Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
|
@@ -6,31 +6,19 @@ from tokenizer_base import Tokenizer
|
|
| 6 |
import gradio as gr
|
| 7 |
|
| 8 |
# Параметры модели
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
tokenizer_base = Tokenizer(
|
| 13 |
|
| 14 |
# Надёжный препроцессинг
|
| 15 |
def preprocess_image(img, img_size):
|
| 16 |
-
# Приводим к RGB (убираем альфу/градации серого)
|
| 17 |
img = img.convert("RGB")
|
| 18 |
-
|
| 19 |
-
# Resize bicubic
|
| 20 |
img = img.resize(img_size, Image.BICUBIC)
|
| 21 |
-
|
| 22 |
-
# В numpy (HWC → float32)
|
| 23 |
img = np.array(img).astype(np.float32) / 255.0
|
| 24 |
-
|
| 25 |
-
# HWC → CHW
|
| 26 |
-
img = np.transpose(img, (2, 0, 1))
|
| 27 |
-
|
| 28 |
-
# Нормализация (x-mean)/std
|
| 29 |
img = (img - 0.5) / 0.5
|
| 30 |
-
|
| 31 |
-
# Добавляем batch размерность
|
| 32 |
-
img = np.expand_dims(img, axis=0)
|
| 33 |
-
|
| 34 |
return img
|
| 35 |
|
| 36 |
# softmax на numpy
|
|
@@ -38,42 +26,27 @@ def softmax(x, axis=-1):
|
|
| 38 |
e_x = np.exp(x - np.max(x, axis=axis, keepdims=True))
|
| 39 |
return e_x / e_x.sum(axis=axis, keepdims=True)
|
| 40 |
|
| 41 |
-
# Инициализация модели
|
| 42 |
def initialize_model(model_file):
|
| 43 |
onnx_model = onnx.load(model_file)
|
| 44 |
onnx.checker.check_model(onnx_model)
|
| 45 |
ort_session = rt.InferenceSession(model_file)
|
| 46 |
return ort_session
|
| 47 |
|
| 48 |
-
#
|
| 49 |
def get_text(img_org):
|
| 50 |
-
x = preprocess_image(img_org,
|
| 51 |
-
|
| 52 |
-
# Предсказание с помощью ONNX
|
| 53 |
ort_inputs = {ort_session.get_inputs()[0].name: x}
|
| 54 |
logits = ort_session.run(None, ort_inputs)[0]
|
| 55 |
-
|
| 56 |
probs = softmax(logits, axis=-1)
|
| 57 |
-
preds,
|
| 58 |
return preds[0]
|
| 59 |
|
| 60 |
-
#
|
| 61 |
-
ort_session = initialize_model(
|
| 62 |
|
| 63 |
# Gradio интерфейс
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
outputs=gr.Textbox(),
|
| 69 |
-
title="Text Captcha Reader",
|
| 70 |
-
description="Распознавание текста на изображениях капчи.",
|
| 71 |
-
examples=[
|
| 72 |
-
"8000.png", "11JW29.png", "2a8486.jpg", "2nbcx.png",
|
| 73 |
-
"000679.png", "000HU.png", "00Uga.png.jpg", "00bAQwhAZU.jpg",
|
| 74 |
-
"00h57kYf.jpg", "0EoHdtVb.png", "0JS21.png", "0p98z.png", "10010.png"
|
| 75 |
-
]
|
| 76 |
-
)
|
| 77 |
-
|
| 78 |
-
# Запуск интерфейса
|
| 79 |
-
gradio_interface().launch()
|
|
|
|
| 6 |
import gradio as gr
|
| 7 |
|
| 8 |
# Параметры модели
|
| 9 |
+
MODEL_FILE = "captcha.onnx"
|
| 10 |
+
IMG_SIZE = (32, 128)
|
| 11 |
+
CHARSET = r"0123456789abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ!\"#$%&'()*+,-./:;<=>?@[\\]^_`{|}~"
|
| 12 |
+
tokenizer_base = Tokenizer(CHARSET)
|
| 13 |
|
| 14 |
# Надёжный препроцессинг
|
| 15 |
def preprocess_image(img, img_size):
|
|
|
|
| 16 |
img = img.convert("RGB")
|
|
|
|
|
|
|
| 17 |
img = img.resize(img_size, Image.BICUBIC)
|
|
|
|
|
|
|
| 18 |
img = np.array(img).astype(np.float32) / 255.0
|
| 19 |
+
img = np.transpose(img, (2, 0, 1)) # HWC → CHW
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
img = (img - 0.5) / 0.5
|
| 21 |
+
img = np.expand_dims(img, axis=0) # batch dim
|
|
|
|
|
|
|
|
|
|
| 22 |
return img
|
| 23 |
|
| 24 |
# softmax на numpy
|
|
|
|
| 26 |
e_x = np.exp(x - np.max(x, axis=axis, keepdims=True))
|
| 27 |
return e_x / e_x.sum(axis=axis, keepdims=True)
|
| 28 |
|
| 29 |
+
# Инициализация модели
|
| 30 |
def initialize_model(model_file):
|
| 31 |
onnx_model = onnx.load(model_file)
|
| 32 |
onnx.checker.check_model(onnx_model)
|
| 33 |
ort_session = rt.InferenceSession(model_file)
|
| 34 |
return ort_session
|
| 35 |
|
| 36 |
+
# Распознавание текста
|
| 37 |
def get_text(img_org):
|
| 38 |
+
x = preprocess_image(img_org, IMG_SIZE)
|
|
|
|
|
|
|
| 39 |
ort_inputs = {ort_session.get_inputs()[0].name: x}
|
| 40 |
logits = ort_session.run(None, ort_inputs)[0]
|
|
|
|
| 41 |
probs = softmax(logits, axis=-1)
|
| 42 |
+
preds, _ = tokenizer_base.decode(probs)
|
| 43 |
return preds[0]
|
| 44 |
|
| 45 |
+
# Загружаем модель один раз
|
| 46 |
+
ort_session = initialize_model(MODEL_FILE)
|
| 47 |
|
| 48 |
# Gradio интерфейс
|
| 49 |
+
demo = gr.Interface(
|
| 50 |
+
fn=get_text,
|
| 51 |
+
inputs=gr.Image(type="pil"),
|
| 52 |
+
outputs=gr.T
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|