Spaces:
Runtime error
Runtime error
| from colordescriptor import ColorDescriptor | |
| from CLIP import CLIPImageEncoder | |
| from LBP import LBPImageEncoder | |
| from helper import chi2_distance, euclidean_distance, merge_features | |
| import gradio as gr | |
| import os | |
| import cv2 | |
| import numpy as np | |
| from datasets import * | |
| dataset = load_dataset("nielsr/CelebA-faces", download_mode='force_redownload') | |
| dataset.cleanup_cache_files() | |
| candidate_subset = dataset["train"].select(range(500)) # This is a small CBIR app! :D | |
| def index_dataset(dataset): | |
| print(dataset) | |
| print("LBP Embeddings") | |
| lbp_model = LBPImageEncoder(8,2) | |
| dataset_with_embeddings = dataset.map(lambda row: {'lbp_embeddings': lbp_model.describe(row["image"])}) | |
| print("Color Embeddings") | |
| cd = ColorDescriptor((8, 12, 3)) | |
| dataset_with_embeddings = dataset_with_embeddings.map(lambda row: {'color_embeddings': cd.describe(row["image"])}) | |
| print("CLIP Embeddings") | |
| clip_model = CLIPImageEncoder() | |
| dataset_with_embeddings = dataset_with_embeddings.map(clip_model.encode_images, batched=True, batch_size=16) | |
| print("LBP and Color") | |
| dataset_with_embeddings = dataset_with_embeddings.map(lambda row: {'lbp_color_embeddings': merge_features(row['lbp_embeddings'], row['color_embeddings'])}) | |
| # Add index | |
| dataset_with_embeddings.add_faiss_index(column='color_embeddings') | |
| dataset_with_embeddings.save_faiss_index('color_embeddings', 'color_index.faiss') | |
| dataset_with_embeddings.add_faiss_index(column='clip_embeddings') | |
| dataset_with_embeddings.add_faiss_index(column='lbp_embeddings') | |
| dataset_with_embeddings.save_faiss_index('clip_embeddings', 'clip_index.faiss') | |
| print(dataset_with_embeddings) | |
| return dataset_with_embeddings | |
| def check_index(ds): | |
| index_path = "my_index.faiss" | |
| if os.path.isfile('color_index.faiss') and os.path.isfile('clip_index.faiss'): | |
| ds.load_faiss_index('color_embeddings', 'color_index.faiss') | |
| return ds.load_faiss_index('clip_embeddings', 'clip_index.faiss') | |
| else: | |
| return index_dataset(ds) | |
| dataset_with_embeddings = check_index(candidate_subset) | |
| # Main function, to find similar images | |
| # TODO: implement different distance measures | |
| def get_neighbors(query_image, selected_descriptor, selected_distance, top_k=5): | |
| """Returns the top k nearest examples to the query image. | |
| Args: | |
| query_image: A PIL object representing the query image. | |
| top_k: An integer representing the number of nearest examples to return. | |
| Returns: | |
| A list of the top_k most similar images as PIL objects. | |
| """ | |
| if "Color Descriptor" == selected_descriptor: | |
| cd = ColorDescriptor((8, 12, 3)) | |
| qi_embedding = cd.describe(query_image) | |
| qi_np = np.array(qi_embedding) | |
| if selected_distance == "FAISS": | |
| scores, retrieved_examples = dataset_with_embeddings.get_nearest_examples( | |
| 'color_embeddings', qi_np, k=top_k) | |
| elif selected_distance == "Chi-squared": | |
| tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': chi2_distance(qi_embedding,row['color_embeddings'])}) | |
| retrieved_examples = tmp_dataset.sort("distance")[:5] | |
| else: | |
| tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': euclidean_distance(qi_embedding,row['color_embeddings'])}) | |
| retrieved_examples = tmp_dataset.sort("distance")[:5] | |
| images = retrieved_examples['image'] #retrieved images is a dict, with images and embeddings | |
| return images | |
| if "CLIP" == selected_descriptor: | |
| clip_model = CLIPImageEncoder() | |
| qi_embedding = clip_model.encode_image(query_image) | |
| if selected_distance == "FAISS": | |
| scores, retrieved_examples = dataset_with_embeddings.get_nearest_examples( | |
| 'clip_embeddings', qi_embedding, k=top_k) | |
| elif selected_distance == "Chi-squared": | |
| tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': chi2_distance(qi_embedding, row['clip_embeddings'])}) | |
| retrieved_examples = tmp_dataset.sort("distance")[:5] | |
| else: | |
| tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': euclidean_distance(qi_embedding, row['clip_embeddings'])}) | |
| retrieved_examples = tmp_dataset.sort("distance")[:5] | |
| images = retrieved_examples['image'] | |
| return images | |
| if "LBP" == selected_descriptor: | |
| lbp_model = LBPImageEncoder(8,2) | |
| qi_embedding = lbp_model.describe(query_image) | |
| if selected_distance == "FAISS": | |
| scores, retrieved_examples = dataset_with_embeddings.get_nearest_examples( | |
| 'lbp_embeddings', qi_embedding, k=top_k) | |
| elif selected_distance == "Chi-squared": | |
| tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': chi2_distance(qi_embedding, row['lbp_embeddings'])}) | |
| retrieved_examples = tmp_dataset.sort("distance")[:5] | |
| else: | |
| tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': euclidean_distance(qi_embedding, row['lbp_embeddings'])}) | |
| retrieved_examples = tmp_dataset.sort("distance")[:5] | |
| images = retrieved_examples['image'] | |
| return images | |
| if "LBPColor" == selected_descriptor: | |
| lbp_model = LBPImageEncoder(8,2) | |
| cd = ColorDescriptor((8, 12, 3)) | |
| qi_embedding = merge_features(lbp_model.describe(query_image), cd.describe(query_image)) | |
| if selected_distance == "FAISS": | |
| scores, retrieved_examples = dataset_with_embeddings.get_nearest_examples( | |
| 'lbp_color_embeddings', qi_embedding, k=top_k) | |
| elif selected_distance == "Chi-squared": | |
| tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': chi2_distance(qi_embedding, row['lbp_color_embeddings'])}) | |
| retrieved_examples = tmp_dataset.sort("distance")[:5] | |
| else: | |
| tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': euclidean_distance(qi_embedding, row['lbp_color_embeddings'])}) | |
| retrieved_examples = tmp_dataset.sort("distance")[:5] | |
| images = retrieved_examples['image'] | |
| return images | |
| else: | |
| print("This descriptor is not yet supported :(") | |
| return [] | |
| # Define the Gradio Interface | |
| with gr.Blocks() as demo: | |
| gr.Markdown(""" | |
| # Welcome to this CBIR app | |
| This is a CBIR app focused on the retrieval of similar faces. | |
| ## Find similar images | |
| Here you can upload an image, that is compared with existing image in our dataset. | |
| """) | |
| with gr.Row(): | |
| image_input = gr.Image(type="pil", label="Please upload your image") | |
| gallery_output = gr.Gallery() | |
| btn = gr.Button(value="Submit") | |
| gr.Markdown(""" | |
| ## Settings | |
| Here you can adjust how the images are found | |
| """) | |
| with gr.Row(): | |
| descr_dropdown = gr.Dropdown(["Color Descriptor", "LBP", "CLIP", "LBPColor"], value="LBP", label="Please choose an descriptor") | |
| dist_dropdown = gr.Dropdown(["FAISS", "Chi-squared", "Euclid"], value="FAISS", label="Please choose a distance measure") | |
| btn.click(get_neighbors, inputs=[image_input, descr_dropdown, dist_dropdown], outputs=[gallery_output]) | |
| if __name__ == "__main__": | |
| demo.launch() | |