ASureevaA
commited on
Commit
·
fa051f7
1
Parent(s):
35e85d1
edit
Browse files- app.py +62 -59
- requirements.txt +1 -1
app.py
CHANGED
|
@@ -8,12 +8,12 @@ import torch
|
|
| 8 |
import gradio as gradio_module
|
| 9 |
from PIL import Image
|
| 10 |
from transformers import (
|
| 11 |
-
TrOCRProcessor,
|
| 12 |
-
VisionEncoderDecoderModel,
|
| 13 |
pipeline,
|
| 14 |
-
VitsTokenizer,
|
| 15 |
VitsModel,
|
|
|
|
| 16 |
)
|
|
|
|
|
|
|
| 17 |
|
| 18 |
# ============================
|
| 19 |
# 1. Настройки устройства
|
|
@@ -26,62 +26,66 @@ device_string: str = "cuda" if torch.cuda.is_available() else "cpu"
|
|
| 26 |
# 2. Модели
|
| 27 |
# ============================
|
| 28 |
|
| 29 |
-
|
| 30 |
-
# Модель: microsoft/trocr-small-printed
|
| 31 |
-
ocr_processor: TrOCRProcessor = TrOCRProcessor.from_pretrained(
|
| 32 |
-
"microsoft/trocr-small-printed"
|
| 33 |
-
)
|
| 34 |
-
ocr_model: VisionEncoderDecoderModel = VisionEncoderDecoderModel.from_pretrained(
|
| 35 |
-
"microsoft/trocr-small-printed"
|
| 36 |
-
)
|
| 37 |
-
ocr_model.to(device_string)
|
| 38 |
|
| 39 |
-
# Суммаризация: английский новостной/общий текст
|
| 40 |
-
# Модель: sshleifer/distilbart-cnn-12-6
|
| 41 |
summary_pipeline = pipeline(
|
| 42 |
task="summarization",
|
| 43 |
model="sshleifer/distilbart-cnn-12-6",
|
| 44 |
)
|
| 45 |
|
| 46 |
-
# TTS: английская MMS VITS
|
| 47 |
-
# Модель: facebook/mms-tts-eng
|
| 48 |
tts_model: VitsModel = VitsModel.from_pretrained("facebook/mms-tts-eng")
|
| 49 |
-
tts_tokenizer:
|
| 50 |
tts_model.to(device_string)
|
| 51 |
|
| 52 |
|
| 53 |
# ============================
|
| 54 |
-
# 3. OCR
|
| 55 |
# ============================
|
| 56 |
|
| 57 |
def run_ocr(image_object: Image.Image) -> str:
|
| 58 |
"""
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 64 |
"""
|
|
|
|
| 65 |
if image_object is None:
|
| 66 |
return ""
|
| 67 |
|
| 68 |
-
|
|
|
|
|
|
|
| 69 |
|
| 70 |
-
|
| 71 |
-
images=rgb_image_object,
|
| 72 |
-
return_tensors="pt",
|
| 73 |
-
)
|
| 74 |
-
pixel_values_tensor = processor_output.pixel_values.to(device_string)
|
| 75 |
|
| 76 |
-
|
| 77 |
-
|
|
|
|
|
|
|
|
|
|
| 78 |
|
| 79 |
-
decoded_text_list = ocr_processor.batch_decode(
|
| 80 |
-
generated_id_tensor,
|
| 81 |
-
skip_special_tokens=True,
|
| 82 |
-
)
|
| 83 |
|
| 84 |
-
|
|
|
|
|
|
|
| 85 |
return recognized_text
|
| 86 |
|
| 87 |
|
|
@@ -103,13 +107,14 @@ def run_summarization(
|
|
| 103 |
|
| 104 |
word_count: int = len(cleaned_text.split())
|
| 105 |
|
| 106 |
-
# Простая адаптация длины под размер текста,
|
| 107 |
-
# чтобы не было бессмысленных max_length >> input_length.
|
| 108 |
dynamic_max_length: int = min(
|
| 109 |
max_summary_tokens,
|
| 110 |
max(32, word_count + 20),
|
| 111 |
)
|
| 112 |
|
|
|
|
|
|
|
|
|
|
| 113 |
summary_result_list = summary_pipeline(
|
| 114 |
cleaned_text,
|
| 115 |
max_length=dynamic_max_length,
|
|
@@ -129,10 +134,8 @@ def run_tts(summary_text: str) -> Optional[str]:
|
|
| 129 |
"""
|
| 130 |
Озвучка английского текста конспекта через VitsModel (facebook/mms-tts-eng).
|
| 131 |
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
- ловим RuntimeError изнутри модели (бывают краши на редких входах);
|
| 135 |
-
в этом случае просто возвращаем None, чтобы не ронять весь Space.
|
| 136 |
"""
|
| 137 |
cleaned_text: str = summary_text.strip()
|
| 138 |
if not cleaned_text:
|
|
@@ -142,7 +145,6 @@ def run_tts(summary_text: str) -> Optional[str]:
|
|
| 142 |
cleaned_text,
|
| 143 |
return_tensors="pt",
|
| 144 |
)
|
| 145 |
-
|
| 146 |
tokenized_inputs = {
|
| 147 |
key: value.to(device_string)
|
| 148 |
for key, value in tokenized_inputs.items()
|
|
@@ -151,14 +153,13 @@ def run_tts(summary_text: str) -> Optional[str]:
|
|
| 151 |
input_ids_tensor = tokenized_inputs.get("input_ids")
|
| 152 |
if input_ids_tensor is None:
|
| 153 |
return None
|
| 154 |
-
|
| 155 |
if input_ids_tensor.numel() == 0 or input_ids_tensor.shape[1] == 0:
|
| 156 |
return None
|
| 157 |
|
| 158 |
try:
|
| 159 |
with torch.no_grad():
|
| 160 |
model_output = tts_model(**tokenized_inputs)
|
| 161 |
-
waveform_tensor = model_output.waveform #
|
| 162 |
except RuntimeError as runtime_error:
|
| 163 |
print(f"[WARN] TTS RuntimeError: {runtime_error}")
|
| 164 |
return None
|
|
@@ -190,9 +191,9 @@ def full_flow(
|
|
| 190 |
) -> Tuple[str, str, Optional[str]]:
|
| 191 |
"""
|
| 192 |
Полный пайплайн:
|
| 193 |
-
1) OCR: изображение -> исходный текст
|
| 194 |
-
2) Суммаризация: текст ->
|
| 195 |
-
3) TTS:
|
| 196 |
"""
|
| 197 |
recognized_text: str = run_ocr(image_object=image_object)
|
| 198 |
|
|
@@ -207,7 +208,7 @@ def full_flow(
|
|
| 207 |
|
| 208 |
|
| 209 |
# ============================
|
| 210 |
-
# 7. Gradio UI
|
| 211 |
# ============================
|
| 212 |
|
| 213 |
gradio_interface = gradio_module.Interface(
|
|
@@ -215,35 +216,37 @@ gradio_interface = gradio_module.Interface(
|
|
| 215 |
inputs=[
|
| 216 |
gradio_module.Image(
|
| 217 |
type="pil",
|
| 218 |
-
label="
|
| 219 |
),
|
| 220 |
gradio_module.Slider(
|
| 221 |
minimum=32,
|
| 222 |
maximum=256,
|
| 223 |
value=128,
|
| 224 |
step=16,
|
| 225 |
-
label="
|
| 226 |
),
|
| 227 |
],
|
| 228 |
outputs=[
|
| 229 |
gradio_module.Textbox(
|
| 230 |
-
label="
|
| 231 |
-
lines=
|
| 232 |
),
|
| 233 |
gradio_module.Textbox(
|
| 234 |
-
label="
|
| 235 |
lines=6,
|
| 236 |
),
|
| 237 |
gradio_module.Audio(
|
| 238 |
-
label="
|
| 239 |
type="filepath",
|
| 240 |
),
|
| 241 |
],
|
| 242 |
-
title="
|
| 243 |
description=(
|
| 244 |
-
"1)
|
| 245 |
-
"2)
|
| 246 |
-
"3)
|
|
|
|
|
|
|
| 247 |
),
|
| 248 |
)
|
| 249 |
|
|
|
|
| 8 |
import gradio as gradio_module
|
| 9 |
from PIL import Image
|
| 10 |
from transformers import (
|
|
|
|
|
|
|
| 11 |
pipeline,
|
|
|
|
| 12 |
VitsModel,
|
| 13 |
+
AutoTokenizer,
|
| 14 |
)
|
| 15 |
+
from nemotron_ocr.inference.pipeline import NemotronOCR # <-- Nemotron OCR v1
|
| 16 |
+
|
| 17 |
|
| 18 |
# ============================
|
| 19 |
# 1. Настройки устройства
|
|
|
|
| 26 |
# 2. Модели
|
| 27 |
# ============================
|
| 28 |
|
| 29 |
+
ocr_engine: NemotronOCR = NemotronOCR()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 30 |
|
|
|
|
|
|
|
| 31 |
summary_pipeline = pipeline(
|
| 32 |
task="summarization",
|
| 33 |
model="sshleifer/distilbart-cnn-12-6",
|
| 34 |
)
|
| 35 |
|
|
|
|
|
|
|
| 36 |
tts_model: VitsModel = VitsModel.from_pretrained("facebook/mms-tts-eng")
|
| 37 |
+
tts_tokenizer: AutoTokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-eng")
|
| 38 |
tts_model.to(device_string)
|
| 39 |
|
| 40 |
|
| 41 |
# ============================
|
| 42 |
+
# 3. OCR через NemotronOCR
|
| 43 |
# ============================
|
| 44 |
|
| 45 |
def run_ocr(image_object: Image.Image) -> str:
|
| 46 |
"""
|
| 47 |
+
OCR для печатного (и вообще любого) английского текста с картины.
|
| 48 |
+
|
| 49 |
+
Используем NemotronOCR из nvidia/nemotron-ocr-v1.
|
| 50 |
+
Модель сама делает:
|
| 51 |
+
- детекцию текстовых блоков,
|
| 52 |
+
- распознавание текста,
|
| 53 |
+
- анализ порядка чтения.
|
| 54 |
+
|
| 55 |
+
На выходе NemotronOCR даёт список dict:
|
| 56 |
+
[
|
| 57 |
+
{
|
| 58 |
+
"text": "...",
|
| 59 |
+
"confidence": float,
|
| 60 |
+
"left": float,
|
| 61 |
+
"upper": float,
|
| 62 |
+
"right": float,
|
| 63 |
+
"lower": float,
|
| 64 |
+
...
|
| 65 |
+
},
|
| 66 |
+
...
|
| 67 |
+
]
|
| 68 |
"""
|
| 69 |
+
|
| 70 |
if image_object is None:
|
| 71 |
return ""
|
| 72 |
|
| 73 |
+
with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as temporary_file:
|
| 74 |
+
image_object.save(temporary_file.name)
|
| 75 |
+
image_path: str = temporary_file.name
|
| 76 |
|
| 77 |
+
predictions = ocr_engine(image_path)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 78 |
|
| 79 |
+
text_parts = []
|
| 80 |
+
for prediction in predictions:
|
| 81 |
+
text_value = prediction.get("text", "")
|
| 82 |
+
if not text_value:
|
| 83 |
+
continue
|
| 84 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 85 |
|
| 86 |
+
text_parts.append(str(text_value))
|
| 87 |
+
|
| 88 |
+
recognized_text: str = "\n".join(text_parts).strip()
|
| 89 |
return recognized_text
|
| 90 |
|
| 91 |
|
|
|
|
| 107 |
|
| 108 |
word_count: int = len(cleaned_text.split())
|
| 109 |
|
|
|
|
|
|
|
| 110 |
dynamic_max_length: int = min(
|
| 111 |
max_summary_tokens,
|
| 112 |
max(32, word_count + 20),
|
| 113 |
)
|
| 114 |
|
| 115 |
+
if word_count < 8:
|
| 116 |
+
return cleaned_text
|
| 117 |
+
|
| 118 |
summary_result_list = summary_pipeline(
|
| 119 |
cleaned_text,
|
| 120 |
max_length=dynamic_max_length,
|
|
|
|
| 134 |
"""
|
| 135 |
Озвучка английского текста конспекта через VitsModel (facebook/mms-tts-eng).
|
| 136 |
|
| 137 |
+
Если модель внутри упадёт (известный баг на некоторых странных инпутах),
|
| 138 |
+
мы просто вернём None и не будем ронять всё приложение.
|
|
|
|
|
|
|
| 139 |
"""
|
| 140 |
cleaned_text: str = summary_text.strip()
|
| 141 |
if not cleaned_text:
|
|
|
|
| 145 |
cleaned_text,
|
| 146 |
return_tensors="pt",
|
| 147 |
)
|
|
|
|
| 148 |
tokenized_inputs = {
|
| 149 |
key: value.to(device_string)
|
| 150 |
for key, value in tokenized_inputs.items()
|
|
|
|
| 153 |
input_ids_tensor = tokenized_inputs.get("input_ids")
|
| 154 |
if input_ids_tensor is None:
|
| 155 |
return None
|
|
|
|
| 156 |
if input_ids_tensor.numel() == 0 or input_ids_tensor.shape[1] == 0:
|
| 157 |
return None
|
| 158 |
|
| 159 |
try:
|
| 160 |
with torch.no_grad():
|
| 161 |
model_output = tts_model(**tokenized_inputs)
|
| 162 |
+
waveform_tensor = model_output.waveform # (batch, n_samples)
|
| 163 |
except RuntimeError as runtime_error:
|
| 164 |
print(f"[WARN] TTS RuntimeError: {runtime_error}")
|
| 165 |
return None
|
|
|
|
| 191 |
) -> Tuple[str, str, Optional[str]]:
|
| 192 |
"""
|
| 193 |
Полный пайплайн:
|
| 194 |
+
1) OCR: изображение -> исходный английский текст
|
| 195 |
+
2) Суммаризация: текст -> конспект (английский)
|
| 196 |
+
3) TTS: конспект -> .wav файл (или None, если TTS не смог)
|
| 197 |
"""
|
| 198 |
recognized_text: str = run_ocr(image_object=image_object)
|
| 199 |
|
|
|
|
| 208 |
|
| 209 |
|
| 210 |
# ============================
|
| 211 |
+
# 7. Gradio UI (на русском)
|
| 212 |
# ============================
|
| 213 |
|
| 214 |
gradio_interface = gradio_module.Interface(
|
|
|
|
| 216 |
inputs=[
|
| 217 |
gradio_module.Image(
|
| 218 |
type="pil",
|
| 219 |
+
label="Изображение с напечатанным английским текстом",
|
| 220 |
),
|
| 221 |
gradio_module.Slider(
|
| 222 |
minimum=32,
|
| 223 |
maximum=256,
|
| 224 |
value=128,
|
| 225 |
step=16,
|
| 226 |
+
label="Максимальная длина конспекта (токены, примерно)",
|
| 227 |
),
|
| 228 |
],
|
| 229 |
outputs=[
|
| 230 |
gradio_module.Textbox(
|
| 231 |
+
label="Распознанный текст (Nemotron OCR)",
|
| 232 |
+
lines=8,
|
| 233 |
),
|
| 234 |
gradio_module.Textbox(
|
| 235 |
+
label="Конспект (английский текст)",
|
| 236 |
lines=6,
|
| 237 |
),
|
| 238 |
gradio_module.Audio(
|
| 239 |
+
label="Озвучка конспекта (английский TTS)",
|
| 240 |
type="filepath",
|
| 241 |
),
|
| 242 |
],
|
| 243 |
+
title="Картинка → Текст → Конспект → Озвучка (Nemotron OCR + английские модели)",
|
| 244 |
description=(
|
| 245 |
+
"1) Nemotron OCR v1 (nvidia/nemotron-ocr-v1) распознаёт текст с документа.\n"
|
| 246 |
+
"2) Английский трансформер суммаризации делает краткий пересказ.\n"
|
| 247 |
+
"3) VITS-модель MMS (facebook/mms-tts-eng) озвучивает конспект.\n\n"
|
| 248 |
+
"Если озвучка не сгенерировалась, значит конкретный текст не понравился TTS-модели "
|
| 249 |
+
"и она упала внутри — пайплайн просто пропустит аудио."
|
| 250 |
),
|
| 251 |
)
|
| 252 |
|
requirements.txt
CHANGED
|
@@ -1,4 +1,4 @@
|
|
| 1 |
-
transformers>=4.
|
| 2 |
torch
|
| 3 |
sentencepiece
|
| 4 |
gradio
|
|
|
|
| 1 |
+
transformers>=4.40.0
|
| 2 |
torch
|
| 3 |
sentencepiece
|
| 4 |
gradio
|