Create app.py

app.py

@@ -0,0 +1,95 @@
+# Import core libraries
+import numpy as np
+import pandas as pd
+import torch
+import gradio as gr
+
+# Import CLIP model and processor
+from transformers import CLIPModel, CLIPProcessor
+
+# Import image handling
+from PIL import Image
+
+
+# Select device (GPU if available)
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+
+# Load pretrained CLIP model
+model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+# Move model to device and set evaluation mode
+model = model.to(device)
+model.eval()
+
+
+# Load precomputed embeddings from file
+emb_df = pd.read_parquet("clip_embeddings_3000.parquet")
+
+# Extract image identifiers
+sampled_ids = emb_df["image_id"].values
+
+# Extract normalized embeddings matrix
+embeddings = emb_df.drop(columns=["image_id"]).values.astype("float32")
+
+
+# Attempt to load sampled indices (optional optimization)
+try:
+    sampled_indices = np.load("sampled_indices_3000.npy")
+except Exception:
+    sampled_indices = None
+
+
+# Convert a user image into a normalized CLIP embedding
+def embed_image(image: Image.Image):
+    # Preprocess image for CLIP
+    inputs = processor(images=image, return_tensors="pt")
+    inputs = {k: v.to(device) for k, v in inputs.items()}
+
+    # Extract image features without gradients
+    with torch.no_grad():
+        features = model.get_image_features(**inputs)
+
+    # Convert embedding to numpy and normalize
+    vec = features.cpu().numpy()[0]
+    vec = vec / np.linalg.norm(vec)
+
+    return vec
+
+
+# Recommend top-3 visually similar images
+def recommend(image):
+    # Embed user input image
+    user_vec = embed_image(image)
+
+    # Compute cosine similarity scores
+    scores = embeddings @ user_vec
+
+    # Retrieve top-4 results (including query itself)
+    top_idx = np.argsort(scores)[::-1][:4]
+
+    # Remove the first result (query image)
+    top_idx = top_idx[1:]
+
+    # Load recommended images
+    results = []
+    for idx in top_idx:
+        img = Image.open(sampled_ids[idx]).convert("RGB")
+        results.append(img)
+
+    return results
+
+
+# Define Gradio interface
+demo = gr.Interface(
+    fn=recommend,
+    inputs=gr.Image(type="pil", label="Upload an image"),
+    outputs=gr.Gallery(label="Top-3 Recommended Images"),
+    title="CLIP Image Recommendation System",
+    description="Upload an image and receive visually similar product recommendations."
+)
+
+
+# Launch the application
+demo.launch()