metrics: BLEU, OTEM, UTEM

    Returns:
        dict with keys: bleu4, bleu1, bleu2, bleu3, otem2, utem4
    """
    target_valid_files = util.fetch_valid_ref_files(target_file)
    if target_valid_files is None:
        return {
            'bleu4': 0.0,
            'bleu1': 0.0,
            'bleu2': 0.0,
            'bleu3': 0.0,
            'otem2': 0.0,
            'utem4': 0.0
        }

    if indices is not None:
        trans = [data[1] for data in sorted(zip(indices, trans), key=lambda x: x[0])]

    # Load references
    references = []
    for ref_file in target_valid_files:
        import tensorflow as tf
        cur_refs = tf.gfile.Open(ref_file).readlines()
        if remove_bpe:
            cur_refs = [line.replace("@@ ", "") for line in cur_refs]
        cur_refs = [line.strip().split() for line in cur_refs]
        references.append(cur_refs)

    references = list(zip(*references))

    # Process translations
    if remove_bpe:
        new_trans = []
        for line in trans:
            line = (' '.join(line)).replace('@@ ', '').split()
            new_trans.append(line)
        trans = new_trans

    # Calculate multiple metrics
    results = {}

    # BLEU-4 (default)
    results['bleu4'] = metric.bleu(trans, references, n=4)

    # BLEU-1, BLEU-2, BLEU-3
    results['bleu1'] = metric.bleu(trans, references, n=1)
    results['bleu2'] = metric.bleu(trans, references, n=2)
    results['bleu3'] = metric.bleu(trans, references, n=3)

    # OTEM-2 (Over-Translation Evaluation Metric)
    results['otem2'] = metric.otem(trans, references, n=2)

    # UTEM-4 (Under-Translation Evaluation Metric)
    results['utem4'] = metric.utem(trans, references, n=4)

    return results


def eval_metrics_full(trans, target_file, indices=None, remove_bpe=False):
    """
    Full evaluation with all available metrics from eval/metrics.py
    Requires sacrebleu and other dependencies

    Returns:
        dict with comprehensive metrics
    """
    try:
        # Import eval metrics (may fail if dependencies not installed)
        from eval import metrics as eval_metrics

        target_valid_files = util.fetch_valid_ref_files(target_file)
        if target_valid_files is None:
            return {}

        if indices is not None:
            trans = [data[1] for data in sorted(zip(indices, trans), key=lambda x: x[0])]

        # Load references as strings
        import tensorflow as tf
        references = []
        for ref_file in target_valid_files:
            cur_refs = [line.strip() for line in tf.gfile.Open(ref_file).readlines()]
            references.append(cur_refs)

        # For single reference, take first one
        references = references[0] if len(references) == 1 else references[0]

        # Convert translations to strings
        hypotheses = [' '.join(t) if isinstance(t, list) else t for t in trans]

        if remove_bpe:
            hypotheses = [h.replace('@@ ', '') for h in hypotheses]
            references = [r.replace('@@ ', '') for r in references]

        results = {}

        # BLEU scores (BLEU-1 to BLEU-4)
        bleu_scores = eval_metrics.bleu(references, hypotheses)
        results.update(bleu_scores)

        # chrF score
        results['chrf'] = eval_metrics.chrf(references, hypotheses)

        # ROUGE score
        results['rouge'] = eval_metrics.rouge(references, hypotheses)

        # Token accuracy
        results['token_accuracy'] = eval_metrics.token_accuracy(references, hypotheses)

        # Sequence accuracy
        results['sequence_accuracy'] = eval_metrics.sequence_accuracy(references, hypotheses)

        # WER and related metrics (for sign language recognition)
        wer_scores = eval_metrics.wer_list(references, hypotheses)
        results.update(wer_scores)

        return results

    except ImportError as e:
        print(f"Warning: Could not import eval.metrics module: {e}")
        print("Falling back to basic metrics...")
        return eval_metrics_multi(trans, target_file, indices, remove_bpe)