Update app.py
Browse files
app.py
CHANGED
@@ -16,17 +16,17 @@ import spaces
|
|
16 |
from typing import List, Dict, Tuple, Optional, Union
|
17 |
|
18 |
# Load model/processor
|
19 |
-
model = CLIPModel.from_pretrained("openai/clip-vit-large-patch14")
|
20 |
-
processor = CLIPProcessor.from_pretrained("openai/clip-vit-large-patch14")
|
21 |
model.eval()
|
22 |
|
23 |
-
DATASET_DIR = Path("dataset")
|
24 |
-
CACHE_FILE = "cache.pkl"
|
25 |
|
26 |
# Define supported image formats
|
27 |
-
IMAGE_EXTENSIONS = ["*.jpg", "*.jpeg", "*.png", "*.bmp", "*.gif", "*.webp", "*.tiff", "*.tif"]
|
28 |
|
29 |
-
def get_all_image_files()
|
30 |
"""
|
31 |
Get all image files from the dataset directory.
|
32 |
|
@@ -35,7 +35,7 @@ def get_all_image_files() -> List[Path]:
|
|
35 |
Returns:
|
36 |
List[Path]: List of Path objects for all found image files
|
37 |
"""
|
38 |
-
image_files = []
|
39 |
for ext in IMAGE_EXTENSIONS:
|
40 |
image_files.extend(DATASET_DIR.glob(ext))
|
41 |
image_files.extend(DATASET_DIR.glob(ext.upper())) # Also check uppercase
|
@@ -59,7 +59,7 @@ def get_embedding(image: Image.Image, device: str = "cpu") -> torch.Tensor:
|
|
59 |
inputs = processor(images=image, return_tensors="pt").to(device)
|
60 |
model_device = model.to(device)
|
61 |
with torch.no_grad():
|
62 |
-
emb = model_device.get_image_features(**inputs)
|
63 |
# L2 normalize the embeddings
|
64 |
emb = emb / emb.norm(p=2, dim=-1, keepdim=True)
|
65 |
return emb
|
@@ -80,29 +80,29 @@ def get_reference_embeddings() -> Dict[str, torch.Tensor]:
|
|
80 |
PermissionError: If unable to write cache file
|
81 |
"""
|
82 |
# Get all current image files
|
83 |
-
current_image_files = get_all_image_files()
|
84 |
-
current_images = set(img_path.name for img_path in current_image_files)
|
85 |
|
86 |
# Load existing cache if it exists
|
87 |
-
cached_embeddings = {}
|
88 |
if os.path.exists(CACHE_FILE):
|
89 |
with open(CACHE_FILE, "rb") as f:
|
90 |
cached_embeddings = pickle.load(f)
|
91 |
|
92 |
# Check if cache is up to date
|
93 |
-
cached_images = set(cached_embeddings.keys())
|
94 |
|
95 |
# If cache is missing images or has extra images, rebuild
|
96 |
if current_images != cached_images:
|
97 |
print(f"Cache outdated. Current: {len(current_images)}, Cached: {len(cached_images)}")
|
98 |
-
embeddings = {}
|
99 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
100 |
|
101 |
for img_path in current_image_files:
|
102 |
print(f"Processing {img_path.name}...")
|
103 |
try:
|
104 |
-
img = Image.open(img_path).convert("RGB")
|
105 |
-
emb = get_embedding(img, device=device)
|
106 |
embeddings[img_path.name] = emb.cpu()
|
107 |
except Exception as e:
|
108 |
print(f"Error processing {img_path.name}: {e}")
|
@@ -117,7 +117,8 @@ def get_reference_embeddings() -> Dict[str, torch.Tensor]:
|
|
117 |
print(f"Using cached embeddings for {len(cached_embeddings)} images")
|
118 |
return cached_embeddings
|
119 |
|
120 |
-
|
|
|
121 |
|
122 |
@spaces.GPU
|
123 |
def search_similar(query_img: Image.Image) -> List[Tuple[str, str]]:
|
@@ -138,21 +139,21 @@ def search_similar(query_img: Image.Image) -> List[Tuple[str, str]]:
|
|
138 |
global reference_embeddings
|
139 |
reference_embeddings = get_reference_embeddings()
|
140 |
|
141 |
-
query_emb = get_embedding(query_img, device="cuda")
|
142 |
-
results = []
|
143 |
|
144 |
for name, ref_emb in reference_embeddings.items():
|
145 |
# Move reference embedding to same device as query
|
146 |
-
ref_emb_gpu = ref_emb.to("cuda")
|
147 |
# Compute cosine similarity
|
148 |
-
sim = torch.nn.functional.cosine_similarity(query_emb, ref_emb_gpu, dim=1).item()
|
149 |
results.append((name, sim))
|
150 |
|
151 |
results.sort(key=lambda x: x[1], reverse=True)
|
152 |
|
153 |
# Filter out low similarity results (adjust threshold as needed)
|
154 |
-
SIMILARITY_THRESHOLD = 0.2 # Only show results above 20% similarity
|
155 |
-
filtered_results = [(name, score) for name, score in results if score > SIMILARITY_THRESHOLD]
|
156 |
|
157 |
if not filtered_results:
|
158 |
return [("No similar images found", "No matches above similarity threshold")]
|
@@ -181,12 +182,12 @@ def add_image(name: str, image: Image.Image) -> str:
|
|
181 |
return "Please provide a valid image name."
|
182 |
|
183 |
# Save as PNG to preserve quality for all input formats
|
184 |
-
path = DATASET_DIR / f"{name}.png"
|
185 |
image.save(path, "PNG")
|
186 |
|
187 |
# Use GPU for consistency if available
|
188 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
189 |
-
emb = get_embedding(image, device=device)
|
190 |
|
191 |
# Add to current embeddings and save cache
|
192 |
reference_embeddings[f"{name}.png"] = emb.cpu()
|
@@ -196,15 +197,37 @@ def add_image(name: str, image: Image.Image) -> str:
|
|
196 |
|
197 |
return f"Image '{name}' added to dataset. Total images: {len(reference_embeddings)}"
|
198 |
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
|
|
|
|
|
|
|
|
|
|
|
203 |
|
204 |
-
add_interface = gr.Interface(
|
205 |
-
|
206 |
-
|
207 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
208 |
|
209 |
-
|
210 |
-
demo.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
from typing import List, Dict, Tuple, Optional, Union
|
17 |
|
18 |
# Load model/processor
|
19 |
+
model: CLIPModel = CLIPModel.from_pretrained("openai/clip-vit-large-patch14")
|
20 |
+
processor: CLIPProcessor = CLIPProcessor.from_pretrained("openai/clip-vit-large-patch14")
|
21 |
model.eval()
|
22 |
|
23 |
+
DATASET_DIR: Path = Path("dataset")
|
24 |
+
CACHE_FILE: str = "cache.pkl"
|
25 |
|
26 |
# Define supported image formats
|
27 |
+
IMAGE_EXTENSIONS: List[str] = ["*.jpg", "*.jpeg", "*.png", "*.bmp", "*.gif", "*.webp", "*.tiff", "*.tif"]
|
28 |
|
29 |
+
def get_all_image_files() -> List[Path]:
|
30 |
"""
|
31 |
Get all image files from the dataset directory.
|
32 |
|
|
|
35 |
Returns:
|
36 |
List[Path]: List of Path objects for all found image files
|
37 |
"""
|
38 |
+
image_files: List[Path] = []
|
39 |
for ext in IMAGE_EXTENSIONS:
|
40 |
image_files.extend(DATASET_DIR.glob(ext))
|
41 |
image_files.extend(DATASET_DIR.glob(ext.upper())) # Also check uppercase
|
|
|
59 |
inputs = processor(images=image, return_tensors="pt").to(device)
|
60 |
model_device = model.to(device)
|
61 |
with torch.no_grad():
|
62 |
+
emb: torch.Tensor = model_device.get_image_features(**inputs)
|
63 |
# L2 normalize the embeddings
|
64 |
emb = emb / emb.norm(p=2, dim=-1, keepdim=True)
|
65 |
return emb
|
|
|
80 |
PermissionError: If unable to write cache file
|
81 |
"""
|
82 |
# Get all current image files
|
83 |
+
current_image_files: List[Path] = get_all_image_files()
|
84 |
+
current_images: set = set(img_path.name for img_path in current_image_files)
|
85 |
|
86 |
# Load existing cache if it exists
|
87 |
+
cached_embeddings: Dict[str, torch.Tensor] = {}
|
88 |
if os.path.exists(CACHE_FILE):
|
89 |
with open(CACHE_FILE, "rb") as f:
|
90 |
cached_embeddings = pickle.load(f)
|
91 |
|
92 |
# Check if cache is up to date
|
93 |
+
cached_images: set = set(cached_embeddings.keys())
|
94 |
|
95 |
# If cache is missing images or has extra images, rebuild
|
96 |
if current_images != cached_images:
|
97 |
print(f"Cache outdated. Current: {len(current_images)}, Cached: {len(cached_images)}")
|
98 |
+
embeddings: Dict[str, torch.Tensor] = {}
|
99 |
+
device: str = "cuda" if torch.cuda.is_available() else "cpu"
|
100 |
|
101 |
for img_path in current_image_files:
|
102 |
print(f"Processing {img_path.name}...")
|
103 |
try:
|
104 |
+
img: Image.Image = Image.open(img_path).convert("RGB")
|
105 |
+
emb: torch.Tensor = get_embedding(img, device=device)
|
106 |
embeddings[img_path.name] = emb.cpu()
|
107 |
except Exception as e:
|
108 |
print(f"Error processing {img_path.name}: {e}")
|
|
|
117 |
print(f"Using cached embeddings for {len(cached_embeddings)} images")
|
118 |
return cached_embeddings
|
119 |
|
120 |
+
# Initialize reference embeddings
|
121 |
+
reference_embeddings: Dict[str, torch.Tensor] = get_reference_embeddings()
|
122 |
|
123 |
@spaces.GPU
|
124 |
def search_similar(query_img: Image.Image) -> List[Tuple[str, str]]:
|
|
|
139 |
global reference_embeddings
|
140 |
reference_embeddings = get_reference_embeddings()
|
141 |
|
142 |
+
query_emb: torch.Tensor = get_embedding(query_img, device="cuda")
|
143 |
+
results: List[Tuple[str, float]] = []
|
144 |
|
145 |
for name, ref_emb in reference_embeddings.items():
|
146 |
# Move reference embedding to same device as query
|
147 |
+
ref_emb_gpu: torch.Tensor = ref_emb.to("cuda")
|
148 |
# Compute cosine similarity
|
149 |
+
sim: float = torch.nn.functional.cosine_similarity(query_emb, ref_emb_gpu, dim=1).item()
|
150 |
results.append((name, sim))
|
151 |
|
152 |
results.sort(key=lambda x: x[1], reverse=True)
|
153 |
|
154 |
# Filter out low similarity results (adjust threshold as needed)
|
155 |
+
SIMILARITY_THRESHOLD: float = 0.2 # Only show results above 20% similarity
|
156 |
+
filtered_results: List[Tuple[str, float]] = [(name, score) for name, score in results if score > SIMILARITY_THRESHOLD]
|
157 |
|
158 |
if not filtered_results:
|
159 |
return [("No similar images found", "No matches above similarity threshold")]
|
|
|
182 |
return "Please provide a valid image name."
|
183 |
|
184 |
# Save as PNG to preserve quality for all input formats
|
185 |
+
path: Path = DATASET_DIR / f"{name}.png"
|
186 |
image.save(path, "PNG")
|
187 |
|
188 |
# Use GPU for consistency if available
|
189 |
+
device: str = "cuda" if torch.cuda.is_available() else "cpu"
|
190 |
+
emb: torch.Tensor = get_embedding(image, device=device)
|
191 |
|
192 |
# Add to current embeddings and save cache
|
193 |
reference_embeddings[f"{name}.png"] = emb.cpu()
|
|
|
197 |
|
198 |
return f"Image '{name}' added to dataset. Total images: {len(reference_embeddings)}"
|
199 |
|
200 |
+
# Create Gradio interfaces
|
201 |
+
search_interface: gr.Interface = gr.Interface(
|
202 |
+
fn=search_similar,
|
203 |
+
inputs=gr.Image(type="pil", label="Query Image"),
|
204 |
+
outputs=gr.Gallery(label="Top Matches", columns=5),
|
205 |
+
allow_flagging="never",
|
206 |
+
title="Image Similarity Search",
|
207 |
+
description="Upload an image to find similar images in the dataset"
|
208 |
+
)
|
209 |
|
210 |
+
add_interface: gr.Interface = gr.Interface(
|
211 |
+
fn=add_image,
|
212 |
+
inputs=[
|
213 |
+
gr.Text(label="Image Name", placeholder="Enter a unique name for your image"),
|
214 |
+
gr.Image(type="pil", label="Product Image")
|
215 |
+
],
|
216 |
+
outputs="text",
|
217 |
+
allow_flagging="never",
|
218 |
+
title="Add Image to Dataset",
|
219 |
+
description="Add a new image to the searchable dataset"
|
220 |
+
)
|
221 |
|
222 |
+
# Create main application
|
223 |
+
demo: gr.TabbedInterface = gr.TabbedInterface(
|
224 |
+
[search_interface, add_interface],
|
225 |
+
tab_names=["Search", "Add Product"],
|
226 |
+
title="CLIP Image Search System",
|
227 |
+
theme=gr.themes.Soft()
|
228 |
+
)
|
229 |
+
|
230 |
+
if __name__ == "__main__":
|
231 |
+
# Ensure dataset directory exists
|
232 |
+
DATASET_DIR.mkdir(exist_ok=True)
|
233 |
+
demo.launch(share=True)
|