The repository contains sample implementations for the Multigrid Neural Memory architectures.
To cite this work, please use:
@INPROCEEDINGS{HuynhICML2020,
author = {Tri Huynh and Michael Maire and Matthew R. Walter},
title = {Multigrid Neural Memory},
booktitle = {International Conference on Machine Learning (ICML)},
year = {2020}
}
- Tensorflow 1.15
- Python 2.7
- OpenCV-Python
- Python-MNIST
Subdirectories contain sample code for corresponding tasks:
- mapping_localization: mapping and localization task. The sample code for an agent moving in spiral motion with 3x3 observations and 3x3 queries is provided.
- mnist_recall: associative recall + classification task on mnist digits.
- The code in each directory is self-contained. Training can be invoked by calling:
python main.py
- Learning curves should look like the following:
MNIST Associative Recall & Classification:
- Inference visualization is generated periodically during training, and saved in CHECKPOINT_DIR/[TIME-STEP]. The interval for saving visualization can be controlled by SAVE_VIS_INTERVAL in constants.py. Sample visualizations are provided below:
Mapping & Localization Task:
See the project page for annotations and more demos.
MNIST Associative Recall & Classification:
Top row visualizes the input sequence. Bottom row shows the query image and the predicted output class of the following digit aligned to their corresponding positions in the top row.
Part of the code structure was adapted from miyosuda's implementation of A3C.