ASureevaA
commited on
Commit
·
8dac34d
1
Parent(s):
a88eb1e
edit
Browse files- app.py +14 -26
- requirements.txt +0 -2
app.py
CHANGED
|
@@ -1,10 +1,8 @@
|
|
| 1 |
from typing import Tuple, Optional, Any
|
| 2 |
|
| 3 |
-
import torch
|
| 4 |
import numpy as numpy
|
| 5 |
-
import gradio as
|
| 6 |
from PIL import Image
|
| 7 |
-
from datasets import load_dataset
|
| 8 |
from transformers import (
|
| 9 |
TrOCRProcessor,
|
| 10 |
VisionEncoderDecoderModel,
|
|
@@ -23,22 +21,15 @@ summary_pipeline = pipeline(
|
|
| 23 |
model="sshleifer/distilbart-cnn-12-6",
|
| 24 |
)
|
| 25 |
|
| 26 |
-
|
| 27 |
task="text-to-speech",
|
| 28 |
-
model="
|
| 29 |
)
|
| 30 |
|
| 31 |
-
speaker_dataset = load_dataset(
|
| 32 |
-
path="Matthijs/cmu-arctic-xvectors",
|
| 33 |
-
split="validation",
|
| 34 |
-
)
|
| 35 |
-
speaker_embedding_tensor: torch.Tensor = torch.tensor(
|
| 36 |
-
speaker_dataset[7306]["xvector"]
|
| 37 |
-
).unsqueeze(0)
|
| 38 |
|
| 39 |
def run_ocr(image_object: Image.Image) -> str:
|
| 40 |
"""
|
| 41 |
-
Распознавание текста с
|
| 42 |
Предполагаем, что на картинке простой напечатанный текст.
|
| 43 |
"""
|
| 44 |
if image_object is None:
|
|
@@ -65,8 +56,8 @@ def run_summarization(
|
|
| 65 |
max_summary_tokens: int = 128,
|
| 66 |
) -> str:
|
| 67 |
"""
|
| 68 |
-
Суммаризация
|
| 69 |
-
|
| 70 |
"""
|
| 71 |
cleaned_text: str = input_text.strip()
|
| 72 |
if not cleaned_text:
|
|
@@ -86,16 +77,13 @@ def run_summarization(
|
|
| 86 |
def run_tts(summary_text: str) -> Optional[Tuple[int, Any]]:
|
| 87 |
"""
|
| 88 |
Озвучка текста конспекта.
|
| 89 |
-
|
| 90 |
"""
|
| 91 |
cleaned_text: str = summary_text.strip()
|
| 92 |
if not cleaned_text:
|
| 93 |
return None
|
| 94 |
|
| 95 |
-
tts_output =
|
| 96 |
-
cleaned_text,
|
| 97 |
-
forward_params={"speaker_embeddings": speaker_embedding_tensor},
|
| 98 |
-
)
|
| 99 |
|
| 100 |
sampling_rate_int: int = int(tts_output["sampling_rate"])
|
| 101 |
audio_array = tts_output["audio"]
|
|
@@ -122,14 +110,14 @@ def full_flow(
|
|
| 122 |
|
| 123 |
return recognized_text, summary_text, audio_tuple
|
| 124 |
|
| 125 |
-
gradio_interface =
|
| 126 |
fn=full_flow,
|
| 127 |
inputs=[
|
| 128 |
-
|
| 129 |
type="pil",
|
| 130 |
label="Изображение с напечатанным текстом (английский)",
|
| 131 |
),
|
| 132 |
-
|
| 133 |
minimum=32,
|
| 134 |
maximum=256,
|
| 135 |
value=128,
|
|
@@ -138,15 +126,15 @@ gradio_interface = gr.Interface(
|
|
| 138 |
),
|
| 139 |
],
|
| 140 |
outputs=[
|
| 141 |
-
|
| 142 |
label="Распознанный текст (OCR)",
|
| 143 |
lines=6,
|
| 144 |
),
|
| 145 |
-
|
| 146 |
label="Конспект (суммаризация)",
|
| 147 |
lines=6,
|
| 148 |
),
|
| 149 |
-
|
| 150 |
label="Озвучка конспекта (TTS)",
|
| 151 |
type="numpy",
|
| 152 |
),
|
|
|
|
| 1 |
from typing import Tuple, Optional, Any
|
| 2 |
|
|
|
|
| 3 |
import numpy as numpy
|
| 4 |
+
import gradio as gradio_module
|
| 5 |
from PIL import Image
|
|
|
|
| 6 |
from transformers import (
|
| 7 |
TrOCRProcessor,
|
| 8 |
VisionEncoderDecoderModel,
|
|
|
|
| 21 |
model="sshleifer/distilbart-cnn-12-6",
|
| 22 |
)
|
| 23 |
|
| 24 |
+
text_to_speech_pipeline = pipeline(
|
| 25 |
task="text-to-speech",
|
| 26 |
+
model="facebook/mms-tts-eng",
|
| 27 |
)
|
| 28 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
|
| 30 |
def run_ocr(image_object: Image.Image) -> str:
|
| 31 |
"""
|
| 32 |
+
Распознавание текста с изображения.
|
| 33 |
Предполагаем, что на картинке простой напечатанный текст.
|
| 34 |
"""
|
| 35 |
if image_object is None:
|
|
|
|
| 56 |
max_summary_tokens: int = 128,
|
| 57 |
) -> str:
|
| 58 |
"""
|
| 59 |
+
Суммаризация текста до короткого конспекта.
|
| 60 |
+
Без разбиения на чанки, поэтому огромные тексты лучше не подавать.
|
| 61 |
"""
|
| 62 |
cleaned_text: str = input_text.strip()
|
| 63 |
if not cleaned_text:
|
|
|
|
| 77 |
def run_tts(summary_text: str) -> Optional[Tuple[int, Any]]:
|
| 78 |
"""
|
| 79 |
Озвучка текста конспекта.
|
| 80 |
+
Используем модель, которой не нужны внешние speaker embeddings.
|
| 81 |
"""
|
| 82 |
cleaned_text: str = summary_text.strip()
|
| 83 |
if not cleaned_text:
|
| 84 |
return None
|
| 85 |
|
| 86 |
+
tts_output = text_to_speech_pipeline(cleaned_text)
|
|
|
|
|
|
|
|
|
|
| 87 |
|
| 88 |
sampling_rate_int: int = int(tts_output["sampling_rate"])
|
| 89 |
audio_array = tts_output["audio"]
|
|
|
|
| 110 |
|
| 111 |
return recognized_text, summary_text, audio_tuple
|
| 112 |
|
| 113 |
+
gradio_interface = gradio_module.Interface(
|
| 114 |
fn=full_flow,
|
| 115 |
inputs=[
|
| 116 |
+
gradio_module.Image(
|
| 117 |
type="pil",
|
| 118 |
label="Изображение с напечатанным текстом (английский)",
|
| 119 |
),
|
| 120 |
+
gradio_module.Slider(
|
| 121 |
minimum=32,
|
| 122 |
maximum=256,
|
| 123 |
value=128,
|
|
|
|
| 126 |
),
|
| 127 |
],
|
| 128 |
outputs=[
|
| 129 |
+
gradio_module.Textbox(
|
| 130 |
label="Распознанный текст (OCR)",
|
| 131 |
lines=6,
|
| 132 |
),
|
| 133 |
+
gradio_module.Textbox(
|
| 134 |
label="Конспект (суммаризация)",
|
| 135 |
lines=6,
|
| 136 |
),
|
| 137 |
+
gradio_module.Audio(
|
| 138 |
label="Озвучка конспекта (TTS)",
|
| 139 |
type="numpy",
|
| 140 |
),
|
requirements.txt
CHANGED
|
@@ -1,8 +1,6 @@
|
|
| 1 |
transformers
|
| 2 |
torch
|
| 3 |
-
datasets
|
| 4 |
sentencepiece
|
| 5 |
-
soundfile
|
| 6 |
gradio
|
| 7 |
Pillow
|
| 8 |
numpy
|
|
|
|
| 1 |
transformers
|
| 2 |
torch
|
|
|
|
| 3 |
sentencepiece
|
|
|
|
| 4 |
gradio
|
| 5 |
Pillow
|
| 6 |
numpy
|