Spaces:
Sleeping
Sleeping
Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,176 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
"""app.ipynb
|
3 |
+
|
4 |
+
Automatically generated by Colab.
|
5 |
+
|
6 |
+
Original file is located at
|
7 |
+
https://colab.research.google.com/drive/17w1I1LKrJAebkjqIeNAKHQDirlY8Xxsw
|
8 |
+
"""
|
9 |
+
|
10 |
+
import torch
|
11 |
+
import torch.nn.functional as F
|
12 |
+
import torchvision
|
13 |
+
import matplotlib.pyplot as plt
|
14 |
+
import zipfile
|
15 |
+
import os
|
16 |
+
import gradio as gr
|
17 |
+
from PIL import Image
|
18 |
+
|
19 |
+
CHARS = "~=" + " abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789,.'-!?:;\""
|
20 |
+
BLANK = 0
|
21 |
+
PAD = 1
|
22 |
+
CHARS_DICT = {c: i for i, c in enumerate(CHARS)}
|
23 |
+
TEXTLEN = 30
|
24 |
+
|
25 |
+
tokens_list = list(CHARS_DICT.keys())
|
26 |
+
silence_token = '|'
|
27 |
+
|
28 |
+
if silence_token not in tokens_list:
|
29 |
+
tokens_list.append(silence_token)
|
30 |
+
|
31 |
+
|
32 |
+
def fit_picture(img):
|
33 |
+
target_height = 32
|
34 |
+
target_width = 400
|
35 |
+
|
36 |
+
# Calculate resize dimensions
|
37 |
+
aspect_ratio = img.width / img.height
|
38 |
+
if aspect_ratio > (target_width / target_height):
|
39 |
+
resize_width = target_width
|
40 |
+
resize_height = int(target_width / aspect_ratio)
|
41 |
+
else:
|
42 |
+
resize_height = target_height
|
43 |
+
resize_width = int(target_height * aspect_ratio)
|
44 |
+
|
45 |
+
# Resize transformation
|
46 |
+
resize_transform = transforms.Resize((resize_height, resize_width))
|
47 |
+
|
48 |
+
# Pad transformation
|
49 |
+
padding_height = (target_height - resize_height) if target_height > resize_height else 0
|
50 |
+
padding_width = (target_width - resize_width) if target_width > resize_width else 0
|
51 |
+
pad_transform = transforms.Pad((0, 0, padding_width, padding_height), fill=0, padding_mode='constant')
|
52 |
+
|
53 |
+
transform = torchvision.transforms.Compose([
|
54 |
+
torchvision.transforms.Grayscale(num_output_channels = 1),
|
55 |
+
torchvision.transforms.ToTensor(),
|
56 |
+
torchvision.transforms.Normalize(0.5,0.5),
|
57 |
+
resize_transform,
|
58 |
+
pad_transform
|
59 |
+
])
|
60 |
+
|
61 |
+
fin_img = transform(img)
|
62 |
+
return fin_img
|
63 |
+
|
64 |
+
def load_model(filename):
|
65 |
+
data = torch.load(filename)
|
66 |
+
recognizer.load_state_dict(data["recognizer"])
|
67 |
+
optimizer.load_state_dict(data["optimizer"])
|
68 |
+
|
69 |
+
def ctc_decode_sequence(seq):
|
70 |
+
"""Removes blanks and repetitions from the sequence."""
|
71 |
+
ret = []
|
72 |
+
prev = BLANK
|
73 |
+
for x in seq:
|
74 |
+
if prev != BLANK and prev != x:
|
75 |
+
ret.append(prev)
|
76 |
+
prev = x
|
77 |
+
if seq[-1] == 66:
|
78 |
+
ret.append(66)
|
79 |
+
return ret
|
80 |
+
|
81 |
+
def ctc_decode(codes):
|
82 |
+
"""Decode a batch of sequences."""
|
83 |
+
ret = []
|
84 |
+
for cs in codes.T:
|
85 |
+
ret.append(ctc_decode_sequence(cs))
|
86 |
+
return ret
|
87 |
+
|
88 |
+
|
89 |
+
def decode_text(codes):
|
90 |
+
chars = [CHARS[c] for c in codes]
|
91 |
+
return ''.join(chars)
|
92 |
+
|
93 |
+
class Residual(torch.nn.Module):
|
94 |
+
def __init__(self, in_channels, out_channels, stride, pdrop = 0.2):
|
95 |
+
super().__init__()
|
96 |
+
self.conv1 = torch.nn.Conv2d(in_channels, out_channels, 3, stride, 1)
|
97 |
+
self.bn1 = torch.nn.BatchNorm2d(out_channels)
|
98 |
+
self.conv2 = torch.nn.Conv2d(out_channels, out_channels, 3, 1, 1)
|
99 |
+
self.bn2 = torch.nn.BatchNorm2d(out_channels)
|
100 |
+
if in_channels != out_channels or stride != 1:
|
101 |
+
self.skip = torch.nn.Conv2d(in_channels, out_channels, 1, stride, 0)
|
102 |
+
else:
|
103 |
+
self.skip = torch.nn.Identity()
|
104 |
+
self.dropout = torch.nn.Dropout2d(pdrop)
|
105 |
+
|
106 |
+
def forward(self, x):
|
107 |
+
y = torch.nn.functional.relu(self.bn1(self.conv1(x)))
|
108 |
+
y = torch.nn.functional.relu(self.bn2(self.conv2(y)) + self.skip(x))
|
109 |
+
y = self.dropout(y)
|
110 |
+
return y
|
111 |
+
|
112 |
+
class TextRecognizer(torch.nn.Module):
|
113 |
+
def __init__(self, labels):
|
114 |
+
super().__init__()
|
115 |
+
self.feature_extractor = torch.nn.Sequential(
|
116 |
+
Residual(1, 32, 1),
|
117 |
+
Residual(32, 32, 2),
|
118 |
+
Residual(32, 32, 1),
|
119 |
+
Residual(32, 64, 2),
|
120 |
+
Residual(64, 64, 1),
|
121 |
+
Residual(64, 128, (2,1)),
|
122 |
+
Residual(128, 128, 1),
|
123 |
+
Residual(128, 128, (2,1)),
|
124 |
+
Residual(128, 128, (2,1)),
|
125 |
+
)
|
126 |
+
self.recurrent = torch.nn.LSTM(128, 128, 1 ,bidirectional = True)
|
127 |
+
self.output = torch.nn.Linear(256, labels)
|
128 |
+
|
129 |
+
def forward(self, x):
|
130 |
+
x = self.feature_extractor(x)
|
131 |
+
x = x.squeeze(2)
|
132 |
+
x = x.permute(2,0,1)
|
133 |
+
x,_ = self.recurrent(x)
|
134 |
+
x = self.output(x)
|
135 |
+
return x
|
136 |
+
|
137 |
+
recognizer = TextRecognizer(len(CHARS))
|
138 |
+
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
139 |
+
print("Device:", DEVICE)
|
140 |
+
LR = 1e-3
|
141 |
+
|
142 |
+
recognizer.to(DEVICE)
|
143 |
+
optimizer = torch.optim.Adam(recognizer.parameters(), lr=LR)
|
144 |
+
|
145 |
+
load_model('model.pt')
|
146 |
+
recognizer.eval()
|
147 |
+
|
148 |
+
def ctc_read(image):
|
149 |
+
imagefin = fit_picture(image)
|
150 |
+
image_tensor = imagefin.unsqueeze(0).to(DEVICE)
|
151 |
+
print(image_tensor.size())
|
152 |
+
|
153 |
+
with torch.no_grad():
|
154 |
+
scores = recognizer(image_tensor)
|
155 |
+
|
156 |
+
predictions = scores.argmax(2).cpu().numpy()
|
157 |
+
|
158 |
+
decoded_sequences = ctc_decode(predictions)
|
159 |
+
|
160 |
+
# Convert decoded sequences to text
|
161 |
+
for i in decoded_sequences:
|
162 |
+
decoded_text = decode_text(i)
|
163 |
+
|
164 |
+
return decoded_text
|
165 |
+
|
166 |
+
|
167 |
+
# Gradio Interface
|
168 |
+
iface = gr.Interface(
|
169 |
+
fn=ctc_read,
|
170 |
+
inputs=gr.Image(type="pil"), # PIL Image input
|
171 |
+
outputs="text", # Text output
|
172 |
+
title="Handwritten Text Recognition",
|
173 |
+
description="Upload an image, and the custome AI will extract the text."
|
174 |
+
)
|
175 |
+
|
176 |
+
iface.launch(share=True)
|