File size: 5,812 Bytes
4c5dfd9
 
d1dd25f
049099a
 
 
dd15286
d1dd25f
049099a
 
 
 
dd15286
049099a
 
dd15286
 
 
 
 
d1dd25f
dd15286
4c5dfd9
d1dd25f
4c5dfd9
 
 
dd15286
4c5dfd9
dd15286
d1dd25f
 
 
dd15286
d1dd25f
 
 
dd15286
 
 
 
 
4c5dfd9
049099a
dd15286
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d1dd25f
dd15286
 
 
 
 
 
 
 
 
 
 
 
 
 
d1dd25f
 
dd15286
d1dd25f
 
 
 
 
 
 
dd15286
 
d1dd25f
dd15286
 
d1dd25f
 
 
 
 
 
 
 
 
049099a
 
 
 
 
 
 
 
 
4c5dfd9
049099a
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
import os
import shutil
from pathlib import Path
from sentence_transformers import CrossEncoder
from nltk import sent_tokenize
import numpy as np
from huggingface_hub import try_to_load_from_cache, snapshot_download
from transformers import AutoConfig

class CrossEncoderSimilarity:
    """
    Uses a cross‑encoder to compute deep semantic similarity between mark and goods.
    Includes robust cache clearing and multiple fallback models.
    """

    def __init__(self,
                 primary_model='cross-encoder/stsb-roberta-large',
                 fallback_model='cross-encoder/stsb-distilroberta-base',
                 second_fallback='cross-encoder/ms-marco-MiniLM-L-6-v2'):
        self.primary_model_name = primary_model
        self.fallback_model_name = fallback_model
        self.second_fallback_name = second_fallback
        self._model = None
        self.current_model_name = None

    @property
    def model(self):
        """Lazy load the cross-encoder model with automatic cache clearing and fallbacks."""
        if self._model is None:
            # Try primary model
            self.current_model_name = self.primary_model_name
            self._model = self._load_model_with_retry(self.primary_model_name)
            if self._model is None:
                print(f"⚠️ Primary model failed. Attempting first fallback: {self.fallback_model_name}")
                self.current_model_name = self.fallback_model_name
                self._model = self._load_model_with_retry(self.fallback_model_name)
                if self._model is None:
                    print(f"⚠️ First fallback failed. Attempting second fallback: {self.second_fallback_name}")
                    self.current_model_name = self.second_fallback_name
                    self._model = self._load_model_with_retry(self.second_fallback_name)
                    if self._model is None:
                        raise RuntimeError("All cross-encoder models failed to load.")
        return self._model

    def _clear_cache_for_model(self, model_name):
        """
        Use huggingface_hub to find and remove the entire snapshot directory for a model.
        Returns True if something was removed, False otherwise.
        """
        # Try to get a cached file (e.g., config.json) to locate the snapshot
        cached_file = try_to_load_from_cache(
            model_name,
            filename="config.json",
            cache_dir=os.environ.get("HF_HOME")
        )
        if cached_file and cached_file != "_CACHED_NOFILE" and os.path.exists(cached_file):
            # The cached_file path is something like:
            #   /tmp/.cache/huggingface/hub/models--org--model/snapshots/abcd1234/config.json
            # We want to remove the entire snapshot directory.
            snapshot_dir = Path(cached_file).parent
            if snapshot_dir.exists() and snapshot_dir.is_dir():
                print(f"🗑️ Removing corrupted snapshot: {snapshot_dir}")
                shutil.rmtree(snapshot_dir)
                return True

        # If that didn't work, try to remove the whole model cache directory
        model_id = model_name.replace("/", "--")
        hf_home = os.environ.get("HF_HOME", "/tmp/.cache/huggingface")
        possible_paths = [
            Path(hf_home) / "hub" / f"models--{model_id}",
            Path(hf_home) / "models--{model_id}",
            Path(hf_home) / model_name.replace("/", "--"),
        ]
        for p in possible_paths:
            if p.exists():
                print(f"🗑️ Removing model cache directory: {p}")
                shutil.rmtree(p)
                return True
        return False

    def _load_model_with_retry(self, model_name):
        """Attempt to load a model, clear cache on failure, and retry with force_download."""
        try:
            print(f"Loading cross-encoder model: {model_name}")
            model = CrossEncoder(model_name, num_labels=1)
            print(f"✅ Cross-encoder model '{model_name}' loaded.")
            return model
        except Exception as e:
            print(f"❌ Error loading model '{model_name}': {e}. Attempting to clear cache...")
            if self._clear_cache_for_model(model_name):
                print("Cache cleared. Retrying model load with force_download...")
                try:
                    # Force a fresh download
                    model = CrossEncoder(model_name, num_labels=1, force_download=True)
                    print(f"✅ Cross-encoder model '{model_name}' loaded after cache clear.")
                    return model
                except Exception as e2:
                    print(f"❌ Still failed after cache clear: {e2}")
                    return None
            else:
                print("Cache directory not found. Cannot clear.")
                return None

    def similarity(self, mark, goods, return_segments=False):
        if not goods:
            return 0.0 if not return_segments else (0.0, None)
        sentences = sent_tokenize(goods)
        if not sentences:
            return 0.0 if not return_segments else (0.0, None)

        pairs = [(mark, sent) for sent in sentences]
        scores = self.model.predict(pairs)
        # Normalize (assuming stsb model output range 0-5)
        scores_norm = [min(1.0, max(0.0, s / 5.0)) for s in scores]
        max_score = max(scores_norm)
        max_idx = int(np.argmax(scores_norm))

        if return_segments:
            return max_score, sentences[max_idx]
        return max_score

    def similarity_with_explanation(self, mark, goods):
        max_score, best_sentence = self.similarity(mark, goods, return_segments=True)
        explanation = f"Highest similarity with segment: '{best_sentence}' (score: {max_score:.2f})"
        return max_score, explanation