VIVEK JAYARAM
commited on
Commit
·
235a140
1
Parent(s):
22a317f
bug fix
Browse files
app.py
CHANGED
@@ -5,7 +5,6 @@ import yaml
|
|
5 |
import os
|
6 |
import numpy as np
|
7 |
from PIL import Image
|
8 |
-
import time
|
9 |
from cdim.noise import get_noise
|
10 |
from cdim.operators import get_operator
|
11 |
from cdim.image_utils import save_to_image
|
@@ -15,13 +14,11 @@ from cdim.diffusion.diffusion_pipeline import run_diffusion
|
|
15 |
from cdim.eta_scheduler import EtaScheduler
|
16 |
from diffusers import DiffusionPipeline
|
17 |
|
18 |
-
|
19 |
-
# Global variables for model and scheduler
|
20 |
model = None
|
21 |
ddim_scheduler = None
|
22 |
model_type = None
|
23 |
|
24 |
-
|
25 |
def load_image(image_path):
|
26 |
"""Process input image to tensor format."""
|
27 |
image = Image.open(image_path)
|
@@ -29,23 +26,30 @@ def load_image(image_path):
|
|
29 |
original_image = torch.from_numpy(original_image).unsqueeze(0).permute(0, 3, 1, 2)
|
30 |
return (original_image / 127.5 - 1.0).to(torch.float)[:, :3]
|
31 |
|
32 |
-
|
33 |
def load_yaml(file_path: str) -> dict:
|
34 |
-
"""Load configurations from a YAML file."""
|
35 |
with open(file_path) as f:
|
36 |
config = yaml.load(f, Loader=yaml.FullLoader)
|
37 |
return config
|
38 |
|
39 |
-
|
40 |
def convert_to_np(torch_image):
|
41 |
return ((torch_image.detach().clamp(-1, 1).cpu().numpy().transpose(1, 2, 0) + 1) * 127.5).astype(np.uint8)
|
42 |
|
43 |
-
|
44 |
@spaces.GPU
|
45 |
-
def
|
46 |
-
"""
|
47 |
-
# Map image choice to path
|
48 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
49 |
|
50 |
image_paths = {
|
51 |
"CelebA HQ 1": "sample_images/celebhq_29999.jpg",
|
@@ -55,70 +59,37 @@ def generate_noisy_image(image_choice, noise_sigma, operator_key):
|
|
55 |
|
56 |
config_paths = {
|
57 |
"Box Inpainting": "operator_configs/box_inpainting_config.yaml",
|
58 |
-
"Random Inpainting": "operator_configs/random_inpainting_config.yaml",
|
59 |
"Super Resolution": "operator_configs/super_resolution_config.yaml",
|
60 |
"Gaussian Deblur": "operator_configs/gaussian_blur_config.yaml"
|
61 |
}
|
62 |
|
|
|
63 |
image_path = image_paths[image_choice]
|
64 |
-
|
65 |
-
# Load image and get noisy version
|
66 |
original_image = load_image(image_path).to(device)
|
|
|
67 |
noise_config = load_yaml("noise_configs/gaussian_noise_config.yaml")
|
68 |
noise_config["sigma"] = noise_sigma
|
69 |
noise_function = get_noise(**noise_config)
|
|
|
70 |
operator_config = load_yaml(config_paths[operator_key])
|
71 |
operator_config["device"] = device
|
72 |
operator = get_operator(**operator_config)
|
73 |
-
|
74 |
noisy_measurement = noise_function(operator(original_image))
|
75 |
noisy_image = Image.fromarray(convert_to_np(noisy_measurement[0]))
|
76 |
|
77 |
-
# Store necessary data for restoration
|
78 |
-
data = {
|
79 |
-
'noisy_measurement': noisy_measurement.cpu(),
|
80 |
-
'operator': operator,
|
81 |
-
'noise_function': noise_function
|
82 |
-
}
|
83 |
-
|
84 |
-
return noisy_image, data # Return the noisy image and data for restoration
|
85 |
-
|
86 |
-
@spaces.GPU
|
87 |
-
def run_restoration(data, T, K):
|
88 |
-
"""Run the restoration process and return the restored image."""
|
89 |
-
global model, ddim_scheduler, model_type
|
90 |
-
|
91 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
92 |
-
|
93 |
-
# Extract stored data
|
94 |
-
noisy_measurement = data['noisy_measurement'].to(device)
|
95 |
-
operator = data['operator']
|
96 |
-
noise_function = data['noise_function']
|
97 |
-
|
98 |
-
# Initialize model if not already done
|
99 |
-
if model is None:
|
100 |
-
model_type = "diffusers"
|
101 |
-
model = DiffusionPipeline.from_pretrained("google/ddpm-celebahq-256").to(device).unet
|
102 |
-
|
103 |
-
ddim_scheduler = DDIMScheduler(
|
104 |
-
num_train_timesteps=1000,
|
105 |
-
beta_start=0.0001,
|
106 |
-
beta_end=0.02,
|
107 |
-
beta_schedule="linear"
|
108 |
-
)
|
109 |
-
|
110 |
# Run restoration
|
111 |
eta_scheduler = EtaScheduler("gradnorm", operator.name, T, K, 'l2', noise_function, None)
|
112 |
output_image = run_diffusion(
|
113 |
model, ddim_scheduler, noisy_measurement, operator, noise_function, device,
|
114 |
eta_scheduler, num_inference_steps=T, K=K, model_type=model_type, loss_type='l2'
|
115 |
)
|
116 |
-
|
117 |
-
# Convert output image for display
|
118 |
output_image = Image.fromarray(convert_to_np(output_image[0]))
|
119 |
-
return output_image
|
120 |
-
|
121 |
|
|
|
122 |
with gr.Blocks() as demo:
|
123 |
gr.Markdown("# Noisy Image Restoration with Diffusion Models")
|
124 |
|
@@ -142,19 +113,13 @@ with gr.Blocks() as demo:
|
|
142 |
run_button = gr.Button("Run Inference")
|
143 |
noisy_image = gr.Image(label="Noisy Image")
|
144 |
restored_image = gr.Image(label="Restored Image")
|
145 |
-
state = gr.State() # To store intermediate data
|
146 |
|
147 |
-
#
|
148 |
run_button.click(
|
149 |
-
fn=
|
150 |
-
inputs=[image_select, noise_sigma, operator_select],
|
151 |
-
outputs=[noisy_image,
|
152 |
-
).then(
|
153 |
-
fn=run_restoration,
|
154 |
-
inputs=[state, T, K],
|
155 |
-
outputs=restored_image
|
156 |
)
|
157 |
|
158 |
-
|
159 |
if __name__ == "__main__":
|
160 |
-
demo.launch(server_name="0.0.0.0", server_port=7860)
|
|
|
5 |
import os
|
6 |
import numpy as np
|
7 |
from PIL import Image
|
|
|
8 |
from cdim.noise import get_noise
|
9 |
from cdim.operators import get_operator
|
10 |
from cdim.image_utils import save_to_image
|
|
|
14 |
from cdim.eta_scheduler import EtaScheduler
|
15 |
from diffusers import DiffusionPipeline
|
16 |
|
17 |
+
# Global variables moved inside GPU-decorated functions
|
|
|
18 |
model = None
|
19 |
ddim_scheduler = None
|
20 |
model_type = None
|
21 |
|
|
|
22 |
def load_image(image_path):
|
23 |
"""Process input image to tensor format."""
|
24 |
image = Image.open(image_path)
|
|
|
26 |
original_image = torch.from_numpy(original_image).unsqueeze(0).permute(0, 3, 1, 2)
|
27 |
return (original_image / 127.5 - 1.0).to(torch.float)[:, :3]
|
28 |
|
|
|
29 |
def load_yaml(file_path: str) -> dict:
|
|
|
30 |
with open(file_path) as f:
|
31 |
config = yaml.load(f, Loader=yaml.FullLoader)
|
32 |
return config
|
33 |
|
|
|
34 |
def convert_to_np(torch_image):
|
35 |
return ((torch_image.detach().clamp(-1, 1).cpu().numpy().transpose(1, 2, 0) + 1) * 127.5).astype(np.uint8)
|
36 |
|
|
|
37 |
@spaces.GPU
|
38 |
+
def process_image(image_choice, noise_sigma, operator_key, T, K):
|
39 |
+
"""Combined function to handle both generation and restoration"""
|
|
|
40 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
41 |
+
|
42 |
+
# Initialize model inside GPU-decorated function
|
43 |
+
global model, ddim_scheduler, model_type
|
44 |
+
if model is None:
|
45 |
+
model_type = "diffusers"
|
46 |
+
model = DiffusionPipeline.from_pretrained("google/ddpm-celebahq-256").to(device).unet
|
47 |
+
ddim_scheduler = DDIMScheduler(
|
48 |
+
num_train_timesteps=1000,
|
49 |
+
beta_start=0.0001,
|
50 |
+
beta_end=0.02,
|
51 |
+
beta_schedule="linear"
|
52 |
+
)
|
53 |
|
54 |
image_paths = {
|
55 |
"CelebA HQ 1": "sample_images/celebhq_29999.jpg",
|
|
|
59 |
|
60 |
config_paths = {
|
61 |
"Box Inpainting": "operator_configs/box_inpainting_config.yaml",
|
62 |
+
"Random Inpainting": "operator_configs/random_inpainting_config.yaml",
|
63 |
"Super Resolution": "operator_configs/super_resolution_config.yaml",
|
64 |
"Gaussian Deblur": "operator_configs/gaussian_blur_config.yaml"
|
65 |
}
|
66 |
|
67 |
+
# Generate noisy image
|
68 |
image_path = image_paths[image_choice]
|
|
|
|
|
69 |
original_image = load_image(image_path).to(device)
|
70 |
+
|
71 |
noise_config = load_yaml("noise_configs/gaussian_noise_config.yaml")
|
72 |
noise_config["sigma"] = noise_sigma
|
73 |
noise_function = get_noise(**noise_config)
|
74 |
+
|
75 |
operator_config = load_yaml(config_paths[operator_key])
|
76 |
operator_config["device"] = device
|
77 |
operator = get_operator(**operator_config)
|
78 |
+
|
79 |
noisy_measurement = noise_function(operator(original_image))
|
80 |
noisy_image = Image.fromarray(convert_to_np(noisy_measurement[0]))
|
81 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
82 |
# Run restoration
|
83 |
eta_scheduler = EtaScheduler("gradnorm", operator.name, T, K, 'l2', noise_function, None)
|
84 |
output_image = run_diffusion(
|
85 |
model, ddim_scheduler, noisy_measurement, operator, noise_function, device,
|
86 |
eta_scheduler, num_inference_steps=T, K=K, model_type=model_type, loss_type='l2'
|
87 |
)
|
88 |
+
|
|
|
89 |
output_image = Image.fromarray(convert_to_np(output_image[0]))
|
90 |
+
return noisy_image, output_image
|
|
|
91 |
|
92 |
+
# Gradio interface
|
93 |
with gr.Blocks() as demo:
|
94 |
gr.Markdown("# Noisy Image Restoration with Diffusion Models")
|
95 |
|
|
|
113 |
run_button = gr.Button("Run Inference")
|
114 |
noisy_image = gr.Image(label="Noisy Image")
|
115 |
restored_image = gr.Image(label="Restored Image")
|
|
|
116 |
|
117 |
+
# Single function call instead of chaining
|
118 |
run_button.click(
|
119 |
+
fn=process_image,
|
120 |
+
inputs=[image_select, noise_sigma, operator_select, T, K],
|
121 |
+
outputs=[noisy_image, restored_image]
|
|
|
|
|
|
|
|
|
122 |
)
|
123 |
|
|
|
124 |
if __name__ == "__main__":
|
125 |
+
demo.launch(server_name="0.0.0.0", server_port=7860)
|