natural-language-joint-query-search

In the project, we support multiple types of query search including text-image, image-image, text2-image, and text+image-image. In order to analyze the result of retrieved images, we also support visualization of text attention. The attention of image will be supported soon!

Colab Demo

Search photos on Unsplash, support for joint image+text queries search.

Attention visualization of CLIP.

Usage

We follow the same environment as the CLIP project:

$ conda install --yes -c pytorch pytorch=1.7.1 torchvision cudatoolkit=11.0
$ pip install ftfy regex tqdm

To visualize the attention of CLIP, we slightly modify the code of CLIP as mention here, so you don’t have to install CLIP via official command. An open-sourced visualization tool is used in our project, you need to clone it into this repo.

$ git clone https://github.com/shashwattrivedi/Attention_visualizer.git

Download the pre-extracted image id and features of Unsplash dataset from Google Drive or just run the following commands, and put them under unsplash-dataset dir, details can be found in natural-language-image-search project.

from pathlib import Path
# Create a folder for the precomputed features
!mkdir unsplash-dataset
# Download from Github Releases
if not Path('unsplash-dataset/photo_ids.csv').exists():
  !wget https://github.com/haltakov/natural-language-image-search/releases/download/1.0.0/photo_ids.csv -O unsplash-dataset/photo_ids.csv
if not Path('unsplash-dataset/features.npy').exists():
  !wget https://github.com/haltakov/natural-language-image-search/releases/download/1.0.0/features.npy -O unsplash-dataset/features.npy

Example of joint query search.

import torch
import numpy as np
import pandas as pd
from PIL import Image
from CLIP.clip import clip
def encode_search_query(search_query):
    with torch.no_grad():
        text_encoded, weight = model.encode_text(clip.tokenize(search_query).to(device))
        text_encoded /= text_encoded.norm(dim=-1, keepdim=True)
        return text_encoded.cpu().numpy()
def find_best_matches(text_features, photo_features, photo_ids, results_count):
  similarities = (photo_features @ text_features.T).squeeze(1)
  best_photo_idx = (-similarities).argsort()
  return [photo_ids[i] for i in best_photo_idx[:results_count]]
device = "cuda" if torch.cuda.is_available() else "cpu"
model, preprocess = clip.load("ViT-B/32", device=device)
photo_ids = pd.read_csv("unsplash-dataset/photo_ids.csv")
photo_ids = list(photo_ids['photo_id'])
photo_features = np.load("unsplash-dataset/features.npy")
# text to image
search_query = "Tokyo Tower at night."
text_features = model.encode_search_query(search_query)
best_photo_ids = find_best_matches(text_features, photo_features, photo_ids, 5)
for photo_id in best_photo_ids:
  print("https://unsplash.com/photos/{}/download".format(photo_id))
# image to image
source_image = "images/borna-hrzina-8IPrifbjo-0-unsplash.jpg"
with torch.no_grad():
  image_feature = model.encode_image(preprocess(Image.open(source_image)).unsqueeze(0).to(device))
  image_feature = (image_feature / image_feature.norm(dim=-1, keepdim=True)).cpu().numpy()
best_photo_ids = find_best_matches(image_feature, photo_features, photo_ids, 5)
for photo_id in best_photo_ids:
  print("https://unsplash.com/photos/{}/download".format(photo_id))
# text+text to image
search_query = "red flower"
search_query_extra = "blue sky"
text_features = encode_search_query(search_query)
text_features_extra = encode_search_query(search_query_extra)
mixed_features = text_features + text_features_extra
best_photo_ids = find_best_matches(mixed_features, photo_features, photo_ids, 5)
for photo_id in best_photo_ids:
  print("https://unsplash.com/photos/{}/download".format(photo_id))
# image+text to image
search_image = "images/borna-hrzina-8IPrifbjo-0-unsplash.jpg"
search_text = "cars"
with torch.no_grad():
  image_feature = model.encode_image(preprocess(Image.open(search_image)).unsqueeze(0).to(device))
  image_feature = (image_feature / image_feature.norm(dim=-1, keepdim=True)).cpu().numpy()
text_feature = encode_search_query(search_text)
modified_feature = image_feature + text_feature
best_photo_ids = find_best_matches(modified_feature, photo_features, photo_ids, 5)
for photo_id in best_photo_ids:
  print("https://unsplash.com/photos/{}/download".format(photo_id))

Example of CLIP attention visualization. You can know which keywords does CLIP use to retrieve the results. To be convenient, all punctuations are removed.

import torch
import numpy as np
import pandas as pd
from PIL import Image
from CLIP.clip import clip
from CLIP.clip import model
from Attention_visualizer.attention_visualizer import *
def find_best_matches(text_features, photo_features, photo_ids, results_count):
  similarities = (photo_features @ text_features.T).squeeze(1)
  best_photo_idx = (-similarities).argsort()
  return [photo_ids[i] for i in best_photo_idx[:results_count]]
device = "cuda" if torch.cuda.is_available() else "cpu"
model, preprocess = clip.load("ViT-B/32", device=device, jit=False)
photo_ids = pd.read_csv("unsplash-dataset/photo_ids.csv")
photo_ids = list(photo_ids['photo_id'])
photo_features = np.load("unsplash-dataset/features.npy")
search_query = "A red flower is under the blue sky and there is a bee on the flower"
with torch.no_grad():
    text_token = clip.tokenize(search_query).to(device)
    text_encoded, weight = model.encode_text(text_token)
    text_encoded /= text_encoded.norm(dim=-1, keepdim=True)
text_features = text_encoded.cpu().numpy()
best_photo_ids = find_best_matches(text_features, photo_features, photo_ids, 5)
for photo_id in best_photo_ids:
  print("https://unsplash.com/photos/{}/download".format(photo_id))
sentence = search_query.split(" ")
attention_weights = list(weight[-1][0][1+len(sentence)].cpu().numpy())[:2+len(sentence)][1:][:-1]
attention_weights = [float(item) for item in attention_weights]
display_attention(sentence,attention_weights)

You can also run these example on Colab via joint-query-search and clip-attention.

Example

Text-to-Image

“Tokyo tower at night.”

“People come and go on the street.”

Image-to-Image

A normal street view. (The left side is the source image)

Text+Text-to-Image

“Flower” + “Blue sky”

“Flower” + “Bee”

Image+Text-to-Image

A normal street view + “cars”

Visualization

“A woman holding an umbrella standing next to a man in a rainy day”

“umbrella”, “standing” and “rainy” receive the most of attention.

“A red flower is under the blue sky and there is a bee on the flower”

“flower”, “sky” and “bee” receive the most of attention.

Acknowledgements

Search photos on Unsplash using natural language descriptions. The search is powered by OpenAI’s CLIP model and the Unsplash Dataset. This project is mostly based on natural-language-image-search.

This project was inspired by these projects:

• OpenAI’s CLIP
• natural-language-image-search