app.py

import gradio as gr
import torch
import pickle
import numpy as np
from PIL import Image
from torchvision import transforms
from huggingface_hub import hf_hub_download
from sklearn.metrics.pairwise import cosine_similarity
import timm

# Load the precomputed fashion embeddings
file_path = hf_hub_download(
    repo_id="Elevi7/MatchMe",
    filename="fashion_embeddings.pkl",
    repo_type="dataset"
)

with open(file_path, "rb") as f:
    embedding_store = pickle.load(f)

# Load the pretrained model (same one used for embeddings)
model = timm.create_model("resnet18", pretrained=True)
model.eval()

# Define image transformation pipeline
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor()
])

# Extract embedding from uploaded image
def extract_embedding(image):
    image = transform(image).unsqueeze(0)
    with torch.no_grad():
        embedding = model(image).squeeze().numpy()
    return embedding

# Recommendation function with filtering
def recommend(image, category, gender):
    query_embedding = extract_embedding(image)

    # Filter embeddings by category and gender
    filtered_store = [
        item for item in embedding_store
        if item.get("category") == category and item.get("gender") == gender
    ]

    if len(filtered_store) == 0:
        print("No matches found for selected category and gender.")
        return ["", "", ""]  # Return empty images

    all_embeddings = np.array([item["embedding"] for item in filtered_store])
    paths = [item["image_path"] for item in filtered_store]

    similarities = cosine_similarity([query_embedding], all_embeddings)[0]
    top_indices = np.argsort(similarities)[-3:][::-1]  # Top 3

    return [paths[i] for i in top_indices]

# Gradio interface
demo = gr.Interface(
    fn=recommend,
    inputs=[
        gr.Image(type="pil", label="Upload Clothing Image"),
        gr.Dropdown(
            choices=["shoes", "tops", "pants", "handbags", "coats_jackets",
                     "sunglasses", "shorts", "skirts", "earrings", "necklaces"],
            label="Category"
        ),
        gr.Dropdown(choices=["men", "women"], label="Gender")
    ],
    outputs=[
        gr.Image(type="filepath", label="Match 1"),
        gr.Image(type="filepath", label="Match 2"),
        gr.Image(type="filepath", label="Match 3")
    ],
    title="MatchMe: Fashion Recommender",
    description="Upload a fashion image and get 3 visually similar items. Filter by category and gender."
)

demo.launch()