Search over large image datasets with natural language and computer vision!


Keywords
computer-vision, image-search, local-search, natural-language, python
License
MIT
Install
pip install memery==0.1.0

Documentation

Memery

Use human language to search your image folders!

What is memery?

meme about having too many memes

The problem: you have a huge folder of images. Memes, screenshots, datasets, product photos, inspo albums, anything. You know that somewhere in that folder is the exact image you want, but you can't remember the filename or what day you saved it. There's nothing you can do but scroll through the folder, skimming hundreds of thumbnails, hoping you don't accidentally miss it, and that you'll recognize it when you do see it.

Humans do this amazingly well. But even with computers, local image search is still a manual effort - you're still sorting through folders of images, like an archivist of old.

Now there's Memery.

The memery package provides natural language search over local images. You can use it to search for things like "a line drawing of a woman facing to the left" and get reasonably good results!

You can do this over thousands of images (it's not optimized for performance yet, but search times scale well under O(n)).

You can view the images in a browser GUI, or pipe them through command line tools.

You can use memery or its modules in Jupyter notebooks, including GUI functions!

Under the hood, memery makes use of CLIP, the Contrastive Language-Image Pretraining transformer, released by OpenAI in 2021. CLIP trains a vision transformer and a language transformer to find the same latent space for images and their captions. This makes it perfect for the purpose of natural language image search. CLIP is a giant achievement, and memery stands on its shoulders.

Outline:

  • Usage
    • Install locally
    • Use GUI
    • Use CLI
    • Use in Jupyter
    • Use the library
  • Development
  • Contributing
    • Who works on this project

Installation

With Python 3.6 or greater:

pip install memery
pip install git+https://github.com/openai/CLIP.git

Currently memery defaults to GPU installation. This will probably be switched in a future version.

For now, if you want to run CPU-only, run the following command after installing memery:

pip install torch==1.7.1+cpu torchvision==0.8.2+cpu torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

Someday memery will be packaged in an easy to use format, but since this is a Python project it is hard to predict when that day will be.

If you want to help develop memery, you'll need to clone the repo. See below.

Usage

What's your use case?

I have images and want to search them with a GUI app

↳ Use the Browser GUI

i have a program/workflow and want to use image search as one part of it

↳ Use in Jupyter

↳ Use as a Python module

↳ Use from command line or shell scripts

i want to improve on and/or contribute to memery development

↳ Start by cloning the repo

Use GUI

Currently memery has a rough browser-based GUI. To launch it, run the following in a command line:

memery serve

or set up a desktop shortcut that points to the above command.

You can pass a directory to open on startup, like so:

memery serve home/user/Pictures/memes

Relative directories will also work:

cd ~/Pictures
memery serve memes

The default directory passed will be ./images, which is memery's example meme directory.

Memery will open in a browser window. The interface is pretty straightforward, but it has some quirks.

screenshot of memery GUI displaying wholesome memes

The sidebar on the left controls the location and query for the search. The "Directory" box requires a full directory path; unfortunately, Streamlit does not yet have a folder-picker component. The path is relative to your current working directory when you run memery serve.

The search will run once you enter a text or image query. If you enter both text and image queries, memery will search for the combination.

Beneath these widgets is the output area for temporary messages displayed with each search. Mostly this can be ignored.

The right hand panel displays the images and associated options. Major errors will appear here as giant stack traces; sometimes, changing variables in the other widgets will fix these errors live. If you get a large error here it's helpful to take a screenshot and share it with us in Github Issues.

Use in Jupyter

If you're in a Jupyter environment, you can summon an ipywidgets GUI directly into an output cell like this:

from memery.gui import appPage

app = appPage()
display(app)
<memery.gui.appPage at 0x7f829c87d290>

screenshot of memery Jupyter GUI displaying wholesome memes

Unfortunately the widgets won't display without an active runtime, so the above is a screenshot. In a Jupyter environment, the GUI works just like the browser-based GUI.

Opening new appPage instances in separate cells can be helpful for exploring a dataset, but it can also cause memory overruns. For this reason the Jupyter interface is tabbed and each search will appear in a new tab.

Use CLI

The memery command line is currently very rudimentary. So far you can use memery on any folder and it will search for images recursively, returning a list object to stdout.

Pass the --n flag to control how many images are returned (default 10).

memery recall PATH/TO/IMAGE/FOLDER 'query' --n 20

I'm not clear yet on what behavior command-line users will expect from it. In the future we may want to make it non-recursive by default, and map the behavior more closely to POSIX standards. I would love feedback or help on this. `

Use as a library

The function currently called query_flow accepts a folder name and a query and returns a ranked list of image files. This is also recursive by default, and prints too much information into stdout. And it calls index_flow every time, which can be annoying if you have corrupted files in the directory, as it will need to rebuild the tree-index each time despite the files not changing. This will be solved in a future release.

from memery.core import query_flow
ranked = query_flow('./images', 'dad joke')

print(ranked[:5])
89it [00:00, 44503.23it/s]
89it [00:00, 37660.72it/s]
0it [00:00, ?it/s]

starting timer
Checking files
Indexing
Skipping bad file: images/memes/corrupted-file.jpeg
due to <class 'PIL.UnidentifiedImageError'>
Skipping bad file: images/memes/.ipynb_checkpoints/corrupted-file-checkpoint.jpeg
due to <class 'PIL.UnidentifiedImageError'>
Loaded 82 encodings
Encoding 0 new images


0it [00:00, ?it/s]

Building treemap





Saving 82 encodings
Converting query
Searching 82 images
Done in 0.5853581428527832 seconds
['images/memes/Wholesome-Meme-68.jpg', 'images/memes/Wholesome-Meme-74.jpg', 'images/memes/Wholesome-Meme-88.jpg', 'images/memes/Wholesome-Meme-78.jpg', 'images/memes/Wholesome-Meme-23.jpg']

Here's the first result from that list:

So that's how to use memery. Let's look at how you can help make it better.

Development

Pull the repo

Clone this repository from Github:

git clone https://github.com/deepfates/memery.git

Install dependencies and memery

Enter the memery folder and install requirements:

cd memery
pip install requirements.txt

We also have to download the CLIP model directly from their git repo, as they decline to upload it to PyPi:

pip install git+https://github.com/openai/CLIP.git

And finally install your local, editable copy of memery with

pip install -e .

Contributing

Memery is open source and you can contribute. See CONTRIBUTING.md for guidelines on how you can help.

Who works on this project

Memery was first written by Max Anton Brewer aka @deepfates in the summer of 2021. Some commits are listed from @robotface-io but that was just me using the wrong account when I first started.

I wrote this to solve my own needs and learn notebook-based development. I hope it helps other people too. If you can help me make it better, please do. I welcome any contribution, guidance or criticism.

The ideal way to get support is to open an issue on Github. However, the fastest way to get a response from me is probably to direct message me on twitter. This is my first attempt to coordinate an open-source project, bear with me.