Lego Sketch utilizes scalable memory-augmented neural networks (MANN) to store data streams, enabling future approximate item-frequency queries—a process known as sketching streams. This research falls into the emerging category of neural sketch methods, aiming to replace traditional rule-based handcrafted sketches with a self-supervised end-to-end training paradigm. By implementing notable architectural improvements, it overcomes the scalability challenges of previous neural sketches, eliminates the need for frequent retraining, and achieves substantially lower prediction errors.
To train the Lego Brick model, navigate to the following directory and execute the command:
cd Lego/ExpCode/Lego_brick
python Lego_seed0.py
Important Note: For efficiency, we use a uniform random number generator as a substitute for the hash addressing during training. However, for evaluation, we provide the CUDA kernel for the actual hash function. If you are using the Lego Sketch architecture for your own tasks, you need to confirm that this substitution, made for efficiency purposes, does not introduce errors.
To facilitate accuracy testing, we provide synthetic datasets with different skewness levels and two Lego Brick instances (trained under random seeds 0 and 1). You can directly test them using the following command, and the results will be saved in the Figure folder.
cd Lego/EvalDataset
python Test_Lego.py