Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,239 +1,128 @@
|
|
1 |
-
# app.py
|
2 |
-
import gradio as gr
|
3 |
import torch
|
4 |
-
import
|
|
|
5 |
from PIL import Image
|
6 |
-
|
7 |
-
import torch.nn.functional as F
|
8 |
from datasets import load_dataset
|
9 |
-
from
|
10 |
-
import
|
11 |
-
from tqdm import tqdm
|
12 |
-
from transformers import AutoProcessor, AutoModelForCausalLM
|
13 |
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
# Create model label
|
32 |
-
model_label = torch.zeros(len(self.model_to_idx))
|
33 |
-
model_label[self.model_to_idx[item['model_name']]] = 1
|
34 |
-
|
35 |
-
# Create prompt label (multi-hot encoding)
|
36 |
-
prompt_label = torch.zeros(len(self.token_to_idx))
|
37 |
-
for token in item['prompt'].split():
|
38 |
-
if token in self.token_to_idx:
|
39 |
-
prompt_label[self.token_to_idx[token]] = 1
|
40 |
-
|
41 |
-
return image_inputs, model_label, prompt_label
|
42 |
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
# Get Florence embeddings
|
53 |
-
outputs = self.florence(pixel_values=pixel_values, output_hidden_states=True)
|
54 |
-
features = outputs.hidden_states[-1].mean(dim=1) # Use mean pooling of last hidden state
|
55 |
-
|
56 |
-
# Generate model and prompt recommendations
|
57 |
-
model_logits = self.model_head(features)
|
58 |
-
prompt_logits = self.prompt_head(features)
|
59 |
-
|
60 |
-
return model_logits, prompt_logits
|
61 |
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
print(f"Using device: {self.device}")
|
66 |
-
|
67 |
-
# Load Florence model and processor
|
68 |
-
print("Loading Florence model and processor...")
|
69 |
-
self.processor = AutoProcessor.from_pretrained(
|
70 |
-
"microsoft/Florence-2-large",
|
71 |
-
trust_remote_code=True
|
72 |
-
)
|
73 |
-
self.florence = AutoModelForCausalLM.from_pretrained(
|
74 |
-
"microsoft/Florence-2-large",
|
75 |
-
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
76 |
-
trust_remote_code=True
|
77 |
-
).to(self.device)
|
78 |
-
|
79 |
-
# Load dataset
|
80 |
-
print("Loading dataset...")
|
81 |
-
self.dataset = load_dataset("thefcraft/civitai-stable-diffusion-337k", split="train")
|
82 |
-
self.dataset = self.dataset.select(range(min(max_samples, len(self.dataset))))
|
83 |
-
print(f"Using {len(self.dataset)} samples from dataset")
|
84 |
-
|
85 |
-
# Create vocabularies for models and tokens
|
86 |
-
self.model_to_idx = self._create_model_vocab()
|
87 |
-
self.token_to_idx = self._create_prompt_vocab()
|
88 |
-
|
89 |
-
# Initialize the recommendation model
|
90 |
-
self.model = SDRecommenderModel(
|
91 |
-
self.florence,
|
92 |
-
len(self.model_to_idx),
|
93 |
-
len(self.token_to_idx)
|
94 |
-
).to(self.device)
|
95 |
-
|
96 |
-
# Load trained weights if available
|
97 |
-
if os.path.exists("recommender_model.pth"):
|
98 |
-
self.model.load_state_dict(torch.load("recommender_model.pth", map_location=self.device))
|
99 |
-
print("Loaded trained model weights")
|
100 |
-
self.model.eval()
|
101 |
-
|
102 |
-
def _create_model_vocab(self):
|
103 |
-
print("Creating model vocabulary...")
|
104 |
-
models = set()
|
105 |
-
for item in self.dataset:
|
106 |
-
models.add(item["model_name"])
|
107 |
-
return {model: idx for idx, model in enumerate(sorted(models))}
|
108 |
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
|
|
|
|
|
|
116 |
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
self.processor,
|
124 |
-
self.model_to_idx,
|
125 |
-
self.token_to_idx
|
126 |
-
)
|
127 |
-
train_loader = DataLoader(
|
128 |
-
train_dataset,
|
129 |
-
batch_size=batch_size,
|
130 |
-
shuffle=True,
|
131 |
-
num_workers=2
|
132 |
-
)
|
133 |
-
|
134 |
-
# Setup optimizer
|
135 |
-
optimizer = torch.optim.AdamW(self.model.parameters(), lr=learning_rate)
|
136 |
-
|
137 |
-
# Training loop
|
138 |
-
self.model.train()
|
139 |
-
for epoch in range(num_epochs):
|
140 |
-
total_loss = 0
|
141 |
-
progress_bar = tqdm(train_loader, desc=f"Epoch {epoch+1}/{num_epochs}")
|
142 |
-
|
143 |
-
for batch_idx, (images, model_labels, prompt_labels) in enumerate(progress_bar):
|
144 |
-
# Move everything to device
|
145 |
-
images = {k: v.to(self.device) for k, v in images.items()}
|
146 |
-
model_labels = model_labels.to(self.device)
|
147 |
-
prompt_labels = prompt_labels.to(self.device)
|
148 |
-
|
149 |
-
# Forward pass
|
150 |
-
model_logits, prompt_logits = self.model(images)
|
151 |
-
|
152 |
-
# Calculate loss
|
153 |
-
model_loss = F.cross_entropy(model_logits, model_labels)
|
154 |
-
prompt_loss = F.binary_cross_entropy_with_logits(prompt_logits, prompt_labels)
|
155 |
-
loss = model_loss + prompt_loss
|
156 |
-
|
157 |
-
# Backward pass
|
158 |
-
optimizer.zero_grad()
|
159 |
-
loss.backward()
|
160 |
-
optimizer.step()
|
161 |
-
|
162 |
-
# Update progress
|
163 |
-
total_loss += loss.item()
|
164 |
-
progress_bar.set_postfix({"loss": total_loss / (batch_idx + 1)})
|
165 |
-
|
166 |
-
# Save trained model
|
167 |
-
torch.save(self.model.state_dict(), "recommender_model.pth")
|
168 |
-
print("Training completed and model saved")
|
169 |
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
# Get top 5 model recommendations
|
185 |
-
model_probs = F.softmax(model_logits, dim=-1)
|
186 |
-
top_models = torch.topk(model_probs, k=5)
|
187 |
-
model_recommendations = [
|
188 |
-
(list(self.model_to_idx.keys())[idx.item()], prob.item())
|
189 |
-
for prob, idx in zip(top_models.values[0], top_models.indices[0])
|
190 |
-
]
|
191 |
-
|
192 |
-
# Get top tokens for prompt recommendations
|
193 |
-
prompt_probs = F.softmax(prompt_logits, dim=-1)
|
194 |
-
top_tokens = torch.topk(prompt_probs, k=20)
|
195 |
-
recommended_tokens = [
|
196 |
-
list(self.token_to_idx.keys())[idx.item()]
|
197 |
-
for idx in top_tokens.indices[0]
|
198 |
-
]
|
199 |
-
|
200 |
-
# Create 5 prompt combinations
|
201 |
-
prompt_recommendations = [
|
202 |
-
" ".join(np.random.choice(recommended_tokens, size=8, replace=False))
|
203 |
-
for _ in range(5)
|
204 |
-
]
|
205 |
-
|
206 |
-
return (
|
207 |
-
"\n".join(f"{model} (confidence: {conf:.2f})" for model, conf in model_recommendations),
|
208 |
-
"\n".join(prompt_recommendations)
|
209 |
-
)
|
210 |
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
|
215 |
-
#
|
216 |
-
if not
|
217 |
-
|
218 |
|
219 |
-
|
220 |
-
|
221 |
-
return model_recs, prompt_recs
|
222 |
|
223 |
-
|
224 |
-
|
225 |
-
|
226 |
-
outputs=[
|
227 |
-
gr.Textbox(label="Recommended Models"),
|
228 |
-
gr.Textbox(label="Recommended Prompts")
|
229 |
-
],
|
230 |
-
title="Stable Diffusion Model & Prompt Recommender",
|
231 |
-
description="Upload an AI-generated image to get model and prompt recommendations",
|
232 |
-
)
|
233 |
|
234 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
235 |
|
236 |
# Launch the interface
|
237 |
-
|
238 |
-
interface = create_interface()
|
239 |
-
interface.launch()
|
|
|
|
|
|
|
1 |
import torch
|
2 |
+
import gradio as gr
|
3 |
+
from transformers import AutoProcessor, AutoModelForCausalLM
|
4 |
from PIL import Image
|
5 |
+
import pandas as pd
|
|
|
6 |
from datasets import load_dataset
|
7 |
+
from sklearn.metrics.pairwise import cosine_similarity
|
8 |
+
import numpy as np
|
|
|
|
|
9 |
|
10 |
+
# Load Florence-2 model and processor
|
11 |
+
model_name = "microsoft/Florence-2-base"
|
12 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
13 |
+
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
|
14 |
+
|
15 |
+
model = AutoModelForCausalLM.from_pretrained(
|
16 |
+
model_name,
|
17 |
+
torch_dtype=torch_dtype,
|
18 |
+
trust_remote_code=True
|
19 |
+
).to(device)
|
20 |
+
processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)
|
21 |
+
|
22 |
+
# Load CivitAI dataset (limited to 1000 samples)
|
23 |
+
dataset = load_dataset("thefcraft/civitai-stable-diffusion-337k", split="train[:1000]")
|
24 |
+
df = pd.DataFrame(dataset)
|
25 |
+
|
26 |
+
# Create cache for embeddings to improve performance
|
27 |
+
text_embedding_cache = {}
|
28 |
+
|
29 |
+
def get_image_embedding(image):
|
30 |
+
inputs = processor(images=image, return_tensors="pt").to(device, torch_dtype)
|
31 |
+
with torch.no_grad():
|
32 |
+
outputs = model.get_image_features(**inputs)
|
33 |
+
return outputs.cpu().numpy()
|
34 |
+
|
35 |
+
def get_text_embedding(text):
|
36 |
+
if text in text_embedding_cache:
|
37 |
+
return text_embedding_cache[text]
|
38 |
|
39 |
+
inputs = processor(text=text, return_tensors="pt").to(device, torch_dtype)
|
40 |
+
with torch.no_grad():
|
41 |
+
outputs = model.get_text_features(**inputs)
|
42 |
+
|
43 |
+
embedding = outputs.cpu().numpy()
|
44 |
+
text_embedding_cache[text] = embedding
|
45 |
+
return embedding
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
46 |
|
47 |
+
# Pre-compute text embeddings for all prompts in the dataset
|
48 |
+
def precompute_embeddings():
|
49 |
+
print("Pre-computing text embeddings...")
|
50 |
+
for idx, row in df.iterrows():
|
51 |
+
if row['prompt'] not in text_embedding_cache:
|
52 |
+
_ = get_text_embedding(row['prompt'])
|
53 |
+
if idx % 100 == 0:
|
54 |
+
print(f"Processed {idx}/1000 embeddings")
|
55 |
+
print("Finished pre-computing embeddings")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
56 |
|
57 |
+
def find_similar_images(uploaded_image, top_k=5):
|
58 |
+
# Get embedding for uploaded image
|
59 |
+
query_embedding = get_image_embedding(uploaded_image)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
|
61 |
+
# Calculate similarities with dataset
|
62 |
+
similarities = []
|
63 |
+
for idx, row in df.iterrows():
|
64 |
+
prompt_embedding = get_text_embedding(row['prompt'])
|
65 |
+
similarity = cosine_similarity(query_embedding, prompt_embedding)[0][0]
|
66 |
+
similarities.append({
|
67 |
+
'similarity': similarity,
|
68 |
+
'model': row['Model'],
|
69 |
+
'prompt': row['prompt']
|
70 |
+
})
|
71 |
|
72 |
+
# Sort by similarity and get top k results
|
73 |
+
sorted_results = sorted(similarities, key=lambda x: x['similarity'], reverse=True)
|
74 |
+
top_models = []
|
75 |
+
top_prompts = []
|
76 |
+
seen_models = set()
|
77 |
+
seen_prompts = set()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
78 |
|
79 |
+
for result in sorted_results:
|
80 |
+
if len(top_models) < top_k and result['model'] not in seen_models:
|
81 |
+
top_models.append(result['model'])
|
82 |
+
seen_models.add(result['model'])
|
83 |
|
84 |
+
if len(top_prompts) < top_k and result['prompt'] not in seen_prompts:
|
85 |
+
top_prompts.append(result['prompt'])
|
86 |
+
seen_prompts.add(result['prompt'])
|
87 |
|
88 |
+
if len(top_models) == top_k and len(top_prompts) == top_k:
|
89 |
+
break
|
90 |
+
|
91 |
+
return top_models, top_prompts
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
92 |
|
93 |
+
def process_image(input_image):
|
94 |
+
if input_image is None:
|
95 |
+
return "Please upload an image.", "Please upload an image."
|
96 |
|
97 |
+
# Convert to PIL Image if needed
|
98 |
+
if not isinstance(input_image, Image.Image):
|
99 |
+
input_image = Image.fromarray(input_image)
|
100 |
|
101 |
+
# Get recommendations
|
102 |
+
recommended_models, recommended_prompts = find_similar_images(input_image)
|
|
|
103 |
|
104 |
+
# Format output
|
105 |
+
models_text = "Recommended Models:\n" + "\n".join([f"{i+1}. {model}" for i, model in enumerate(recommended_models)])
|
106 |
+
prompts_text = "Recommended Prompts:\n" + "\n".join([f"{i+1}. {prompt}" for i, prompt in enumerate(recommended_prompts)])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
107 |
|
108 |
+
return models_text, prompts_text
|
109 |
+
|
110 |
+
# Pre-compute embeddings when starting the application
|
111 |
+
precompute_embeddings()
|
112 |
+
|
113 |
+
# Create Gradio interface
|
114 |
+
iface = gr.Interface(
|
115 |
+
fn=process_image,
|
116 |
+
inputs=gr.Image(type="pil", label="Upload AI-generated image"),
|
117 |
+
outputs=[
|
118 |
+
gr.Textbox(label="Recommended Models", lines=6),
|
119 |
+
gr.Textbox(label="Recommended Prompts", lines=6)
|
120 |
+
],
|
121 |
+
title="AI Image Model & Prompt Recommender",
|
122 |
+
description="Upload an AI-generated image to get recommendations for Stable Diffusion models and prompts.",
|
123 |
+
examples=[],
|
124 |
+
cache_examples=False
|
125 |
+
)
|
126 |
|
127 |
# Launch the interface
|
128 |
+
iface.launch()
|
|
|
|