Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# app.py
|
2 |
+
|
3 |
+
import os
|
4 |
+
import torch
|
5 |
+
import torch.nn as nn
|
6 |
+
from torchvision import transforms
|
7 |
+
from PIL import Image
|
8 |
+
import numpy as np
|
9 |
+
import gradio as gr
|
10 |
+
import timm
|
11 |
+
import matplotlib.pyplot as plt
|
12 |
+
import matplotlib.patches as patches
|
13 |
+
|
14 |
+
# Optional: If integrating OCR
|
15 |
+
# import pytesseract
|
16 |
+
|
17 |
+
# Define the Detection Model Architecture
|
18 |
+
class ViTDetectionModel(nn.Module):
|
19 |
+
def __init__(self, num_queries=100, hidden_dim=768):
|
20 |
+
"""
|
21 |
+
Initializes the ViTDetectionModel.
|
22 |
+
|
23 |
+
Args:
|
24 |
+
num_queries (int, optional): Number of detection queries. Defaults to 100.
|
25 |
+
hidden_dim (int, optional): Hidden dimension size. Defaults to 768.
|
26 |
+
"""
|
27 |
+
super(ViTDetectionModel, self).__init__()
|
28 |
+
# Configure the ViT model to output features only
|
29 |
+
self.vit = timm.create_model(
|
30 |
+
'vit_base_patch16_224',
|
31 |
+
pretrained=False, # Set to False since we are loading a trained model
|
32 |
+
num_classes=0, # Disable classification head
|
33 |
+
features_only=True, # Return feature maps
|
34 |
+
out_indices=(11,) # Get the last feature map
|
35 |
+
)
|
36 |
+
self.query_embed = nn.Embedding(num_queries, hidden_dim)
|
37 |
+
self.fc_bbox = nn.Linear(hidden_dim, 8) # 4 points (x, y) for quadrilateral
|
38 |
+
self.fc_class = nn.Linear(hidden_dim, 1) # Binary classification
|
39 |
+
|
40 |
+
def forward(self, x):
|
41 |
+
"""
|
42 |
+
Forward pass of the detection model.
|
43 |
+
|
44 |
+
Args:
|
45 |
+
x (Tensor): Input images [batch, 3, H, W].
|
46 |
+
|
47 |
+
Returns:
|
48 |
+
Tuple[Tensor, Tensor]: Predicted bounding boxes and class scores.
|
49 |
+
"""
|
50 |
+
# Retrieve the feature map
|
51 |
+
features = self.vit(x)[0] # [batch, hidden_dim, H*W]
|
52 |
+
|
53 |
+
if features.dim() == 3:
|
54 |
+
batch_size, hidden_dim, num_patches = features.shape
|
55 |
+
grid_size = int(np.sqrt(num_patches))
|
56 |
+
if grid_size * grid_size != num_patches:
|
57 |
+
raise ValueError(f"Number of patches {num_patches} is not a perfect square.")
|
58 |
+
H, W = grid_size, grid_size
|
59 |
+
features = features.view(batch_size, hidden_dim, H, W)
|
60 |
+
elif features.dim() == 4:
|
61 |
+
batch_size, hidden_dim, H, W = features.shape
|
62 |
+
else:
|
63 |
+
raise ValueError(f"Unexpected feature dimensions: {features.dim()}, expected 3 or 4.")
|
64 |
+
|
65 |
+
# Flatten the spatial dimensions
|
66 |
+
features = features.flatten(2).transpose(1, 2) # [batch, H*W, hidden_dim]
|
67 |
+
|
68 |
+
# Prepare query embeddings
|
69 |
+
queries = self.query_embed.weight.unsqueeze(0).repeat(batch_size, 1, 1) # [batch, num_queries, hidden_dim]
|
70 |
+
|
71 |
+
# Compute attention weights
|
72 |
+
attn = torch.matmul(features, queries.transpose(-1, -2)) # [batch, H*W, num_queries]
|
73 |
+
attn = torch.softmax(attn, dim=1) # Softmax over patches
|
74 |
+
|
75 |
+
# Aggregate features based on attention
|
76 |
+
output = torch.matmul(attn.transpose(-1, -2), features) # [batch, num_queries, hidden_dim]
|
77 |
+
|
78 |
+
# Predict bounding boxes and classes
|
79 |
+
bboxes = self.fc_bbox(output) # [batch, num_queries, 8]
|
80 |
+
classes = self.fc_class(output) # [batch, num_queries, 1]
|
81 |
+
|
82 |
+
return bboxes, classes
|
83 |
+
|
84 |
+
# Function to Load the Trained Model
|
85 |
+
def load_model(model_path, device):
|
86 |
+
"""
|
87 |
+
Loads the trained detection model.
|
88 |
+
|
89 |
+
Args:
|
90 |
+
model_path (str): Path to the saved model state dictionary.
|
91 |
+
device (torch.device): Device to load the model on.
|
92 |
+
|
93 |
+
Returns:
|
94 |
+
nn.Module: Loaded detection model.
|
95 |
+
"""
|
96 |
+
model = ViTDetectionModel(num_queries=100, hidden_dim=768).to(device)
|
97 |
+
model.load_state_dict(torch.load(model_path, map_location=device))
|
98 |
+
model.eval()
|
99 |
+
return model
|
100 |
+
|
101 |
+
# Function to Perform Text Detection on an Image
|
102 |
+
def detect_text(image, model, device, max_boxes=100, confidence_threshold=0.5):
|
103 |
+
"""
|
104 |
+
Detects text in the input image using the detection model.
|
105 |
+
|
106 |
+
Args:
|
107 |
+
image (PIL Image): Input image.
|
108 |
+
model (nn.Module): Trained detection model.
|
109 |
+
device (torch.device): Device to run the model on.
|
110 |
+
max_boxes (int, optional): Maximum number of bounding boxes to return. Defaults to 100.
|
111 |
+
confidence_threshold (float, optional): Threshold to filter detections. Defaults to 0.5.
|
112 |
+
|
113 |
+
Returns:
|
114 |
+
PIL Image: Image with detected bounding boxes drawn.
|
115 |
+
"""
|
116 |
+
# Define transformation
|
117 |
+
transform = transforms.Compose([
|
118 |
+
transforms.Resize((224, 224)),
|
119 |
+
transforms.ToTensor(),
|
120 |
+
])
|
121 |
+
|
122 |
+
# Preprocess the image
|
123 |
+
input_tensor = transform(image).unsqueeze(0).to(device) # [1, 3, 224, 224]
|
124 |
+
|
125 |
+
# Perform detection
|
126 |
+
with torch.no_grad():
|
127 |
+
pred_bboxes, pred_classes = model(input_tensor) # [1, num_queries, 8], [1, num_queries, 1]
|
128 |
+
|
129 |
+
# Process predictions
|
130 |
+
pred_bboxes = pred_bboxes.squeeze(0) # [num_queries, 8]
|
131 |
+
pred_classes = pred_classes.squeeze(0) # [num_queries, 1]
|
132 |
+
pred_classes_sigmoid = torch.sigmoid(pred_classes)
|
133 |
+
high_conf_indices = (pred_classes_sigmoid > confidence_threshold).squeeze(1).nonzero(as_tuple=False).squeeze(1)
|
134 |
+
selected_indices = high_conf_indices[:max_boxes]
|
135 |
+
selected_bboxes = pred_bboxes[selected_indices] # [selected, 8]
|
136 |
+
|
137 |
+
# Denormalize bounding boxes to original image size
|
138 |
+
width, height = image.size
|
139 |
+
scale_x = width / 224
|
140 |
+
scale_y = height / 224
|
141 |
+
boxes = selected_bboxes.cpu().numpy() * np.array([scale_x, scale_y] * 4) # [selected, 8]
|
142 |
+
|
143 |
+
# Draw bounding boxes on the image
|
144 |
+
fig, ax = plt.subplots(1, figsize=(12, 12))
|
145 |
+
ax.imshow(image)
|
146 |
+
|
147 |
+
for box in boxes:
|
148 |
+
polygon = patches.Polygon(box.reshape(-1, 2), linewidth=2, edgecolor='r', facecolor='none')
|
149 |
+
ax.add_patch(polygon)
|
150 |
+
|
151 |
+
plt.axis('off')
|
152 |
+
# Convert Matplotlib figure to PIL Image
|
153 |
+
fig.canvas.draw()
|
154 |
+
img_with_boxes = Image.frombytes('RGB', fig.canvas.get_width_height(), fig.canvas.tostring_rgb())
|
155 |
+
plt.close(fig)
|
156 |
+
|
157 |
+
return img_with_boxes
|
158 |
+
|
159 |
+
# Optional: If integrating OCR with pytesseract
|
160 |
+
# def detect_and_recognize_text(image, model, device, max_boxes=100, confidence_threshold=0.5):
|
161 |
+
# # Similar to detect_text but includes OCR steps
|
162 |
+
# pass
|
163 |
+
|
164 |
+
# Initialize the model
|
165 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
166 |
+
model_path = "detection_model.pth" # Ensure this path matches where the model is stored
|
167 |
+
model = load_model(model_path, device)
|
168 |
+
print("Model loaded successfully.")
|
169 |
+
|
170 |
+
# Define the Gradio Interface Function
|
171 |
+
def gradio_detect(image):
|
172 |
+
"""
|
173 |
+
Gradio interface function for text detection.
|
174 |
+
|
175 |
+
Args:
|
176 |
+
image (PIL Image): Uploaded image.
|
177 |
+
|
178 |
+
Returns:
|
179 |
+
PIL Image: Image with detected bounding boxes.
|
180 |
+
"""
|
181 |
+
result_image = detect_text(image, model, device)
|
182 |
+
return result_image
|
183 |
+
|
184 |
+
# Create Gradio Interface
|
185 |
+
iface = gr.Interface(
|
186 |
+
fn=gradio_detect,
|
187 |
+
inputs=gr.Image(type="pil"),
|
188 |
+
outputs=gr.Image(type="pil"),
|
189 |
+
title="Text Detection with ViT",
|
190 |
+
description="Upload an image, and the model will detect and highlight text regions.",
|
191 |
+
examples=[
|
192 |
+
# You can add URLs or paths to example images here
|
193 |
+
# "https://example.com/image1.jpg",
|
194 |
+
# "https://example.com/image2.jpg",
|
195 |
+
],
|
196 |
+
allow_flagging="never"
|
197 |
+
)
|
198 |
+
|
199 |
+
# Launch the Gradio App (Optional for local testing)
|
200 |
+
# if __name__ == "__main__":
|
201 |
+
# iface.launch()
|