git clone https://github.com/yourusername/532Project.git
cd 532Project
pip install -r requirements.txtImages to be processed and stored in the pinecone database must be put into the images/ directory. We used Flikr 30k Images for testing.
Install Docker onto your computer and run daemon thread. Run following command once Docker is installed to setup the Kafka server:
docker compose upTo begine pyspark preprocessing of images directory:
python preprocessing/process_pipeline.pyRun the following to start the image consumer. The image consumer will consume messages from the kafka topic, create embeddings from incoming image vectors, and send image embeddings to the pinecone database.
python CLIP_server/image_consumer.pyFirst, start the query server so user queries can be embedded using the Text CLIP model:
python CLIP_server/query_server.py --port <port>Next, start the query pipeline to retrieve an image based on a natural language query. For an image to actually open on your computer, you must have it downloaded on your local computer with the same name as the one stored on pinecone. The query "otter swimming" should work with the test image in images/.
python main_client.pyBy running the preprocessing pipeline with image_consumer.py and process_pipeline.py as explained above, it will automatically complete the pipleline with test images in the images/directory with no other setup.
Python script is set up to test the query server with 10,000 randomized queries with 1000 threads at a time. Test results will be printed out once all queries are processed by the server.
python CLIP_server/test_query_server.pyTo test the full user-side pipeline from user query to returned image, set the custom_query variable in main_client.py to False. This will run a test which sends 50 requests randomly in 5 second intervals, in order to simulate realistic user loads. Tests results are printed out after completion.
python main_client.pyThe image and text CLIP models will log their throughput automatically every 5 seconds, so the query server and image consumers will both display throughput over time. To visualize the image mdoel's throughput over time after doing image preprocessing, run the following script which will analyze the latest Kafka logs and plot throughput over time.
python plot_throughput.py