entropy25 commited on
Commit
bc367e1
·
verified ·
1 Parent(s): 8efd9d9

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +126 -66
app.py CHANGED
@@ -4,25 +4,39 @@ from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, BitsAndBytesConfi
4
  from peft import PeftModel
5
  from functools import lru_cache
6
  import os
 
 
7
 
8
- base_model_name = "facebook/nllb-200-distilled-600M"
9
- adapter_en_to_no = "entropy25/mt_en_no_oil"
10
- adapter_no_to_en = "entropy25/mt_no_en_oil"
 
 
 
 
 
 
 
 
 
 
 
 
11
 
12
- tokenizer = AutoTokenizer.from_pretrained(base_model_name)
13
 
14
  print("Loading shared base model with 8-bit quantization...")
15
  quantization_config = BitsAndBytesConfig(load_in_8bit=True)
16
 
17
  base_model = AutoModelForSeq2SeqLM.from_pretrained(
18
- base_model_name,
19
  quantization_config=quantization_config,
20
  low_cpu_mem_usage=True
21
  )
22
 
23
  print("Loading adapters...")
24
- model = PeftModel.from_pretrained(base_model, adapter_en_to_no, adapter_name="en_to_no")
25
- model.load_adapter(adapter_no_to_en, adapter_name="no_to_en")
26
  model.eval()
27
 
28
  QUALITY_PRESETS = {
@@ -74,10 +88,13 @@ QUALITY_TEST_CASES = {
74
  }
75
 
76
  MAX_FILE_SIZE = 1024 * 1024
 
77
 
78
  def translate_core(text, source_lang, target_lang, quality_preset):
79
  if not text.strip() or source_lang == target_lang:
80
- return text
 
 
81
 
82
  if source_lang == "English" and target_lang == "Norwegian":
83
  model.set_adapter("en_to_no")
@@ -86,63 +103,74 @@ def translate_core(text, source_lang, target_lang, quality_preset):
86
  model.set_adapter("no_to_en")
87
  tgt_code = "eng_Latn"
88
  else:
89
- return "Unsupported language pair"
90
 
91
  preset = QUALITY_PRESETS[quality_preset]
92
 
93
- lines = text.split('\n')
94
- non_empty_lines = [line for line in lines if line.strip()]
95
-
96
- if not non_empty_lines:
97
- return text
98
 
99
- batch_size = preset["batch_size"]
100
- all_results = []
101
-
102
- for i in range(0, len(non_empty_lines), batch_size):
103
- batch = non_empty_lines[i:i+batch_size]
104
 
105
- inputs = tokenizer(
106
- batch,
107
- return_tensors="pt",
108
- padding=True,
109
- truncation=True,
110
- max_length=preset["max_length"]
111
- )
112
 
113
- if hasattr(model, 'device'):
114
- inputs = {k: v.to(model.device) for k, v in inputs.items()}
115
 
116
- with torch.inference_mode():
117
- outputs = model.generate(
118
- **inputs,
119
- forced_bos_token_id=tokenizer.convert_tokens_to_ids(tgt_code),
120
- max_length=preset["max_length"],
121
- num_beams=preset["num_beams"],
122
- early_stopping=True
 
 
123
  )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
124
 
125
- batch_results = tokenizer.batch_decode(outputs, skip_special_tokens=True)
126
- all_results.extend(batch_results)
127
-
128
- result_iter = iter(all_results)
129
- final_lines = []
130
- for line in lines:
131
- if line.strip():
132
- final_lines.append(next(result_iter))
133
- else:
134
- final_lines.append("")
135
 
136
- return '\n'.join(final_lines)
 
137
 
138
  @lru_cache(maxsize=512)
139
  def translate_cached(text, source_lang, target_lang, quality_preset):
140
- return translate_core(text, source_lang, target_lang, quality_preset)
 
141
 
142
  def translate(text, source_lang, target_lang, quality_preset):
143
- if len(text) > 10000:
144
- return "Error: Text too long (max 10,000 characters)"
145
- return translate_cached(text, source_lang, target_lang, quality_preset)
 
 
 
 
 
 
 
 
 
 
 
146
 
147
  def run_quality_tests():
148
  results = []
@@ -157,7 +185,7 @@ def run_quality_tests():
157
  results.append(f"\n{src_lang} to {tgt_lang}\n")
158
 
159
  for i, case in enumerate(test_cases, 1):
160
- translation = translate_core(case["input"], src_lang, tgt_lang, "Professional (Best Quality)")
161
 
162
  passed_checks = []
163
  failed_checks = []
@@ -195,26 +223,26 @@ def swap_languages(src, tgt, input_txt, output_txt):
195
 
196
  def load_file(file):
197
  if file is None:
198
- return ""
199
 
200
  try:
201
  if os.path.getsize(file.name) > MAX_FILE_SIZE:
202
- return "Error: File too large (max 1MB)"
203
 
204
  with open(file.name, 'r', encoding='utf-8') as f:
205
  content = f.read()
206
- if len(content) > 10000:
207
- return "Error: File content too long (max 10,000 characters)"
208
- return content
209
  except:
210
  try:
211
  with open(file.name, 'r', encoding='latin-1') as f:
212
  content = f.read()
213
- if len(content) > 10000:
214
- return "Error: File content too long (max 10,000 characters)"
215
- return content
216
  except Exception as e:
217
- return f"Error reading file: {str(e)}"
218
 
219
  EXAMPLES_EN = {
220
  "drilling_short": "Mud weight adjusted to 1.82 specific gravity at 3,247 meters depth.",
@@ -312,7 +340,7 @@ custom_css = """
312
  border-color: #0f6fff !important;
313
  }
314
  .translate-btn {
315
- background: #0f6fff !important;
316
  color: white !important;
317
  border: none !important;
318
  padding: 12px 24px !important;
@@ -322,7 +350,14 @@ custom_css = """
322
  cursor: pointer !important;
323
  }
324
  .translate-btn:hover {
325
- background: #0d5dd9 !important;
 
 
 
 
 
 
 
326
  }
327
  .footer-info {
328
  text-align: center !important;
@@ -335,10 +370,27 @@ custom_css = """
335
  border: 1px solid #0f6fff !important;
336
  border-radius: 4px !important;
337
  }
 
 
 
 
 
 
 
 
 
338
  """
339
 
340
  with gr.Blocks(css=custom_css, theme=gr.themes.Default()) as demo:
341
- gr.HTML("<div style='height: 20px'></div>")
 
 
 
 
 
 
 
 
342
 
343
  with gr.Row():
344
  quality_preset = gr.Radio(
@@ -398,7 +450,15 @@ with gr.Blocks(css=custom_css, theme=gr.themes.Default()) as demo:
398
  with gr.Row():
399
  translate_btn = gr.Button("Translate", variant="primary", elem_classes="translate-btn", size="lg")
400
 
401
- gr.HTML("<div class='footer-info'>Oil & Gas Translation • English ↔ Norwegian • Optimized for HF Space</div>")
 
 
 
 
 
 
 
 
402
 
403
  with gr.Accordion("Example Sentences", open=True):
404
  with gr.Row():
@@ -460,7 +520,7 @@ with gr.Blocks(css=custom_css, theme=gr.themes.Default()) as demo:
460
  translate_btn.click(
461
  fn=translate,
462
  inputs=[input_text, source_lang, target_lang, quality_preset],
463
- outputs=output_text
464
  )
465
 
466
  swap_btn.click(
@@ -469,6 +529,6 @@ with gr.Blocks(css=custom_css, theme=gr.themes.Default()) as demo:
469
  outputs=[source_lang, target_lang, input_text, output_text]
470
  )
471
 
472
- file_input.change(fn=load_file, inputs=file_input, outputs=input_text)
473
 
474
  demo.queue().launch()
 
4
  from peft import PeftModel
5
  from functools import lru_cache
6
  import os
7
+ import time
8
+ import nltk
9
 
10
+ try:
11
+ nltk.data.find('tokenizers/punkt')
12
+ except LookupError:
13
+ print("Downloading NLTK punkt tokenizer...")
14
+ nltk.download('punkt')
15
+ try:
16
+ nltk.download('punkt_tab')
17
+ except:
18
+ pass
19
+
20
+ from nltk.tokenize import sent_tokenize
21
+
22
+ BASE_MODEL = os.getenv("BASE_MODEL", "facebook/nllb-200-distilled-600M")
23
+ ADAPTER_EN_TO_NO = os.getenv("ADAPTER_EN_TO_NO", "entropy25/mt_en_no_oil")
24
+ ADAPTER_NO_TO_EN = os.getenv("ADAPTER_NO_TO_EN", "entropy25/mt_no_en_oil")
25
 
26
+ tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
27
 
28
  print("Loading shared base model with 8-bit quantization...")
29
  quantization_config = BitsAndBytesConfig(load_in_8bit=True)
30
 
31
  base_model = AutoModelForSeq2SeqLM.from_pretrained(
32
+ BASE_MODEL,
33
  quantization_config=quantization_config,
34
  low_cpu_mem_usage=True
35
  )
36
 
37
  print("Loading adapters...")
38
+ model = PeftModel.from_pretrained(base_model, ADAPTER_EN_TO_NO, adapter_name="en_to_no")
39
+ model.load_adapter(ADAPTER_NO_TO_EN, adapter_name="no_to_en")
40
  model.eval()
41
 
42
  QUALITY_PRESETS = {
 
88
  }
89
 
90
  MAX_FILE_SIZE = 1024 * 1024
91
+ MAX_TEXT_LENGTH = 10000
92
 
93
  def translate_core(text, source_lang, target_lang, quality_preset):
94
  if not text.strip() or source_lang == target_lang:
95
+ return text, 0.0
96
+
97
+ start_time = time.time()
98
 
99
  if source_lang == "English" and target_lang == "Norwegian":
100
  model.set_adapter("en_to_no")
 
103
  model.set_adapter("no_to_en")
104
  tgt_code = "eng_Latn"
105
  else:
106
+ return "Unsupported language pair", 0.0
107
 
108
  preset = QUALITY_PRESETS[quality_preset]
109
 
110
+ original_paragraphs = text.split('\n')
111
+ final_translated_paragraphs = []
 
 
 
112
 
113
+ for paragraph in original_paragraphs:
114
+ if not paragraph.strip():
115
+ final_translated_paragraphs.append("")
116
+ continue
 
117
 
118
+ sentences = sent_tokenize(paragraph)
 
 
 
 
 
 
119
 
120
+ batch_size = preset["batch_size"]
121
+ paragraph_results = []
122
 
123
+ for i in range(0, len(sentences), batch_size):
124
+ batch = sentences[i:i+batch_size]
125
+
126
+ inputs = tokenizer(
127
+ batch,
128
+ return_tensors="pt",
129
+ padding=True,
130
+ truncation=True,
131
+ max_length=preset["max_length"]
132
  )
133
+
134
+ if hasattr(model, 'device'):
135
+ inputs = {k: v.to(model.device) for k, v in inputs.items()}
136
+
137
+ with torch.inference_mode():
138
+ outputs = model.generate(
139
+ **inputs,
140
+ forced_bos_token_id=tokenizer.convert_tokens_to_ids(tgt_code),
141
+ max_length=preset["max_length"],
142
+ num_beams=preset["num_beams"],
143
+ early_stopping=True
144
+ )
145
+
146
+ batch_texts = tokenizer.batch_decode(outputs, skip_special_tokens=True)
147
+ paragraph_results.extend(batch_texts)
148
 
149
+ final_translated_paragraphs.append(" ".join(paragraph_results))
 
 
 
 
 
 
 
 
 
150
 
151
+ elapsed_time = time.time() - start_time
152
+ return '\n'.join(final_translated_paragraphs), elapsed_time
153
 
154
  @lru_cache(maxsize=512)
155
  def translate_cached(text, source_lang, target_lang, quality_preset):
156
+ result, elapsed = translate_core(text, source_lang, target_lang, quality_preset)
157
+ return result, elapsed
158
 
159
  def translate(text, source_lang, target_lang, quality_preset):
160
+ try:
161
+ if len(text) > MAX_TEXT_LENGTH:
162
+ return f"Error: Text too long (max {MAX_TEXT_LENGTH:,} characters)", ""
163
+
164
+ if not text.strip():
165
+ return "", ""
166
+
167
+ result, elapsed = translate_cached(text, source_lang, target_lang, quality_preset)
168
+
169
+ time_info = f"Translation completed in {elapsed:.2f}s"
170
+ return result, time_info
171
+
172
+ except Exception as e:
173
+ return f"Translation error: {str(e)}. Please try again.", ""
174
 
175
  def run_quality_tests():
176
  results = []
 
185
  results.append(f"\n{src_lang} to {tgt_lang}\n")
186
 
187
  for i, case in enumerate(test_cases, 1):
188
+ translation, _ = translate_core(case["input"], src_lang, tgt_lang, "Professional (Best Quality)")
189
 
190
  passed_checks = []
191
  failed_checks = []
 
223
 
224
  def load_file(file):
225
  if file is None:
226
+ return "", ""
227
 
228
  try:
229
  if os.path.getsize(file.name) > MAX_FILE_SIZE:
230
+ return "Error: File too large (max 1MB)", ""
231
 
232
  with open(file.name, 'r', encoding='utf-8') as f:
233
  content = f.read()
234
+ if len(content) > MAX_TEXT_LENGTH:
235
+ return f"Error: File content too long (max {MAX_TEXT_LENGTH:,} characters)", ""
236
+ return content, ""
237
  except:
238
  try:
239
  with open(file.name, 'r', encoding='latin-1') as f:
240
  content = f.read()
241
+ if len(content) > MAX_TEXT_LENGTH:
242
+ return f"Error: File content too long (max {MAX_TEXT_LENGTH:,} characters)", ""
243
+ return content, ""
244
  except Exception as e:
245
+ return f"Error reading file: {str(e)}", ""
246
 
247
  EXAMPLES_EN = {
248
  "drilling_short": "Mud weight adjusted to 1.82 specific gravity at 3,247 meters depth.",
 
340
  border-color: #0f6fff !important;
341
  }
342
  .translate-btn {
343
+ background: #ff8c00 !important;
344
  color: white !important;
345
  border: none !important;
346
  padding: 12px 24px !important;
 
350
  cursor: pointer !important;
351
  }
352
  .translate-btn:hover {
353
+ background: #e67e00 !important;
354
+ }
355
+ .time-info {
356
+ text-align: center !important;
357
+ color: #666 !important;
358
+ font-size: 13px !important;
359
+ padding: 10px !important;
360
+ font-style: italic !important;
361
  }
362
  .footer-info {
363
  text-align: center !important;
 
370
  border: 1px solid #0f6fff !important;
371
  border-radius: 4px !important;
372
  }
373
+ .disclaimer {
374
+ background: #fff9e6 !important;
375
+ border-left: 4px solid #ff8c00 !important;
376
+ padding: 15px 20px !important;
377
+ margin: 20px 0 !important;
378
+ border-radius: 4px !important;
379
+ font-size: 13px !important;
380
+ color: #666 !important;
381
+ }
382
  """
383
 
384
  with gr.Blocks(css=custom_css, theme=gr.themes.Default()) as demo:
385
+
386
+ gr.HTML("""
387
+ <div class='disclaimer'>
388
+ <strong>Privacy & Compliance:</strong> This model is fine-tuned on public domain oil & gas terminology.
389
+ No proprietary data was used in training. Local inference ensures data privacy and GDPR compliance.
390
+ <br><br>
391
+ <strong>Terminology Enforcement:</strong> Advanced lexical constraints for critical technical terms are currently in development.
392
+ </div>
393
+ """)
394
 
395
  with gr.Row():
396
  quality_preset = gr.Radio(
 
450
  with gr.Row():
451
  translate_btn = gr.Button("Translate", variant="primary", elem_classes="translate-btn", size="lg")
452
 
453
+ with gr.Row():
454
+ time_display = gr.Textbox(
455
+ show_label=False,
456
+ container=False,
457
+ interactive=False,
458
+ elem_classes="time-info"
459
+ )
460
+
461
+ gr.HTML("<div class='footer-info'>Oil & Gas Translation • English ↔ Norwegian • Production-Ready System</div>")
462
 
463
  with gr.Accordion("Example Sentences", open=True):
464
  with gr.Row():
 
520
  translate_btn.click(
521
  fn=translate,
522
  inputs=[input_text, source_lang, target_lang, quality_preset],
523
+ outputs=[output_text, time_display]
524
  )
525
 
526
  swap_btn.click(
 
529
  outputs=[source_lang, target_lang, input_text, output_text]
530
  )
531
 
532
+ file_input.change(fn=load_file, inputs=file_input, outputs=[input_text, time_display])
533
 
534
  demo.queue().launch()