Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,102 +1,126 @@
|
|
1 |
-
|
2 |
-
import
|
3 |
-
import
|
4 |
-
|
5 |
-
import
|
6 |
-
import
|
7 |
-
|
8 |
-
|
|
|
9 |
|
10 |
|
|
|
11 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
12 |
|
13 |
-
#
|
14 |
class CompressedStudentModel(nn.Module):
|
15 |
def __init__(self):
|
|
|
16 |
super(CompressedStudentModel, self).__init__()
|
|
|
17 |
self.encoder = nn.Sequential(
|
18 |
-
nn.Conv2d(3, 64, kernel_size=3, padding=1),
|
|
|
|
|
19 |
nn.ReLU(),
|
20 |
-
nn.
|
|
|
21 |
nn.ReLU(),
|
22 |
-
nn.
|
23 |
-
nn.Conv2d(64, 128, kernel_size=3, padding=1),
|
24 |
nn.ReLU(),
|
25 |
-
nn.
|
|
|
26 |
nn.ReLU(),
|
27 |
-
nn.
|
28 |
-
nn.Conv2d(128, 256, kernel_size=3, padding=1),
|
29 |
-
nn.ReLU(),
|
30 |
-
nn.Conv2d(256, 256, kernel_size=3, padding=1),
|
31 |
nn.ReLU(),
|
32 |
)
|
|
|
33 |
self.decoder = nn.Sequential(
|
34 |
-
nn.ConvTranspose2d(256, 128, kernel_size=3, stride=2, padding=1, output_padding=1),
|
35 |
nn.ReLU(),
|
36 |
-
nn.ConvTranspose2d(128, 64, kernel_size=3, stride=2, padding=1, output_padding=1),
|
37 |
nn.ReLU(),
|
38 |
-
nn.Conv2d(64, 1, kernel_size=3, padding=1),
|
39 |
)
|
40 |
|
41 |
def forward(self, x):
|
|
|
42 |
features = self.encoder(x)
|
|
|
43 |
depth = self.decoder(features)
|
44 |
return depth
|
45 |
|
46 |
-
#
|
47 |
-
model = CompressedStudentModel().to(device)
|
48 |
-
model.load_state_dict(torch.load("huntrezz_depth_v2.pt", map_location=device))
|
49 |
-
model.eval()
|
50 |
|
|
|
51 |
processor = DPTImageProcessor.from_pretrained("Intel/dpt-swinv2-tiny-256")
|
52 |
|
53 |
def preprocess_image(image):
|
|
|
54 |
image = cv2.resize(image, (200, 200))
|
|
|
55 |
image = torch.from_numpy(image).permute(2, 0, 1).unsqueeze(0).float().to(device)
|
|
|
56 |
return image / 255.0
|
57 |
|
58 |
def plot_depth_map(depth_map, original_image):
|
|
|
59 |
fig = plt.figure(figsize=(16, 9))
|
|
|
60 |
ax = fig.add_subplot(111, projection='3d')
|
|
|
61 |
x, y = np.meshgrid(range(depth_map.shape[1]), range(depth_map.shape[0]))
|
62 |
|
63 |
-
# Normalize depth
|
64 |
norm = plt.Normalize(depth_map.min(), depth_map.max())
|
65 |
colors = plt.cm.viridis(norm(depth_map))
|
66 |
|
|
|
67 |
ax.plot_surface(x, y, depth_map, facecolors=colors, shade=False)
|
68 |
-
ax.set_zlim(0, 1)
|
69 |
|
70 |
-
#
|
71 |
ax.view_init(elev=70, azim=90)
|
72 |
-
plt.axis('off')
|
73 |
-
plt.close(fig)
|
74 |
|
|
|
75 |
fig.canvas.draw()
|
76 |
img = np.frombuffer(fig.canvas.tostring_rgb(), dtype=np.uint8)
|
77 |
img = img.reshape(fig.canvas.get_width_height()[::-1] + (3,))
|
78 |
|
79 |
return img
|
80 |
|
81 |
-
@torch.inference_mode()
|
82 |
def process_frame(image):
|
|
|
83 |
if image is None:
|
84 |
return None
|
|
|
85 |
preprocessed = preprocess_image(image)
|
|
|
86 |
predicted_depth = model(preprocessed).squeeze().cpu().numpy()
|
87 |
|
|
|
88 |
depth_map = (predicted_depth - predicted_depth.min()) / (predicted_depth.max() - predicted_depth.min())
|
89 |
|
|
|
90 |
if image.shape[2] == 3:
|
91 |
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
|
92 |
|
|
|
93 |
return plot_depth_map(depth_map, image)
|
94 |
|
|
|
95 |
interface = gr.Interface(
|
96 |
-
fn=process_frame,
|
97 |
-
inputs=gr.Image(sources="webcam", streaming=True),
|
98 |
-
outputs="image",
|
99 |
-
live=True
|
100 |
)
|
101 |
|
|
|
102 |
interface.launch()
|
|
|
1 |
+
# Import required libraries for image processing, deep learning, and visualization
|
2 |
+
import cv2 # OpenCV for image processing
|
3 |
+
import torch # PyTorch deep learning framework
|
4 |
+
import numpy as np # NumPy for numerical operations
|
5 |
+
from transformers import DPTImageProcessor # Hugging Face image processor for depth estimation
|
6 |
+
import gradio as gr # Gradio for creating web interfaces
|
7 |
+
import matplotlib.pyplot as plt # Matplotlib for plotting
|
8 |
+
from mpl_toolkits.mplot3d import Axes3D # 3D plotting tools
|
9 |
+
import torch.nn as nn # Neural network modules from PyTorch
|
10 |
|
11 |
|
12 |
+
# Set up device - will use GPU if available, otherwise CPU
|
13 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
14 |
|
15 |
+
# Define my compressed student model architecture for depth estimation
|
16 |
class CompressedStudentModel(nn.Module):
|
17 |
def __init__(self):
|
18 |
+
# Initialize parent class
|
19 |
super(CompressedStudentModel, self).__init__()
|
20 |
+
# Define encoder network that extracts features from input image
|
21 |
self.encoder = nn.Sequential(
|
22 |
+
nn.Conv2d(3, 64, kernel_size=3, padding=1), # First conv layer: RGB -> 64 channels
|
23 |
+
nn.ReLU(), # Activation function
|
24 |
+
nn.Conv2d(64, 64, kernel_size=3, padding=1), # Second conv: 64 -> 64 channels
|
25 |
nn.ReLU(),
|
26 |
+
nn.MaxPool2d(2), # Reduce spatial dimensions by 2
|
27 |
+
nn.Conv2d(64, 128, kernel_size=3, padding=1), # Third conv: 64 -> 128 channels
|
28 |
nn.ReLU(),
|
29 |
+
nn.Conv2d(128, 128, kernel_size=3, padding=1), # Fourth conv: 128 -> 128 channels
|
|
|
30 |
nn.ReLU(),
|
31 |
+
nn.MaxPool2d(2), # Further reduce spatial dimensions
|
32 |
+
nn.Conv2d(128, 256, kernel_size=3, padding=1), # Fifth conv: 128 -> 256 channels
|
33 |
nn.ReLU(),
|
34 |
+
nn.Conv2d(256, 256, kernel_size=3, padding=1), # Sixth conv: 256 -> 256 channels
|
|
|
|
|
|
|
35 |
nn.ReLU(),
|
36 |
)
|
37 |
+
# Define decoder network that upsamples features back to original resolution
|
38 |
self.decoder = nn.Sequential(
|
39 |
+
nn.ConvTranspose2d(256, 128, kernel_size=3, stride=2, padding=1, output_padding=1), # First upsample: 256 -> 128
|
40 |
nn.ReLU(),
|
41 |
+
nn.ConvTranspose2d(128, 64, kernel_size=3, stride=2, padding=1, output_padding=1), # Second upsample: 128 -> 64
|
42 |
nn.ReLU(),
|
43 |
+
nn.Conv2d(64, 1, kernel_size=3, padding=1), # Final conv: 64 -> 1 channel depth map
|
44 |
)
|
45 |
|
46 |
def forward(self, x):
|
47 |
+
# Pass input through encoder to get features
|
48 |
features = self.encoder(x)
|
49 |
+
# Pass features through decoder to get depth map
|
50 |
depth = self.decoder(features)
|
51 |
return depth
|
52 |
|
53 |
+
# Load my trained model and prepare it for inference
|
54 |
+
model = CompressedStudentModel().to(device) # Create model instance and move to device
|
55 |
+
model.load_state_dict(torch.load("huntrezz_depth_v2.pt", map_location=device)) # Load trained weights
|
56 |
+
model.eval() # Set model to evaluation mode
|
57 |
|
58 |
+
# Initialize the image processor from Hugging Face
|
59 |
processor = DPTImageProcessor.from_pretrained("Intel/dpt-swinv2-tiny-256")
|
60 |
|
61 |
def preprocess_image(image):
|
62 |
+
# Resize image to 200x200 for consistent processing
|
63 |
image = cv2.resize(image, (200, 200))
|
64 |
+
# Convert image to PyTorch tensor and move to device
|
65 |
image = torch.from_numpy(image).permute(2, 0, 1).unsqueeze(0).float().to(device)
|
66 |
+
# Normalize pixel values to [0,1] range
|
67 |
return image / 255.0
|
68 |
|
69 |
def plot_depth_map(depth_map, original_image):
|
70 |
+
# Create new figure with specific size
|
71 |
fig = plt.figure(figsize=(16, 9))
|
72 |
+
# Add 3D subplot
|
73 |
ax = fig.add_subplot(111, projection='3d')
|
74 |
+
# Create coordinate grids for 3D plot
|
75 |
x, y = np.meshgrid(range(depth_map.shape[1]), range(depth_map.shape[0]))
|
76 |
|
77 |
+
# Normalize depth values for coloring
|
78 |
norm = plt.Normalize(depth_map.min(), depth_map.max())
|
79 |
colors = plt.cm.viridis(norm(depth_map))
|
80 |
|
81 |
+
# Create 3D surface plot
|
82 |
ax.plot_surface(x, y, depth_map, facecolors=colors, shade=False)
|
83 |
+
ax.set_zlim(0, 1) # Set z-axis limits
|
84 |
|
85 |
+
# Set viewing angle for better visualization
|
86 |
ax.view_init(elev=70, azim=90)
|
87 |
+
plt.axis('off') # Hide axes
|
88 |
+
plt.close(fig) # Close the figure to free memory
|
89 |
|
90 |
+
# Convert matplotlib figure to numpy array
|
91 |
fig.canvas.draw()
|
92 |
img = np.frombuffer(fig.canvas.tostring_rgb(), dtype=np.uint8)
|
93 |
img = img.reshape(fig.canvas.get_width_height()[::-1] + (3,))
|
94 |
|
95 |
return img
|
96 |
|
97 |
+
@torch.inference_mode() # Disable gradient computation for inference
|
98 |
def process_frame(image):
|
99 |
+
# Check if image is valid
|
100 |
if image is None:
|
101 |
return None
|
102 |
+
# Preprocess input image
|
103 |
preprocessed = preprocess_image(image)
|
104 |
+
# Get depth prediction from model
|
105 |
predicted_depth = model(preprocessed).squeeze().cpu().numpy()
|
106 |
|
107 |
+
# Normalize depth values to [0,1] range
|
108 |
depth_map = (predicted_depth - predicted_depth.min()) / (predicted_depth.max() - predicted_depth.min())
|
109 |
|
110 |
+
# Convert BGR to RGB if needed
|
111 |
if image.shape[2] == 3:
|
112 |
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
|
113 |
|
114 |
+
# Create and return 3D visualization
|
115 |
return plot_depth_map(depth_map, image)
|
116 |
|
117 |
+
# Create Gradio interface for webcam input
|
118 |
interface = gr.Interface(
|
119 |
+
fn=process_frame, # Processing function
|
120 |
+
inputs=gr.Image(sources="webcam", streaming=True), # Webcam input
|
121 |
+
outputs="image", # Image output
|
122 |
+
live=True # Enable live updates
|
123 |
)
|
124 |
|
125 |
+
# Launch the interface
|
126 |
interface.launch()
|