VQGAN_CLIP / app.py
Sanshruth's picture
Update app.py
31573b4 verified
import zipfile
import time
import threading
def unzip_content():
try:
# First try using Python's zipfile
print("Attempting to unzip content using Python...")
with zipfile.ZipFile('./content.zip', 'r') as zip_ref:
zip_ref.extractall('.')
except Exception as e:
print(f"Python unzip failed: {str(e)}")
try:
# Fallback to system unzip command
print("Attempting to unzip content using system command...")
subprocess.run(['unzip', '-o', './content.zip'], check=True)
except Exception as e:
print(f"System unzip failed: {str(e)}")
raise Exception("Failed to unzip content using both methods")
print("Content successfully unzipped!")
# Try to unzip content at startup
try:
unzip_content()
except Exception as e:
print(f"Warning: Could not unzip content: {str(e)}")
import gradio as gr
import numpy as np
import torch
import torchvision
import torchvision.transforms
import torchvision.transforms.functional
import PIL
import matplotlib.pyplot as plt
import yaml
from omegaconf import OmegaConf
from CLIP import clip
import os
import sys
#os.chdir('./taming-transformers')
#from taming.models.vqgan import VQModel
#os.chdir('..')
taming_path = os.path.join(os.getcwd(), 'taming-transformers')
sys.path.append(taming_path)
from taming.models.vqgan import VQModel
from PIL import Image
import cv2
import imageio
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def create_video(images_list):
"""Create video from a list of image tensors"""
if not images_list:
print("No images provided.")
return None
# Create a unique filename in the current directory
video_path = os.path.join(os.getcwd(), f"output_{int(time.time())}.mp4")
try:
video_writer = imageio.get_writer(video_path, fps=10, codec='libx264', quality=7,
output_params=['-movflags', 'faststart'])
for img_tensor in images_list:
# Convert tensor to numpy array
img = img_tensor.cpu().numpy().transpose((1, 2, 0))
img = (img * 255).astype('uint8')
video_writer.append_data(img)
video_writer.close()
return video_path
except Exception as e:
if os.path.exists(video_path):
os.remove(video_path)
raise e
def save_from_tensors(tensor):
"""Process tensor and return the processed version"""
img = tensor.clone()
img = img.mul(255).byte()
img = img.cpu().numpy().transpose((1, 2, 0))
return img
def norm_data(data):
return (data.clip(-1, 1) + 1) / 2
def setup_clip_model():
model, _ = clip.load('ViT-B/32', jit=False)
model.eval().to(device)
return model
def setup_vqgan_model(config_path, checkpoint_path):
config = OmegaConf.load(config_path)
model = VQModel(**config.model.params)
state_dict = torch.load(checkpoint_path, map_location="cpu")["state_dict"]
model.load_state_dict(state_dict, strict=False)
return model.eval().to(device)
def generator(x, model):
x = model.post_quant_conv(x)
x = model.decoder(x)
return x
def encode_text(text, clip_model):
t = clip.tokenize(text).to(device)
return clip_model.encode_text(t).detach().clone()
def create_encoding(include, exclude, extras, clip_model):
include_enc = [encode_text(text, clip_model) for text in include]
exclude_enc = [encode_text(text, clip_model) for text in exclude]
extras_enc = [encode_text(text, clip_model) for text in extras]
return include_enc, exclude_enc, extras_enc
def create_crops(img, num_crops=32, size1=225, noise_factor=0.05):
aug_transform = torch.nn.Sequential(
torchvision.transforms.RandomHorizontalFlip(),
torchvision.transforms.RandomAffine(30, translate=(0.1, 0.1), fill=0)
).to(device)
p = size1 // 2
img = torch.nn.functional.pad(img, (p, p, p, p), mode='constant', value=0)
img = aug_transform(img)
crop_set = []
for _ in range(num_crops):
gap1 = int(torch.normal(1.2, .3, ()).clip(.43, 1.9) * size1)
offsetx = torch.randint(0, int(size1 * 2 - gap1), ())
offsety = torch.randint(0, int(size1 * 2 - gap1), ())
crop = img[:, :, offsetx:offsetx + gap1, offsety:offsety + gap1]
crop = torch.nn.functional.interpolate(crop, (224, 224), mode='bilinear', align_corners=True)
crop_set.append(crop)
img_crops = torch.cat(crop_set, 0)
randnormal = torch.randn_like(img_crops, requires_grad=False)
randstotal = torch.rand((img_crops.shape[0], 1, 1, 1)).to(device)
img_crops = img_crops + noise_factor * randstotal * randnormal
return img_crops
def optimize_result(params, prompt, vqgan_model, clip_model, w1, w2, extras_enc, exclude_enc):
alpha = 1
beta = 0.5
out = generator(params, vqgan_model)
out = norm_data(out)
out = create_crops(out)
out = torchvision.transforms.Normalize((0.48145466, 0.4578275, 0.40821073),
(0.26862954, 0.26130258, 0.27577711))(out)
img_enc = clip_model.encode_image(out)
final_enc = w1 * prompt + w2 * extras_enc[0]
final_text_include_enc = final_enc / final_enc.norm(dim=-1, keepdim=True)
final_text_exclude_enc = exclude_enc[0]
main_loss = torch.cosine_similarity(final_text_include_enc, img_enc, dim=-1)
penalize_loss = torch.cosine_similarity(final_text_exclude_enc, img_enc, dim=-1)
return -alpha * main_loss.mean() + beta * penalize_loss.mean()
def optimize(params, optimizer, prompt, vqgan_model, clip_model, w1, w2, extras_enc, exclude_enc):
loss = optimize_result(params, prompt, vqgan_model, clip_model, w1, w2, extras_enc, exclude_enc)
optimizer.zero_grad()
loss.backward()
optimizer.step()
return loss
def training_loop(params, optimizer, include_enc, exclude_enc, extras_enc, vqgan_model, clip_model, w1, w2,
total_iter=200, show_step=1):
res_img = []
res_z = []
for prompt in include_enc:
for it in range(total_iter):
loss = optimize(params, optimizer, prompt, vqgan_model, clip_model, w1, w2, extras_enc, exclude_enc)
if it >= 0 and it % show_step == 0:
with torch.no_grad():
generated = generator(params, vqgan_model)
new_img = norm_data(generated[0].to(device))
res_img.append(new_img)
res_z.append(params.clone().detach())
print(f"loss: {loss.item():.4f}\nno. of iteration: {it}")
torch.cuda.empty_cache()
return res_img, res_z
def generate_art(include_text, exclude_text, extras_text, num_iterations):
try:
# Process the input prompts
include = [x.strip() for x in include_text.split(',')]
exclude = [x.strip() for x in exclude_text.split(',')]
extras = [x.strip() for x in extras_text.split(',')]
w1, w2 = 1.0, 0.9
# Setup models
clip_model = setup_clip_model()
vqgan_model = setup_vqgan_model("./models/vqgan_imagenet_f16_16384/configs/model.yaml",
"./models/vqgan_imagenet_f16_16384/checkpoints/last.ckpt")
# Parameters
learning_rate = 0.1
batch_size = 1
wd = 0.1
size1, size2 = 225, 400
# Initialize parameters
initial_image = PIL.Image.open('./gradient1.png')
initial_image = initial_image.resize((size2, size1))
initial_image = torchvision.transforms.ToTensor()(initial_image).unsqueeze(0).to(device)
with torch.no_grad():
z, _, _ = vqgan_model.encode(initial_image)
params = torch.nn.Parameter(z).to(device)
optimizer = torch.optim.AdamW([params], lr=learning_rate, weight_decay=wd)
params.data = params.data * 0.6 + torch.randn_like(params.data) * 0.4
# Encode prompts
include_enc, exclude_enc, extras_enc = create_encoding(include, exclude, extras, clip_model)
# Run training loop
res_img, res_z = training_loop(params, optimizer, include_enc, exclude_enc, extras_enc,
vqgan_model, clip_model, w1, w2, total_iter=num_iterations)
# Create video directly from tensors
video_path = create_video(res_img)
return video_path
except Exception as e:
raise e
def gradio_interface(include_text, exclude_text, extras_text, num_iterations):
video_path = None
try:
video_path = generate_art(include_text, exclude_text, extras_text, int(num_iterations))
if not os.path.exists(video_path):
return "Video generation failed"
# Create a copy of the video path before scheduling deletion
response_path = video_path
# Schedule the video file for deletion after a delay
def cleanup():
try:
if os.path.exists(video_path):
os.remove(video_path)
except:
pass
threading.Timer(10.0, cleanup).start()
return response_path
except Exception as e:
if video_path and os.path.exists(video_path):
os.remove(video_path)
return f"An error occurred: {str(e)}"
# Try to unzip content at startup
try:
unzip_content()
except Exception as e:
print(f"Warning: Could not unzip content: {str(e)}")
# Define and launch the Gradio app
iface = gr.Interface(
fn=gradio_interface,
inputs=[
gr.Textbox(label="Include Prompts (comma-separated)",
value="desert, heavy rain, cactus"),
gr.Textbox(label="Exclude Prompts (comma-separated)",
value="confusing, blurry"),
gr.Textbox(label="Extra Style Prompts (comma-separated)",
value="desert, clear, detailed, beautiful, good shape, detailed"),
gr.Number(label="Number of Iterations",
value=200, minimum=1, maximum=1000)
],
outputs=gr.Video(label="Generated Morphing Video", format="mp4", autoplay=True),
title="VQGAN-CLIP Art Generator",
css="allow",
allow_flagging="never",
description = """
Generate artistic videos using VQGAN-CLIP. Enter your prompts separated by commas and adjust the number of iterations. The model will generate a morphing video based on your inputs.
Note: This application requires GPU access. Please either:
1. Use the Colab notebook available at https://github.com/SanshruthR/VQGAN-CLIP
2. Clone this space and enable GPU in your personal copy.
""")
if __name__ == "__main__":
print("Checking GPU availability:", "GPU AVAILABLE" if torch.cuda.is_available() else "NO GPU FOUND")
iface.launch()