File size: 8,835 Bytes
a7d24e3
 
 
 
 
 
a503604
a7d24e3
 
a503604
 
 
 
 
 
 
 
 
 
a7d24e3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a503604
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a7d24e3
a503604
 
 
 
a7d24e3
a503604
a7d24e3
a503604
a7d24e3
a503604
 
 
 
a7d24e3
 
 
 
 
 
a503604
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a7d24e3
a503604
 
 
 
a7d24e3
a503604
a7d24e3
a503604
a7d24e3
a503604
 
 
 
a7d24e3
 
06eadfa
 
 
 
 
 
 
 
700ea8e
 
 
06eadfa
 
 
 
 
 
700ea8e
 
 
 
 
06eadfa
 
 
 
700ea8e
 
06eadfa
700ea8e
 
06eadfa
 
 
 
 
 
 
 
 
 
a7d24e3
 
 
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
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
"""
Translation service - handles model loading and translation logic
"""

import os
import time
import warnings
from typing import Tuple, Optional
from huggingface_hub import hf_hub_download

# Handle NumPy compatibility issues
try:
    import numpy as np
    # Suppress NumPy 2.0 warnings for compatibility
    warnings.filterwarnings("ignore", message=".*copy.*", category=np.VisibleDeprecationWarning)
    warnings.filterwarnings("ignore", message=".*copy.*", category=UserWarning)
except ImportError:
    pass

import ctranslate2
import sentencepiece as spm
import fasttext

from ..core.config import settings
from ..core.logging import get_logger

logger = get_logger()

# Global model instances
lang_model: Optional[fasttext.FastText._FastText] = None
sp_model: Optional[spm.SentencePieceProcessor] = None
translator: Optional[ctranslate2.Translator] = None


def get_model_paths() -> Tuple[str, str, str]:
    """Get model paths from HuggingFace cache (models pre-downloaded in Docker)"""
    logger.info("loading_models_from_cache")

    try:
        # Check if we're in offline mode (Docker environment)
        offline_mode = os.environ.get("HF_HUB_OFFLINE", "0") == "1"

        if offline_mode:
            logger.info("running_in_offline_mode")
            # In offline mode, models are already downloaded and cached
            spm_path = hf_hub_download(
                repo_id=settings.model_repo_id,
                filename="spm.model",
                local_files_only=True
            )

            ft_path = hf_hub_download(
                repo_id=settings.model_repo_id,
                filename="lid218e.bin",
                local_files_only=True
            )

            # Get the translation model path
            model_bin_path = hf_hub_download(
                repo_id=settings.model_repo_id,
                filename=f"translation_models/{settings.translation_model}/model.bin",
                local_files_only=True
            )

            ct_model_full_path = os.path.dirname(model_bin_path)

        else:
            logger.info("running_in_online_mode")
            # Online mode - download models (for local development)
            spm_path = hf_hub_download(
                repo_id=settings.model_repo_id,
                filename="spm.model"
            )

            ft_path = hf_hub_download(
                repo_id=settings.model_repo_id,
                filename="lid218e.bin"
            )

            # Download all necessary CTranslate2 files
            model_bin_path = hf_hub_download(
                repo_id=settings.model_repo_id,
                filename=f"translation_models/{settings.translation_model}/model.bin"
            )

            hf_hub_download(
                repo_id=settings.model_repo_id,
                filename=f"translation_models/{settings.translation_model}/config.json"
            )

            hf_hub_download(
                repo_id=settings.model_repo_id,
                filename=f"translation_models/{settings.translation_model}/shared_vocabulary.txt"
            )

            ct_model_full_path = os.path.dirname(model_bin_path)

        logger.info(
            "model_paths_resolved",
            spm_path=spm_path,
            ft_path=ft_path,
            ct_model_path=ct_model_full_path
        )

        return spm_path, ft_path, ct_model_full_path

    except Exception as e:
        logger.error("model_path_resolution_failed", error=str(e))
        raise e


def load_models():
    """Load all models into memory"""
    global lang_model, sp_model, translator

    logger.info("starting_model_loading")

    # Get model paths
    spm_path, ft_path, ct_model_path = get_model_paths()

    # Suppress fasttext warnings
    fasttext.FastText.eprint = lambda x: None

    try:
        # Load language detection model
        logger.info("loading_language_detection_model")
        lang_model = fasttext.load_model(ft_path)

        # Load SentencePiece model
        logger.info("loading_sentencepiece_model")
        sp_model = spm.SentencePieceProcessor()
        sp_model.load(spm_path)

        # Load translation model
        logger.info("loading_translation_model")
        translator = ctranslate2.Translator(ct_model_path, settings.device)

        logger.info("all_models_loaded_successfully")

    except Exception as e:
        logger.error("model_loading_failed", error=str(e))
        raise e


def translate_with_detection(text: str, target_lang: str) -> Tuple[str, str, float]:
    """Translate text with automatic source language detection"""
    start_time = time.time()

    try:
        # Prepare input
        source_sents = [text.strip()]
        target_prefix = [[target_lang]]

        # Detect source language
        predictions = lang_model.predict(text.replace('\n', ' '), k=1)
        source_lang = predictions[0][0].replace('__label__', '')

        # Tokenize source text
        source_sents_subworded = sp_model.encode(source_sents, out_type=str)
        source_sents_subworded = [[source_lang] + sent + ["</s>"] for sent in source_sents_subworded]

        # Translate
        translations = translator.translate_batch(
            source_sents_subworded,
            batch_type="tokens",
            max_batch_size=2048,
            beam_size=settings.beam_size,
            target_prefix=target_prefix,
        )

        # Decode translation
        translations = [translation[0]['tokens'] for translation in translations]
        translations_desubword = sp_model.decode(translations)
        translated_text = translations_desubword[0][len(target_lang):]

        inference_time = time.time() - start_time

        return source_lang, translated_text, inference_time

    except Exception as e:
        logger.error("translation_with_detection_failed", error=str(e), error_type=type(e).__name__)
        # Re-raise the exception to be handled by the endpoint
        raise e


def translate_with_source(text: str, source_lang: str, target_lang: str) -> Tuple[str, float]:
    """Translate text with provided source language"""
    start_time = time.time()

    try:
        # Prepare input
        source_sents = [text.strip()]
        target_prefix = [[target_lang]]

        # Tokenize source text
        source_sents_subworded = sp_model.encode(source_sents, out_type=str)
        source_sents_subworded = [[source_lang] + sent + ["</s>"] for sent in source_sents_subworded]

        # Translate
        translations = translator.translate_batch(
            source_sents_subworded,
            batch_type="tokens",
            max_batch_size=2048,
            beam_size=settings.beam_size,
            target_prefix=target_prefix
        )

        # Decode translation
        translations = [translation[0]['tokens'] for translation in translations]
        translations_desubword = sp_model.decode(translations)
        translated_text = translations_desubword[0][len(target_lang):]

        inference_time = time.time() - start_time

        return translated_text, inference_time

    except Exception as e:
        logger.error("translation_with_source_failed", error=str(e), error_type=type(e).__name__)
        # Re-raise the exception to be handled by the endpoint
        raise e


def detect_language(text: str) -> Tuple[str, float]:
    """
    Detect the language of input text

    Returns:
        Tuple of (language_code, confidence_score)
    """
    try:
        # Clean and normalize text for better detection
        # FastText models work better with lowercase text
        cleaned_text = text.replace('\n', ' ').strip().lower()

        # Get predictions with confidence scores
        predictions = lang_model.predict(cleaned_text, k=1)

        # Extract language code and confidence
        language_code = predictions[0][0].replace('__label__', '')
        raw_confidence = float(predictions[1][0])

        # Normalize confidence to ensure it's within [0.0, 1.0]
        # FastText sometimes returns values slightly above 1.0
        confidence = min(raw_confidence, 1.0)

        logger.info(
            "language_detected",
            text_length=len(text),
            original_text_sample=text[:50] + "..." if len(text) > 50 else text,
            cleaned_text_sample=cleaned_text[:50] + "..." if len(cleaned_text) > 50 else cleaned_text,
            detected_language=language_code,
            raw_confidence=raw_confidence,
            normalized_confidence=confidence
        )

        return language_code, confidence

    except Exception as e:
        logger.error("language_detection_failed", error=str(e), error_type=type(e).__name__)
        # Re-raise the exception to be handled by the endpoint
        raise e


def models_loaded() -> bool:
    """Check if all models are loaded"""
    return all([lang_model, sp_model, translator])