File size: 7,170 Bytes
bde1c71
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
import torch
from torch.utils.data import DataLoader
from datasets import concatenate_datasets, load_dataset,Dataset
from typing import Dict, Any, List, Optional
import warnings

from ..base_evaluator import BaseEvaluator



SUPPORTED_DATASETS = {
    "tsac": {
        "path": "tunis-ai/tsac",
        "text_column": "sentence",
        "label_column": "target",
        "label_map": {0: 0, 1: 1},  # already binary
        "trust_remote_code": True,
        "split": "test"
    },

    

}


class SentimentAnalysisEvaluator(BaseEvaluator):
    """
    Unified evaluator for Tunisian sentiment analysis.
    Supports multiple datasets, harmonizes labels to binary (0=neg, 1=pos).
    Neutral/mapped-to-invalid labels are filtered out.
    """

    def __init__(
        self,
        datasets: Optional[List[str]] = None,
        max_samples_per_dataset: int = 500,
        batch_size: int = 16
    ):
        """
        Args:
            datasets: List of dataset keys from SUPPORTED_DATASETS.
                      If None, uses all available.
            max_samples_per_dataset: Limit samples per dataset for faster eval.
            batch_size: Inference batch size.
        """
        if datasets is None:
            self.dataset_keys = list(SUPPORTED_DATASETS.keys())
        else:
            for d in datasets:
                if d not in SUPPORTED_DATASETS:
                    raise ValueError(f"Dataset '{d}' not in supported list: {list(SUPPORTED_DATASETS.keys())}")
            self.dataset_keys = datasets
        
        self.max_samples_per_dataset = max_samples_per_dataset
        self.batch_size = batch_size

    @property
    def task_name(self) -> str:
        return "Sentiment Analysis"

    def load_dataset(self) -> Dataset:
        """Load and harmonize all configured sentiment datasets."""
        print("\n=== Loading Tunisian Sentiment Datasets ===")
        all_datasets = []

        for key in self.dataset_keys:
            cfg = SUPPORTED_DATASETS[key]
            print(f"\nLoading '{key}': {cfg.get('description', "No description available.")}")
            
            try:
                ds = load_dataset(
                    cfg["path"],
                    split=cfg["split"],
                    trust_remote_code=cfg.get("trust_remote_code", False)
                )
                print(f"  Raw size: {len(ds)}")
            except Exception as e:
                warnings.warn(f"Failed to load {key}: {e}. Skipping.")
                continue

            # Harmonize to {"text": str, "label": int in {0,1}}
            def harmonize(example):
                # print(cfg)
                try:
                    text = example[cfg["text_column"]]
                    orig_label = example[cfg["label_column"]]
                    
                    if orig_label not in cfg["label_map"]:
                        return None
                    
                    new_label = cfg["label_map"][orig_label]
                    if new_label not in [0, 1]:
                        return None  # skip neutral/invalid
                    
                    return {"text": text, "label": new_label}
                except Exception:
                    return None
            print("  Harmonizing and filtering...")
            ds = ds.map(
                harmonize,
                load_from_cache_file=False,
                desc=f"Harmonizing {key}"
            )
            # print(ds)
            
            
            print("  Filtering invalid/neutral samples...")
            ds = ds.filter(lambda x: x is not None, load_from_cache_file=False)
            print(f"  Valid binary samples: {len(ds)}")

            if self.max_samples_per_dataset and len(ds) > self.max_samples_per_dataset:
                ds = ds.select(range(self.max_samples_per_dataset))
                print(f"  Trimmed to {self.max_samples_per_dataset} samples")

            if len(ds) > 0:
                all_datasets.append(ds)

        if not all_datasets:
            raise ValueError("No valid sentiment data found!")

        # Combine all datasets
        combined = concatenate_datasets(all_datasets)
        print(f"\n✅ Total Tunisian sentiment samples: {len(combined)}")
        return combined

    def _tokenize_batch(self, examples, tokenizer):
        return tokenizer(
            examples["sentence"],
            padding=True,
            truncation=True,
            max_length=512,
            return_tensors=None
        )

    def _collate_fn(self, batch):
        input_ids = torch.stack([torch.tensor(b["input_ids"]) for b in batch])
        attention_mask = torch.stack([torch.tensor(b["attention_mask"]) for b in batch])
        labels = torch.tensor([b["labels"] for b in batch], dtype=torch.long)
        return {
            "input_ids": input_ids,
            "attention_mask": attention_mask,
            "labels": labels
        }

    def evaluate(self, model, tokenizer, device: str = "cuda") -> Dict[str, Any]:
        """Evaluate model on unified Tunisian sentiment task."""
        print(f"\n=== Evaluating {self.task_name} ===")
        print(f"Model: {model.__class__.__name__} | Device: {device}")
        print(f"Datasets: {self.dataset_keys}")

        # Load and prepare data
        raw_dataset = self.load_dataset()
        tokenized = raw_dataset.map(
            lambda ex: self._tokenize_batch(ex, tokenizer),
            batched=True,
            remove_columns=raw_dataset.column_names
        )
        tokenized.set_format(type="torch", columns=["input_ids", "attention_mask"])
        tokenized = tokenized.add_column("labels", raw_dataset["label"])
        print(tokenized.column_names)

        dataloader = DataLoader(
            tokenized,
            batch_size=self.batch_size,
            shuffle=False,
            collate_fn=self._collate_fn
        )

        # Inference
        model.eval()
        all_preds, all_labels = [], []

        with torch.no_grad():
            for i, batch in enumerate(dataloader):
                inputs = {
                    k: v.to(device) for k, v in batch.items()
                    if k in ["input_ids", "attention_mask"]
                }
                labels = batch["labels"].to(device)

                outputs = model(**inputs)
                logits = outputs.logits if hasattr(outputs, "logits") else outputs[0]

                if logits.dim() == 3:  # [B, L, C]
                    logits = logits[:, 0, :]

                preds = logits.argmax(dim=-1).cpu().tolist()
                trues = labels.cpu().tolist()

                all_preds.extend(preds)
                all_labels.extend(trues)

        # Metrics
        correct = sum(p == t for p, t in zip(all_preds, all_labels))
        total = len(all_preds)
        accuracy = correct / total if total > 0 else 0.0

        print(f"\n✅ {self.task_name} Results:")
        print(f"  Accuracy: {accuracy:.4f} ({correct}/{total})")

        return {
            "task": self.task_name,
            "accuracy": accuracy,
            "main_metric": accuracy,
            "total_samples": total,
            "datasets_used": self.dataset_keys
        }