Upload folder using huggingface_hub
Browse files- gradio_demo.py +359 -0
gradio_demo.py
ADDED
@@ -0,0 +1,359 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import numpy as np
|
5 |
+
import gradio as gr
|
6 |
+
from PIL import Image
|
7 |
+
import math
|
8 |
+
import torch.nn.functional as F
|
9 |
+
import os
|
10 |
+
import tempfile
|
11 |
+
import time
|
12 |
+
import threading
|
13 |
+
|
14 |
+
from utils.hatropeamp import HATNOUP_ROPE_AMP
|
15 |
+
from utils.fea2gsropeamp import Fea2GS_ROPE_AMP
|
16 |
+
from utils.edsrbaseline import EDSRNOUP
|
17 |
+
from utils.hatropeamp import HATNOUP_ROPE_AMP
|
18 |
+
from utils.rdn import RDNNOUP
|
19 |
+
from utils.swinir import SwinIRNOUP
|
20 |
+
from utils.fea2gsropeamp import Fea2GS_ROPE_AMP
|
21 |
+
from utils.gaussian_splatting import generate_2D_gaussian_splatting_step
|
22 |
+
from utils.split_and_joint_image import split_and_joint_image
|
23 |
+
from huggingface_hub import hf_hub_download
|
24 |
+
import subprocess
|
25 |
+
import sys
|
26 |
+
import spaces
|
27 |
+
|
28 |
+
|
29 |
+
|
30 |
+
# Device setup
|
31 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
32 |
+
|
33 |
+
# Global stop flag for interrupting inference
|
34 |
+
stop_inference = False
|
35 |
+
inference_lock = threading.Lock()
|
36 |
+
|
37 |
+
def load_model(
|
38 |
+
pretrained_model_name_or_path: str = "mutou0308/GSASR",
|
39 |
+
model_name: str = "HATL_SA1B",
|
40 |
+
device: str | torch.device = "cuda"
|
41 |
+
):
|
42 |
+
enc_path = hf_hub_download(
|
43 |
+
repo_id=pretrained_model_name_or_path, filename=os.path.join('GSASR_enhenced_ultra', model_name, 'encoder.pth')
|
44 |
+
)
|
45 |
+
dec_path = hf_hub_download(
|
46 |
+
repo_id=pretrained_model_name_or_path, filename=os.path.join('GSASR_enhenced_ultra', model_name, 'decoder.pth')
|
47 |
+
)
|
48 |
+
|
49 |
+
enc_weight = torch.load(enc_path, weights_only=True)['params_ema']
|
50 |
+
dec_weight = torch.load(dec_path, weights_only=True)['params_ema']
|
51 |
+
|
52 |
+
if model_name in ['EDSR_DIV2K', 'EDSR_DF2K']:
|
53 |
+
encoder = EDSRNOUP()
|
54 |
+
decoder = Fea2GS_ROPE_AMP()
|
55 |
+
elif model_name in ['RDN_DIV2K', 'RDN_DF2K']:
|
56 |
+
encoder = RDNNOUP()
|
57 |
+
decoder = Fea2GS_ROPE_AMP(num_crossattn_blocks = 2)
|
58 |
+
elif model_name in ['SwinIR_DIV2K', 'SwinIR_DF2K']:
|
59 |
+
encoder = SwinIRNOUP()
|
60 |
+
decoder = Fea2GS_ROPE_AMP(num_crossattn_blocks=2, num_crossattn_layers=4, num_gs_seed=256, window_size=16)
|
61 |
+
elif model_name in ['HATL_SA1B']:
|
62 |
+
encoder = HATNOUP_ROPE_AMP()
|
63 |
+
decoder = Fea2GS_ROPE_AMP(channel=192, num_crossattn_blocks=4, num_crossattn_layers=4, num_selfattn_blocks=8, num_selfattn_layers=6,
|
64 |
+
num_gs_seed=256, window_size=16)
|
65 |
+
else:
|
66 |
+
raise ValueError(f"args.model-{model_name} must be in ['EDSR_DIV2K', 'EDSR_DF2K', 'RDN_DIV2K', 'RDN_DF2K', 'SwinIR_DIV2K', 'SwinIR_DF2K', 'HATL_SA1B']")
|
67 |
+
|
68 |
+
encoder.load_state_dict(enc_weight, strict=True)
|
69 |
+
decoder.load_state_dict(dec_weight, strict=True)
|
70 |
+
encoder.eval()
|
71 |
+
decoder.eval()
|
72 |
+
encoder = encoder.to(device)
|
73 |
+
decoder = decoder.to(device)
|
74 |
+
return encoder, decoder
|
75 |
+
|
76 |
+
|
77 |
+
def preprocess(x, denominator=16):
|
78 |
+
"""Preprocess image to ensure dimensions are multiples of denominator"""
|
79 |
+
_, c, h, w = x.shape
|
80 |
+
if h % denominator > 0:
|
81 |
+
pad_h = denominator - h % denominator
|
82 |
+
else:
|
83 |
+
pad_h = 0
|
84 |
+
if w % denominator > 0:
|
85 |
+
pad_w = denominator - w % denominator
|
86 |
+
else:
|
87 |
+
pad_w = 0
|
88 |
+
x_new = F.pad(x, (0, pad_w, 0, pad_h), 'reflect')
|
89 |
+
return x_new
|
90 |
+
|
91 |
+
def postprocess(x, gt_size_h, gt_size_w):
|
92 |
+
"""Post-process by cropping to target size"""
|
93 |
+
x_new = x[:, :, :gt_size_h, :gt_size_w]
|
94 |
+
return x_new
|
95 |
+
|
96 |
+
def should_use_tile(image_height, image_width, threshold=1024):
|
97 |
+
"""Determine if tile processing should be used based on image resolution"""
|
98 |
+
return max(image_height, image_width) > threshold
|
99 |
+
|
100 |
+
def set_stop_flag():
|
101 |
+
"""Set the global stop flag to interrupt inference"""
|
102 |
+
global stop_inference
|
103 |
+
with inference_lock:
|
104 |
+
stop_inference = True
|
105 |
+
return "π Stopping inference...", gr.update(interactive=False)
|
106 |
+
|
107 |
+
def reset_stop_flag():
|
108 |
+
"""Reset the global stop flag"""
|
109 |
+
global stop_inference
|
110 |
+
with inference_lock:
|
111 |
+
stop_inference = False
|
112 |
+
|
113 |
+
def check_stop_flag():
|
114 |
+
"""Check if inference should be stopped"""
|
115 |
+
global stop_inference
|
116 |
+
with inference_lock:
|
117 |
+
return stop_inference
|
118 |
+
|
119 |
+
@spaces.GPU
|
120 |
+
def super_resolution_inference(image, scale=4.0):
|
121 |
+
"""Super-resolution inference function with automatic tile processing"""
|
122 |
+
|
123 |
+
# Check if gscuda setup has been run
|
124 |
+
setup_marker = ".setup_complete"
|
125 |
+
if not os.path.exists(setup_marker):
|
126 |
+
print("First run detected, installing dependencies...")
|
127 |
+
try:
|
128 |
+
# subprocess.check_call(["pip", "install", "-e", "."])
|
129 |
+
subprocess.check_call(["pip", "install", "dist/gscuda-0.0.0-cp310-cp310-linux_x86_64.whl"])
|
130 |
+
# Create marker file to indicate setup is complete
|
131 |
+
with open(setup_marker, "w") as f:
|
132 |
+
f.write("Setup completed")
|
133 |
+
print("Setup completed successfully!")
|
134 |
+
except subprocess.CalledProcessError as e:
|
135 |
+
return None, f"β Setup failed with error: {e}", None
|
136 |
+
|
137 |
+
|
138 |
+
|
139 |
+
if image is None:
|
140 |
+
return None, "Please upload an image", None
|
141 |
+
|
142 |
+
# Load model
|
143 |
+
encoder, decoder = load_model(model_name="HATL_SA1B")
|
144 |
+
|
145 |
+
# Reset stop flag at the beginning
|
146 |
+
reset_stop_flag()
|
147 |
+
|
148 |
+
# Fixed parameters
|
149 |
+
tile_overlap = 16 # Fixed overlap size
|
150 |
+
crop_size = 8 # Fixed crop size
|
151 |
+
tile_size = 1024 # Fixed tile size for large images
|
152 |
+
|
153 |
+
try:
|
154 |
+
# Check for interruption
|
155 |
+
if check_stop_flag():
|
156 |
+
return None, "β Inference interrupted", None
|
157 |
+
|
158 |
+
# Convert PIL image to numpy array
|
159 |
+
img_np = np.array(image)
|
160 |
+
if len(img_np.shape) == 3:
|
161 |
+
img_np = img_np[:, :, [2, 1, 0]] # RGB to BGR
|
162 |
+
|
163 |
+
# Convert to tensor
|
164 |
+
img = torch.from_numpy(np.transpose(img_np.astype(np.float32) / 255., (2, 0, 1))).float()
|
165 |
+
img = img.unsqueeze(0).to(device)
|
166 |
+
|
167 |
+
# Check for interruption
|
168 |
+
if check_stop_flag():
|
169 |
+
return None, "β Inference interrupted", None
|
170 |
+
|
171 |
+
# Calculate target size
|
172 |
+
gt_size = [math.floor(scale * img.shape[2]), math.floor(scale * img.shape[3])]
|
173 |
+
|
174 |
+
# Determine if tile processing should be used
|
175 |
+
use_tile = should_use_tile(img.shape[2], img.shape[3])
|
176 |
+
|
177 |
+
# Force AMP mixed precision
|
178 |
+
with torch.inference_mode():
|
179 |
+
with torch.amp.autocast(device_type='cuda', dtype=torch.bfloat16):
|
180 |
+
# Check for interruption before main processing
|
181 |
+
if check_stop_flag():
|
182 |
+
return None, "β Inference interrupted", None
|
183 |
+
|
184 |
+
if use_tile:
|
185 |
+
# Use tile processing
|
186 |
+
assert tile_size % 16 == 0, f"tile_size-{tile_size} must be divisible by 16"
|
187 |
+
assert 2 * tile_overlap < tile_size, f"2 * tile_overlap must be less than tile_size"
|
188 |
+
assert 2 * crop_size <= tile_overlap, f"2 * crop_size must be less than or equal to tile_overlap"
|
189 |
+
|
190 |
+
with torch.no_grad():
|
191 |
+
output = split_and_joint_image(
|
192 |
+
lq=img,
|
193 |
+
scale_factor=scale,
|
194 |
+
split_size=tile_size,
|
195 |
+
overlap_size=tile_overlap,
|
196 |
+
model_g=encoder,
|
197 |
+
model_fea2gs=decoder,
|
198 |
+
crop_size=crop_size,
|
199 |
+
scale_modify=torch.tensor([scale, scale]),
|
200 |
+
default_step_size=1.2,
|
201 |
+
cuda_rendering=True,
|
202 |
+
mode='scale_modify',
|
203 |
+
if_dmax=True,
|
204 |
+
dmax_mode='fix',
|
205 |
+
dmax=0.1
|
206 |
+
)
|
207 |
+
else:
|
208 |
+
# Direct processing without tiles
|
209 |
+
lq_pad = preprocess(img, 16) # denominator=16 for HATL
|
210 |
+
gt_size_pad = torch.tensor([math.floor(scale * lq_pad.shape[2]),
|
211 |
+
math.floor(scale * lq_pad.shape[3])])
|
212 |
+
gt_size_pad = gt_size_pad.unsqueeze(0)
|
213 |
+
|
214 |
+
with torch.no_grad():
|
215 |
+
# Check for interruption before encoder
|
216 |
+
if check_stop_flag():
|
217 |
+
return None, "β Inference interrupted", None
|
218 |
+
|
219 |
+
# Encoder output
|
220 |
+
encoder_output = encoder(lq_pad) # b,c,h,w
|
221 |
+
|
222 |
+
# Check for interruption before decoder
|
223 |
+
if check_stop_flag():
|
224 |
+
return None, "β Inference interrupted", None
|
225 |
+
|
226 |
+
scale_vector = torch.tensor(scale, dtype=torch.float32).unsqueeze(0).to(device)
|
227 |
+
|
228 |
+
# Decoder output
|
229 |
+
batch_gs_parameters = decoder(encoder_output, scale_vector)
|
230 |
+
gs_parameters = batch_gs_parameters[0, :]
|
231 |
+
|
232 |
+
# Check for interruption before gaussian rendering
|
233 |
+
if check_stop_flag():
|
234 |
+
return None, "β Inference interrupted", None
|
235 |
+
|
236 |
+
# Gaussian rendering
|
237 |
+
b_output = generate_2D_gaussian_splatting_step(
|
238 |
+
gs_parameters=gs_parameters,
|
239 |
+
sr_size=gt_size_pad[0],
|
240 |
+
scale=scale,
|
241 |
+
sample_coords=None,
|
242 |
+
scale_modify=torch.tensor([scale, scale]),
|
243 |
+
default_step_size=1.2,
|
244 |
+
cuda_rendering=True,
|
245 |
+
mode='scale_modify',
|
246 |
+
if_dmax=True,
|
247 |
+
dmax_mode='fix',
|
248 |
+
dmax=0.1
|
249 |
+
)
|
250 |
+
output = b_output.unsqueeze(0)
|
251 |
+
|
252 |
+
# Check for interruption before post-processing
|
253 |
+
if check_stop_flag():
|
254 |
+
return None, "β Inference interrupted", None
|
255 |
+
|
256 |
+
# Post-processing
|
257 |
+
output = postprocess(output, gt_size[0], gt_size[1])
|
258 |
+
|
259 |
+
# Convert back to PIL image format
|
260 |
+
output = output.data.squeeze().float().cpu().clamp_(0, 1).numpy()
|
261 |
+
output = np.transpose(output[[2, 1, 0], :, :], (1, 2, 0)) # BGR to RGB
|
262 |
+
output = (output * 255.0).round().astype(np.uint8)
|
263 |
+
|
264 |
+
# Convert to PIL image
|
265 |
+
output_pil = Image.fromarray(output)
|
266 |
+
|
267 |
+
# Generate result information
|
268 |
+
original_size = f"{img.shape[3]}x{img.shape[2]}"
|
269 |
+
output_size = f"{output.shape[1]}x{output.shape[0]}"
|
270 |
+
tile_info = f"Tile processing enabled (size: {tile_size})" if use_tile else "Direct processing (no tiles)"
|
271 |
+
result_info = f"β
Processing completed successfully!\nOriginal size: {original_size}\nSuper-resolution size: {output_size}\nScale factor: {scale:.2f}x\nProcessing mode: {tile_info}\nAMP acceleration: Force enabled\nOverlap size: {tile_overlap}\nCrop size: {crop_size}"
|
272 |
+
|
273 |
+
return output_pil, result_info, output_pil
|
274 |
+
|
275 |
+
except Exception as e:
|
276 |
+
if check_stop_flag():
|
277 |
+
return None, "β Inference interrupted", None
|
278 |
+
return None, f"β Error during processing: {str(e)}", None
|
279 |
+
|
280 |
+
def predict(image, scale):
|
281 |
+
"""Gradio prediction function"""
|
282 |
+
output_image, info, download_image = super_resolution_inference(image, scale)
|
283 |
+
|
284 |
+
# If processing successful, save image for download
|
285 |
+
if output_image is not None:
|
286 |
+
# Create temporary filename
|
287 |
+
timestamp = int(time.time())
|
288 |
+
temp_filename = f"GSASR_SR_result_{scale}x_{timestamp}.png"
|
289 |
+
temp_path = os.path.join(tempfile.gettempdir(), temp_filename)
|
290 |
+
|
291 |
+
# Save image
|
292 |
+
output_image.save(temp_path, "PNG")
|
293 |
+
|
294 |
+
return output_image, temp_path, "β
Ready", gr.update(interactive=True)
|
295 |
+
else:
|
296 |
+
return output_image, None, info if info else "β Processing failed", gr.update(interactive=True)
|
297 |
+
|
298 |
+
# Create Gradio interface
|
299 |
+
with gr.Blocks(title="π GSASR (2D Gaussian Splatting Super-Resolution)") as demo:
|
300 |
+
gr.Markdown("# **π GSASR (Generalized and efficient 2d gaussian splatting for arbitrary-scale super-resolution)**")
|
301 |
+
gr.Markdown("Official demo for GSASR. Please refer to our [paper](https://arxiv.org/pdf/2501.06838), [project page](https://mt-cly.github.io/GSASR.github.io/), and [github](https://github.com/ChrisDud0257/GSASR) for more details.")
|
302 |
+
|
303 |
+
with gr.Row():
|
304 |
+
with gr.Column():
|
305 |
+
input_image = gr.Image(type="pil", label="Input Image")
|
306 |
+
|
307 |
+
# Scale parameters
|
308 |
+
with gr.Group():
|
309 |
+
gr.Markdown("### SR Scale")
|
310 |
+
scale_slider = gr.Slider(minimum=1.0, maximum=30.0, value=4.0, step=0.1, label="SR Scale")
|
311 |
+
|
312 |
+
# Control buttons
|
313 |
+
with gr.Row():
|
314 |
+
submit_btn = gr.Button("π Start Super-Resolution", variant="primary")
|
315 |
+
stop_btn = gr.Button("π Stop Inference", variant="stop")
|
316 |
+
|
317 |
+
with gr.Column():
|
318 |
+
output_image = gr.Image(type="pil", label="Super-Resolution Result")
|
319 |
+
|
320 |
+
# Status display
|
321 |
+
status_text = gr.Textbox(label="Status", value="β
Ready", interactive=False)
|
322 |
+
|
323 |
+
# Download component
|
324 |
+
with gr.Group():
|
325 |
+
gr.Markdown("### π₯ Download Super-Resolution Result")
|
326 |
+
download_btn = gr.File(visible=True)
|
327 |
+
|
328 |
+
# Event handlers
|
329 |
+
submit_event = submit_btn.click(
|
330 |
+
fn=predict,
|
331 |
+
inputs=[input_image, scale_slider],
|
332 |
+
outputs=[output_image, download_btn, status_text, stop_btn]
|
333 |
+
)
|
334 |
+
|
335 |
+
stop_btn.click(
|
336 |
+
fn=set_stop_flag,
|
337 |
+
inputs=[],
|
338 |
+
outputs=[status_text, stop_btn],
|
339 |
+
cancels=[submit_event]
|
340 |
+
)
|
341 |
+
|
342 |
+
# Example images
|
343 |
+
gr.Markdown("### π Example Images")
|
344 |
+
gr.Markdown("Try these examples with different scales:")
|
345 |
+
|
346 |
+
gr.Examples(
|
347 |
+
examples=[
|
348 |
+
["assets/0846x4.png", 1.5],
|
349 |
+
["assets/0892x4.png", 2.8],
|
350 |
+
["assets/0873x4_cropped_120x120.png", 30.0]
|
351 |
+
],
|
352 |
+
inputs=[input_image, scale_slider],
|
353 |
+
examples_per_page=3,
|
354 |
+
cache_examples=False,
|
355 |
+
label="Examples"
|
356 |
+
)
|
357 |
+
|
358 |
+
if __name__ == "__main__":
|
359 |
+
demo.launch(share=True, server_name="0.0.0.0")
|