Spaces:
Sleeping
Sleeping
initial_commit
Browse files
app.py
ADDED
@@ -0,0 +1,298 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import zipfile
|
2 |
+
def unzip_content():
|
3 |
+
try:
|
4 |
+
# First try using Python's zipfile
|
5 |
+
print("Attempting to unzip content using Python...")
|
6 |
+
with zipfile.ZipFile('./content.zip', 'r') as zip_ref:
|
7 |
+
zip_ref.extractall('.')
|
8 |
+
except Exception as e:
|
9 |
+
print(f"Python unzip failed: {str(e)}")
|
10 |
+
try:
|
11 |
+
# Fallback to system unzip command
|
12 |
+
print("Attempting to unzip content using system command...")
|
13 |
+
subprocess.run(['unzip', '-o', './content.zip'], check=True)
|
14 |
+
except Exception as e:
|
15 |
+
print(f"System unzip failed: {str(e)}")
|
16 |
+
raise Exception("Failed to unzip content using both methods")
|
17 |
+
print("Content successfully unzipped!")
|
18 |
+
|
19 |
+
# Try to unzip content at startup
|
20 |
+
try:
|
21 |
+
unzip_content()
|
22 |
+
except Exception as e:
|
23 |
+
print(f"Warning: Could not unzip content: {str(e)}")
|
24 |
+
|
25 |
+
import gradio as gr
|
26 |
+
import numpy as np
|
27 |
+
import torch
|
28 |
+
import torchvision
|
29 |
+
import torchvision.transforms
|
30 |
+
import torchvision.transforms.functional
|
31 |
+
import PIL
|
32 |
+
import matplotlib.pyplot as plt
|
33 |
+
import yaml
|
34 |
+
from omegaconf import OmegaConf
|
35 |
+
from CLIP import clip
|
36 |
+
import os
|
37 |
+
os.chdir('./taming-transformers')
|
38 |
+
from taming.models.vqgan import VQModel
|
39 |
+
os.chdir('..')
|
40 |
+
from PIL import Image
|
41 |
+
import cv2
|
42 |
+
import imageio
|
43 |
+
|
44 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
45 |
+
|
46 |
+
def create_video(image_folder='./generated', video_name='morphing_video.mp4'):
|
47 |
+
images = sorted([img for img in os.listdir(image_folder) if img.endswith(".png") or img.endswith(".jpg")])
|
48 |
+
if len(images) == 0:
|
49 |
+
print("No images found in the folder.")
|
50 |
+
return None
|
51 |
+
|
52 |
+
frame = cv2.imread(os.path.join(image_folder, images[0]))
|
53 |
+
height, width, layers = frame.shape
|
54 |
+
video_writer = imageio.get_writer(video_name, fps=10)
|
55 |
+
|
56 |
+
for image in images:
|
57 |
+
img_path = os.path.join(image_folder, image)
|
58 |
+
img = imageio.imread(img_path)
|
59 |
+
video_writer.append_data(img)
|
60 |
+
|
61 |
+
video_writer.close()
|
62 |
+
return video_name
|
63 |
+
|
64 |
+
def save_from_tensors(tensor, output_dir, filename):
|
65 |
+
img = tensor.clone()
|
66 |
+
img = img.mul(255).byte()
|
67 |
+
img = img.cpu().numpy().transpose((1, 2, 0))
|
68 |
+
os.makedirs(output_dir, exist_ok=True)
|
69 |
+
Image.fromarray(img).save(os.path.join(output_dir, filename))
|
70 |
+
|
71 |
+
def norm_data(data):
|
72 |
+
return (data.clip(-1, 1) + 1) / 2
|
73 |
+
|
74 |
+
def setup_clip_model():
|
75 |
+
model, _ = clip.load('ViT-B/32', jit=False)
|
76 |
+
model.eval().to(device)
|
77 |
+
return model
|
78 |
+
|
79 |
+
def setup_vqgan_model(config_path, checkpoint_path):
|
80 |
+
config = OmegaConf.load(config_path)
|
81 |
+
model = VQModel(**config.model.params)
|
82 |
+
state_dict = torch.load(checkpoint_path, map_location="cpu")["state_dict"]
|
83 |
+
model.load_state_dict(state_dict, strict=False)
|
84 |
+
return model.eval().to(device)
|
85 |
+
|
86 |
+
def generator(x, model):
|
87 |
+
x = model.post_quant_conv(x)
|
88 |
+
x = model.decoder(x)
|
89 |
+
return x
|
90 |
+
|
91 |
+
def encode_text(text, clip_model):
|
92 |
+
t = clip.tokenize(text).to(device)
|
93 |
+
return clip_model.encode_text(t).detach().clone()
|
94 |
+
|
95 |
+
def create_encoding(include, exclude, extras, clip_model):
|
96 |
+
include_enc = [encode_text(text, clip_model) for text in include]
|
97 |
+
exclude_enc = [encode_text(text, clip_model) for text in exclude]
|
98 |
+
extras_enc = [encode_text(text, clip_model) for text in extras]
|
99 |
+
return include_enc, exclude_enc, extras_enc
|
100 |
+
|
101 |
+
def create_crops(img, num_crops=32, size1=225, noise_factor=0.05):
|
102 |
+
aug_transform = torch.nn.Sequential(
|
103 |
+
torchvision.transforms.RandomHorizontalFlip(),
|
104 |
+
torchvision.transforms.RandomAffine(30, translate=(0.1, 0.1), fill=0)
|
105 |
+
).to(device)
|
106 |
+
|
107 |
+
p = size1 // 2
|
108 |
+
img = torch.nn.functional.pad(img, (p, p, p, p), mode='constant', value=0)
|
109 |
+
img = aug_transform(img)
|
110 |
+
|
111 |
+
crop_set = []
|
112 |
+
for _ in range(num_crops):
|
113 |
+
gap1 = int(torch.normal(1.2, .3, ()).clip(.43, 1.9) * size1)
|
114 |
+
offsetx = torch.randint(0, int(size1 * 2 - gap1), ())
|
115 |
+
offsety = torch.randint(0, int(size1 * 2 - gap1), ())
|
116 |
+
crop = img[:, :, offsetx:offsetx + gap1, offsety:offsety + gap1]
|
117 |
+
crop = torch.nn.functional.interpolate(crop, (224, 224), mode='bilinear', align_corners=True)
|
118 |
+
crop_set.append(crop)
|
119 |
+
|
120 |
+
img_crops = torch.cat(crop_set, 0)
|
121 |
+
randnormal = torch.randn_like(img_crops, requires_grad=False)
|
122 |
+
randstotal = torch.rand((img_crops.shape[0], 1, 1, 1)).to(device)
|
123 |
+
img_crops = img_crops + noise_factor * randstotal * randnormal
|
124 |
+
|
125 |
+
return img_crops
|
126 |
+
|
127 |
+
def optimize_result(params, prompt, vqgan_model, clip_model, w1, w2, extras_enc, exclude_enc):
|
128 |
+
alpha = 1
|
129 |
+
beta = 0.5
|
130 |
+
out = generator(params, vqgan_model)
|
131 |
+
out = norm_data(out)
|
132 |
+
out = create_crops(out)
|
133 |
+
out = torchvision.transforms.Normalize((0.48145466, 0.4578275, 0.40821073),
|
134 |
+
(0.26862954, 0.26130258, 0.27577711))(out)
|
135 |
+
|
136 |
+
img_enc = clip_model.encode_image(out)
|
137 |
+
final_enc = w1 * prompt + w2 * extras_enc[0]
|
138 |
+
final_text_include_enc = final_enc / final_enc.norm(dim=-1, keepdim=True)
|
139 |
+
final_text_exclude_enc = exclude_enc[0]
|
140 |
+
|
141 |
+
main_loss = torch.cosine_similarity(final_text_include_enc, img_enc, dim=-1)
|
142 |
+
penalize_loss = torch.cosine_similarity(final_text_exclude_enc, img_enc, dim=-1)
|
143 |
+
|
144 |
+
return -alpha * main_loss.mean() + beta * penalize_loss.mean()
|
145 |
+
|
146 |
+
def optimize(params, optimizer, prompt, vqgan_model, clip_model, w1, w2, extras_enc, exclude_enc):
|
147 |
+
loss = optimize_result(params, prompt, vqgan_model, clip_model, w1, w2, extras_enc, exclude_enc)
|
148 |
+
optimizer.zero_grad()
|
149 |
+
loss.backward()
|
150 |
+
optimizer.step()
|
151 |
+
return loss
|
152 |
+
|
153 |
+
def training_loop(params, optimizer, include_enc, exclude_enc, extras_enc, vqgan_model, clip_model, w1, w2,
|
154 |
+
total_iter=200, show_step=1):
|
155 |
+
res_img = []
|
156 |
+
res_z = []
|
157 |
+
|
158 |
+
for prompt in include_enc:
|
159 |
+
for it in range(total_iter):
|
160 |
+
loss = optimize(params, optimizer, prompt, vqgan_model, clip_model, w1, w2, extras_enc, exclude_enc)
|
161 |
+
|
162 |
+
if it >= 0 and it % show_step == 0:
|
163 |
+
with torch.no_grad():
|
164 |
+
generated = generator(params, vqgan_model)
|
165 |
+
new_img = norm_data(generated[0].to(device))
|
166 |
+
res_img.append(new_img)
|
167 |
+
res_z.append(params.clone().detach())
|
168 |
+
print(f"loss: {loss.item():.4f}\nno. of iteration: {it}")
|
169 |
+
|
170 |
+
torch.cuda.empty_cache()
|
171 |
+
return res_img, res_z
|
172 |
+
|
173 |
+
def generate_art(include_text, exclude_text, extras_text, num_iterations):
|
174 |
+
try:
|
175 |
+
# Process the input prompts
|
176 |
+
include = [x.strip() for x in include_text.split(',')]
|
177 |
+
exclude = [x.strip() for x in exclude_text.split(',')]
|
178 |
+
extras = [x.strip() for x in extras_text.split(',')]
|
179 |
+
|
180 |
+
w1, w2 = 1.0, 0.9
|
181 |
+
|
182 |
+
# Setup models
|
183 |
+
clip_model = setup_clip_model()
|
184 |
+
vqgan_model = setup_vqgan_model("./models/vqgan_imagenet_f16_16384/configs/model.yaml",
|
185 |
+
"./models/vqgan_imagenet_f16_16384/checkpoints/last.ckpt")
|
186 |
+
|
187 |
+
# Parameters
|
188 |
+
learning_rate = 0.1
|
189 |
+
batch_size = 1
|
190 |
+
wd = 0.1
|
191 |
+
size1, size2 = 225, 400
|
192 |
+
|
193 |
+
# Initialize parameters
|
194 |
+
initial_image = PIL.Image.open('./gradient1.png')
|
195 |
+
initial_image = initial_image.resize((size2, size1))
|
196 |
+
initial_image = torchvision.transforms.ToTensor()(initial_image).unsqueeze(0).to(device)
|
197 |
+
|
198 |
+
with torch.no_grad():
|
199 |
+
z, _, _ = vqgan_model.encode(initial_image)
|
200 |
+
|
201 |
+
params = torch.nn.Parameter(z).to(device)
|
202 |
+
optimizer = torch.optim.AdamW([params], lr=learning_rate, weight_decay=wd)
|
203 |
+
params.data = params.data * 0.6 + torch.randn_like(params.data) * 0.4
|
204 |
+
|
205 |
+
# Encode prompts
|
206 |
+
include_enc, exclude_enc, extras_enc = create_encoding(include, exclude, extras, clip_model)
|
207 |
+
|
208 |
+
# Run training loop
|
209 |
+
res_img, res_z = training_loop(params, optimizer, include_enc, exclude_enc, extras_enc,
|
210 |
+
vqgan_model, clip_model, w1, w2, total_iter=num_iterations)
|
211 |
+
|
212 |
+
# Save results
|
213 |
+
output_dir = "generated"
|
214 |
+
# Create output directory if it doesn't exist
|
215 |
+
os.makedirs(output_dir, exist_ok=True)
|
216 |
+
|
217 |
+
# Clear any existing files in the output directory
|
218 |
+
for file in os.listdir(output_dir):
|
219 |
+
file_path = os.path.join(output_dir, file)
|
220 |
+
if os.path.isfile(file_path):
|
221 |
+
os.remove(file_path)
|
222 |
+
|
223 |
+
for i, img in enumerate(res_img):
|
224 |
+
save_from_tensors(img, output_dir, f"generated_image_{i:03d}.png")
|
225 |
+
|
226 |
+
# Create video
|
227 |
+
video_path = create_video()
|
228 |
+
|
229 |
+
# Delete the generated folder and its contents after creating the video
|
230 |
+
import shutil
|
231 |
+
shutil.rmtree(output_dir)
|
232 |
+
|
233 |
+
return video_path
|
234 |
+
|
235 |
+
except Exception as e:
|
236 |
+
# If there's an error, ensure the generated folder is cleaned up
|
237 |
+
if os.path.exists("generated"):
|
238 |
+
import shutil
|
239 |
+
shutil.rmtree("generated")
|
240 |
+
raise e # Re-raise the exception to be handled by the calling function
|
241 |
+
def gradio_interface(include_text, exclude_text, extras_text, num_iterations):
|
242 |
+
try:
|
243 |
+
video_path = generate_art(include_text, exclude_text, extras_text, int(num_iterations))
|
244 |
+
return video_path
|
245 |
+
except Exception as e:
|
246 |
+
return f"An error occurred: {str(e)}"
|
247 |
+
|
248 |
+
# Define and launch the Gradio app
|
249 |
+
iface = gr.Interface(
|
250 |
+
fn=gradio_interface,
|
251 |
+
inputs=[
|
252 |
+
gr.Textbox(label="Include Prompts (comma-separated)",
|
253 |
+
value="desert, heavy rain, cactus"),
|
254 |
+
gr.Textbox(label="Exclude Prompts (comma-separated)",
|
255 |
+
value="confusing, blurry"),
|
256 |
+
gr.Textbox(label="Extra Style Prompts (comma-separated)",
|
257 |
+
value="desert, clear, detailed, beautiful, good shape, detailed"),
|
258 |
+
gr.Number(label="Number of Iterations",
|
259 |
+
value=200, minimum=1, maximum=1000)
|
260 |
+
],
|
261 |
+
outputs=gr.Video(label="Generated Morphing Video"),
|
262 |
+
title="VQGAN-CLIP Art Generator",
|
263 |
+
description="""
|
264 |
+
[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1ivRYvTaX90PRghQIqAdOyEawkY0YLefa?authuser=0#scrollTo=WE7aPQ0t1hd2)
|
265 |
+
[![Clone Space](https://huggingface.co/datasets/huggingface/badges/raw/main/clone-space-lg.svg)](https://huggingface.co/spaces/your-username/your-space-name?duplicate=true)
|
266 |
+
|
267 |
+
Generate artistic videos using VQGAN-CLIP.
|
268 |
+
Enter your prompts separated by commas and adjust the number of iterations.
|
269 |
+
The model will generate a morphing video based on your inputs.
|
270 |
+
|
271 |
+
**Note:** This application requires GPU access. Please either:
|
272 |
+
1. Use the Colab notebook (click the Colab badge above) with GPU runtime
|
273 |
+
2. Clone this space (click Clone Space badge) and enable GPU in your personal copy""",
|
274 |
+
css="""
|
275 |
+
.gradio-container {
|
276 |
+
font-family: 'IBM Plex Sans', sans-serif;
|
277 |
+
}
|
278 |
+
.gr-button {
|
279 |
+
color: white;
|
280 |
+
border-radius: 7px;
|
281 |
+
background: linear-gradient(45deg, #7747FF, #FF3557);
|
282 |
+
border: none;
|
283 |
+
height: 46px;
|
284 |
+
}
|
285 |
+
a {
|
286 |
+
text-decoration: none;
|
287 |
+
}
|
288 |
+
.maintenance-msg {
|
289 |
+
color: #FF0000;
|
290 |
+
font-size: 14px;
|
291 |
+
margin-top: 10px;
|
292 |
+
}
|
293 |
+
"""
|
294 |
+
)
|
295 |
+
|
296 |
+
if __name__ == "__main__":
|
297 |
+
print("Checking GPU availability:", "GPU AVAILABLE" if torch.cuda.is_available() else "NO GPU FOUND")
|
298 |
+
iface.launch()
|