GemFire Image Search

1. Click “Create Index”

2. Click “Load Embeddings” the application will then

   • Determine if the /static/images directory exists in the backend directory and if there are any files in that directory.
     • If the directory DOES NOT exist, the application will create the directory, download the images dataset (~1.9 GB), unzip the file, and extract the images to the directory (~25,000 images)
     • If the directory DOES exist, then it will skip the above step.
   • Determine if the precomputed embeddings file exists. If it does not, then it will download the file.
   • The precomputed embeddings are then loaded into GemFire.

3. Now that the index has been created, images are available, and the embeddings have been loaded into GemFire, you can search for images.

The full image set can also be downloaded here if needed: Image Data Set

If you would like to use your own images data set:

• The images must be placed in the backend/static/images directory.
• The images must be .jpg files.
• You must set use_precomputed_embeddings (~line 20) in the embeddings_processor.py to False.
  • The application will randomly select and encode 100 images from this folder. This is only because the encoding takes time and memory. You can adjust this number if needed in the embeddings_processor.py file, in the load_embeddings method (~line 33).
• With 1500 images on a local machine it can take about ~20 minutes to encode them. This number will vary for each user.

When searching images

• The application filters out results that have a score less than .60. This is set in the app.py file, in the /searchImages endpoint.
• The search query sets top-k:100 - this means it will return the top 100 results from the query. This can be changed in the