Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,114 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import spaces
|
3 |
+
import gradio as gr
|
4 |
+
import numpy as np
|
5 |
+
import tensorflow as tf
|
6 |
+
from keras.applications.mobilenet_v2 import MobileNetV2, preprocess_input
|
7 |
+
from keras.models import Model
|
8 |
+
import matplotlib.pyplot as plt
|
9 |
+
import logging
|
10 |
+
from skimage.transform import resize
|
11 |
+
from PIL import Image, ImageEnhance, ImageFilter
|
12 |
+
from tqdm import tqdm
|
13 |
+
|
14 |
+
# Disable GPU usage by default
|
15 |
+
os.environ['CUDA_VISIBLE_DEVICES'] = ''
|
16 |
+
|
17 |
+
# HDC Encoding and Decoding Functions
|
18 |
+
DIM = 1000 # Hypervector dimensionality
|
19 |
+
|
20 |
+
def pixel_to_hypervector(pixel):
|
21 |
+
"""Convert a pixel intensity to a high-dimensional binary hypervector."""
|
22 |
+
np.random.seed(int(pixel))
|
23 |
+
return np.random.choice([1, -1], size=(DIM,))
|
24 |
+
|
25 |
+
def image_to_hdc(image):
|
26 |
+
"""Encode the entire image into hypervectors (by pixel intensity)."""
|
27 |
+
return np.array([pixel_to_hypervector(p) for p in image.flatten()])
|
28 |
+
|
29 |
+
def hdc_to_image(hdc_vectors, shape):
|
30 |
+
"""Decode hypervectors back into an image."""
|
31 |
+
decoded_pixels = np.mean(hdc_vectors, axis=1) # Aggregate hypervector values
|
32 |
+
decoded_pixels = np.clip((decoded_pixels + 1) / 2 * 255, 0, 255) # Rescale to [0, 255]
|
33 |
+
return decoded_pixels.reshape(shape).astype(np.uint8)
|
34 |
+
|
35 |
+
class SwarmAgent:
|
36 |
+
def __init__(self, position, velocity):
|
37 |
+
self.position = position
|
38 |
+
self.velocity = velocity
|
39 |
+
|
40 |
+
class SwarmNeuralNetwork:
|
41 |
+
def __init__(self, num_agents, image_shape, target_image_path):
|
42 |
+
self.image_shape = image_shape
|
43 |
+
self.agents = [SwarmAgent(self.random_position(), self.random_velocity()) for _ in range(num_agents)]
|
44 |
+
self.target_image = self.load_target_image(target_image_path)
|
45 |
+
self.mobilenet = self.load_mobilenet_model()
|
46 |
+
|
47 |
+
def random_position(self):
|
48 |
+
return np.random.randn(*self.image_shape)
|
49 |
+
|
50 |
+
def random_velocity(self):
|
51 |
+
return np.random.randn(*self.image_shape) * 0.01
|
52 |
+
|
53 |
+
def load_target_image(self, img_path):
|
54 |
+
img = Image.open(img_path).resize((self.image_shape[1], self.image_shape[0]))
|
55 |
+
return np.array(img) / 127.5 - 1
|
56 |
+
|
57 |
+
def load_mobilenet_model(self):
|
58 |
+
mobilenet = MobileNetV2(weights='imagenet', include_top=False, input_shape=(128, 128, 3))
|
59 |
+
return Model(inputs=mobilenet.input, outputs=mobilenet.get_layer('block_13_expand_relu').output)
|
60 |
+
|
61 |
+
def update_agents(self, timestep):
|
62 |
+
for agent in self.agents:
|
63 |
+
# Convert agent's position and target image into HDC space
|
64 |
+
agent_hdc = image_to_hdc(agent.position)
|
65 |
+
target_hdc = image_to_hdc(self.target_image)
|
66 |
+
|
67 |
+
# Compute similarity between the agent's position and the target image
|
68 |
+
similarity = np.mean(agent_hdc * target_hdc, axis=1) # Cosine-like similarity
|
69 |
+
attention = similarity / np.sum(similarity)
|
70 |
+
|
71 |
+
# Adjust the agent's position based on HDC-guided noise reduction
|
72 |
+
noise = np.random.randn(*self.image_shape) * 0.1
|
73 |
+
agent.position += attention.reshape(self.image_shape) * noise
|
74 |
+
|
75 |
+
# Clip values to ensure valid range
|
76 |
+
agent.position = np.clip(agent.position, -1, 1)
|
77 |
+
|
78 |
+
def generate_image(self):
|
79 |
+
generated_image = np.mean([agent.position for agent in self.agents], axis=0)
|
80 |
+
return (generated_image + 1) / 2
|
81 |
+
|
82 |
+
def train(self, epochs):
|
83 |
+
for epoch in tqdm(range(epochs), desc="Training"):
|
84 |
+
self.update_agents(epoch)
|
85 |
+
generated_image = self.generate_image()
|
86 |
+
|
87 |
+
# Display the generated image
|
88 |
+
self.display_image(generated_image, title=f'Epoch {epoch}')
|
89 |
+
|
90 |
+
def display_image(self, image, title=''):
|
91 |
+
plt.imshow(image)
|
92 |
+
plt.title(title)
|
93 |
+
plt.axis('off')
|
94 |
+
plt.show()
|
95 |
+
|
96 |
+
# Gradio Interface
|
97 |
+
def train_snn(image_path, num_agents, epochs):
|
98 |
+
snn = SwarmNeuralNetwork(num_agents=num_agents, image_shape=(128, 128, 3), target_image_path=image_path)
|
99 |
+
snn.train(epochs=epochs)
|
100 |
+
generated_image = snn.generate_image()
|
101 |
+
return generated_image
|
102 |
+
|
103 |
+
interface = gr.Interface(
|
104 |
+
fn=train_snn,
|
105 |
+
inputs=[
|
106 |
+
gr.Image(type="filepath", label="Upload Target Image"),
|
107 |
+
gr.Slider(minimum=100, maximum=1000, value=500, label="Number of Agents"),
|
108 |
+
gr.Slider(minimum=5, maximum=20, value=10, label="Number of Epochs")
|
109 |
+
],
|
110 |
+
outputs=gr.Image(type="numpy", label="Generated Image"),
|
111 |
+
title="HDC Swarm Neural Network Image Generation"
|
112 |
+
)
|
113 |
+
|
114 |
+
interface.launch()
|