Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,159 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from torch.utils.data import DataLoader, Dataset
|
3 |
+
from torchvision import transforms
|
4 |
+
from PIL import Image
|
5 |
+
from diffusers import StableDiffusionPipeline
|
6 |
+
from transformers import CLIPTokenizer
|
7 |
+
import os
|
8 |
+
import zipfile
|
9 |
+
import gradio as gr
|
10 |
+
|
11 |
+
# Define the device
|
12 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
13 |
+
|
14 |
+
# Define your custom dataset
|
15 |
+
class CustomImageDataset(Dataset):
|
16 |
+
def __init__(self, images, prompts, transform=None):
|
17 |
+
self.images = images
|
18 |
+
self.prompts = prompts
|
19 |
+
self.transform = transform
|
20 |
+
|
21 |
+
def __len__(self):
|
22 |
+
return len(self.images)
|
23 |
+
|
24 |
+
def __getitem__(self, idx):
|
25 |
+
image = self.images[idx]
|
26 |
+
if self.transform:
|
27 |
+
image = self.transform(image)
|
28 |
+
prompt = self.prompts[idx]
|
29 |
+
return image, prompt
|
30 |
+
|
31 |
+
# Function to fine-tune the model
|
32 |
+
def fine_tune_model(images, prompts, model_save_path, num_epochs=3):
|
33 |
+
transform = transforms.Compose([
|
34 |
+
transforms.Resize((512, 512)),
|
35 |
+
transforms.ToTensor(),
|
36 |
+
transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]),
|
37 |
+
])
|
38 |
+
dataset = CustomImageDataset(images, prompts, transform)
|
39 |
+
dataloader = DataLoader(dataset, batch_size=4, shuffle=True)
|
40 |
+
|
41 |
+
# Load Stable Diffusion model
|
42 |
+
pipeline = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2").to(device)
|
43 |
+
|
44 |
+
# Load model components
|
45 |
+
vae = pipeline.vae.to(device)
|
46 |
+
unet = pipeline.unet.to(device)
|
47 |
+
text_encoder = pipeline.text_encoder.to(device)
|
48 |
+
tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32") # Ensure correct tokenizer is used
|
49 |
+
optimizer = torch.optim.AdamW(unet.parameters(), lr=5e-6) # Define the optimizer
|
50 |
+
|
51 |
+
# Define timestep range for training
|
52 |
+
timesteps = torch.linspace(0, 1, steps=5).to(device)
|
53 |
+
|
54 |
+
# Fine-tuning loop
|
55 |
+
for epoch in range(num_epochs):
|
56 |
+
for i, (images, prompts) in enumerate(dataloader):
|
57 |
+
images = images.to(device) # Move images to GPU if available
|
58 |
+
|
59 |
+
# Tokenize the prompts
|
60 |
+
inputs = tokenizer(list(prompts), padding=True, return_tensors="pt", truncation=True).to(device)
|
61 |
+
|
62 |
+
latents = vae.encode(images).latent_dist.sample() * 0.18215
|
63 |
+
text_embeddings = text_encoder(inputs.input_ids).last_hidden_state
|
64 |
+
|
65 |
+
noise = torch.randn_like(latents).to(device)
|
66 |
+
noisy_latents = latents + noise
|
67 |
+
|
68 |
+
# Pass text embeddings and timestep to UNet
|
69 |
+
timestep = torch.randint(0, len(timesteps), (latents.size(0),), device=device).float()
|
70 |
+
pred_noise = unet(noisy_latents, timestep=timestep, encoder_hidden_states=text_embeddings).sample
|
71 |
+
|
72 |
+
loss = torch.nn.functional.mse_loss(pred_noise, noise)
|
73 |
+
optimizer.zero_grad()
|
74 |
+
loss.backward()
|
75 |
+
optimizer.step()
|
76 |
+
|
77 |
+
# Save the fine-tuned model
|
78 |
+
pipeline.save_pretrained(model_save_path)
|
79 |
+
|
80 |
+
# Function to convert tensor to PIL Image
|
81 |
+
def tensor_to_pil(tensor):
|
82 |
+
tensor = tensor.squeeze().cpu().clamp(0, 1) # Remove batch dimension if necessary
|
83 |
+
tensor = transforms.ToPILImage()(tensor)
|
84 |
+
return tensor
|
85 |
+
|
86 |
+
# Function to generate images
|
87 |
+
def generate_images(pipeline, prompt):
|
88 |
+
with torch.no_grad():
|
89 |
+
# Generate image from the prompt
|
90 |
+
output = pipeline(prompt)
|
91 |
+
|
92 |
+
# Convert the output to PIL Image
|
93 |
+
image = output.images[0] # Get the first generated image
|
94 |
+
return image
|
95 |
+
|
96 |
+
# Function to zip the fine-tuned model
|
97 |
+
def zip_model(model_path):
|
98 |
+
zip_path = f"{model_path}.zip"
|
99 |
+
with zipfile.ZipFile(zip_path, "w") as zipf:
|
100 |
+
for root, _, files in os.walk(model_path):
|
101 |
+
for file in files:
|
102 |
+
zipf.write(os.path.join(root, file), os.path.relpath(os.path.join(root, file), model_path))
|
103 |
+
return zip_path
|
104 |
+
|
105 |
+
# Gradio interface functions
|
106 |
+
def start_fine_tuning(uploaded_files, prompts, num_epochs):
|
107 |
+
images = [Image.open(file).convert("RGB") for file in uploaded_files]
|
108 |
+
model_save_path = "fine_tuned_model"
|
109 |
+
fine_tune_model(images, prompts, model_save_path, num_epochs=int(num_epochs))
|
110 |
+
return "Fine-tuning completed! Model is ready for download."
|
111 |
+
|
112 |
+
def download_model():
|
113 |
+
model_save_path = "fine_tuned_model"
|
114 |
+
if os.path.exists(model_save_path):
|
115 |
+
return zip_model(model_save_path)
|
116 |
+
else:
|
117 |
+
return None
|
118 |
+
|
119 |
+
def generate_new_image(prompt):
|
120 |
+
model_save_path = "fine_tuned_model"
|
121 |
+
if os.path.exists(model_save_path):
|
122 |
+
pipeline = StableDiffusionPipeline.from_pretrained(model_save_path).to(device)
|
123 |
+
else:
|
124 |
+
pipeline = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2").to(device)
|
125 |
+
image = generate_images(pipeline, prompt)
|
126 |
+
image_path = "generated_image.png"
|
127 |
+
image.save(image_path)
|
128 |
+
return image_path
|
129 |
+
|
130 |
+
# Gradio interface
|
131 |
+
with gr.Blocks() as demo:
|
132 |
+
gr.Markdown("# Fine-Tune Stable Diffusion and Generate Images")
|
133 |
+
|
134 |
+
with gr.Tab("Fine-Tune Model"):
|
135 |
+
with gr.Row():
|
136 |
+
uploaded_files = gr.File(label="Upload Images", file_types=[".png", ".jpg", ".jpeg"], file_count="multiple")
|
137 |
+
with gr.Row():
|
138 |
+
prompts = gr.Textbox(label="Enter Prompts (comma-separated)")
|
139 |
+
num_epochs = gr.Number(label="Number of Epochs", value=3)
|
140 |
+
with gr.Row():
|
141 |
+
fine_tune_button = gr.Button("Start Fine-Tuning")
|
142 |
+
fine_tune_output = gr.Textbox(label="Output")
|
143 |
+
|
144 |
+
fine_tune_button.click(start_fine_tuning, [uploaded_files, prompts, num_epochs], fine_tune_output)
|
145 |
+
|
146 |
+
with gr.Tab("Download Fine-Tuned Model"):
|
147 |
+
download_button = gr.Button("Download Fine-Tuned Model")
|
148 |
+
download_output = gr.File()
|
149 |
+
|
150 |
+
download_button.click(download_model, [], download_output)
|
151 |
+
|
152 |
+
with gr.Tab("Generate New Images"):
|
153 |
+
prompt_input = gr.Textbox(label="Enter a Prompt")
|
154 |
+
generate_button = gr.Button("Generate Image")
|
155 |
+
generated_image = gr.Image(label="Generated Image")
|
156 |
+
|
157 |
+
generate_button.click(generate_new_image, [prompt_input], generated_image)
|
158 |
+
|
159 |
+
demo.launch()
|