Skip to content

Latest commit

 

History

History
63 lines (33 loc) · 3.99 KB

README.md

File metadata and controls

63 lines (33 loc) · 3.99 KB

Grid cells for 3D Object Recall

Structure

.py files in the base folder contain working code for Grid cells (data processing and SDR creation). The notebooks/ folder contains experimental files and notebooks for exploring grid cell algorithms, training, inference, and validation of data. Use these notebooks with caution.

Overview

All the core code for Grid cells have been moved to the Monty frameworks. The main experiment class that uses grid cells is in frameworks/experiments/htm_experiment.py. The separate L4 and L6a (i.e. TM and grid cell layer) algorithms can be found in frameworks/models/htm.py. The Learning Module that uses both of these layers and creates a joint L4 / L6a algorithm can be found in frameworks/models/htm_learning_modules.py.

Please run python tbp.monty/projects/monty_runs/run.py -e <experiment name>. Experiment configs can be found here in projects/monty_runs/experiments/htm.py -- please edit these configs as needed to run any custom experiments. The run.py script will fail if the processed dataset and curvature SDRs are not found. If it is not found, then follow the script's instructions to generate this dataset (with the help of scripts like process_data.py, curvature_encoder.py, generate_paths.py and/or generate_random.py).

Tutorial

Data generation

Test to see if the conda installation is successful by checking to see if the following import works:

python

from nupic.bindings.math import SparseMatrixConnections

Ensure that trimesh is the correct version: pip install trimesh==3.10.8. Later versions of trimesh don't always produce the correct results for these experiments.

Flags

sdr_p: where to save the dataset

ycb_p: root directory of the YCB path (most likely ~/tbp/data/habitat/objects/ycb)

-r: hash radius in Cartesian Coordinate space (ultimately determines how much one SDR overlaps with another)

-objects: list of objects (separated by a space) to save data for

-n: the total number of bits in the SDR in the case of the curvature_encoder.py

-w: total number of ON bits in the SDR in the case of the curvature_encoder.py

-num_paths: number of paths to enerate for training and testing in the case of generate_paths.py

-path_size: length of each path to generate for training and test in the case of generate_paths.py

-num_points: number of uniformly random points to generate for training and testing in the case of generate_random.py

Pre-process the data

Update <YOU> in the following lines to reflect your user ID. Replace other arguments with < > and specify your own. The following are single-line commands.

python ~/tbp/tbp.monty/projects/grid_cells/process_data.py -sdr_p /Users/<YOU>/tbp/tbp.monty/projects/grid_cells/grid_dataset -ycb_p ~/tbp/data/habitat/objects/ycb

python ~/tbp/tbp.monty/projects/grid_cells/curvature_encoder.py -sdr_p /Users/<YOU>/tbp/tbp.monty/projects/grid_cells/grid_dataset -r 5 -objects <0 1 2 3...> -n 1024 -w 11

Run either generate_paths.py to train on sequences of somewhat continuous paths along an object's surface. Or run generate_random.py to train on uniformly random points along an object's surface.

python ~/tbp/tbp.monty/projects/grid_cells/generate_paths.py -sdr_p /Users/<YOU>/tbp/tbp.monty/projects/grid_cells/grid_dataset -objects <0 1 2 3...> -num_paths 50 -path_size 10

python ~/tbp/tbp.monty/projects/grid_cells/generate_random.py -sdr_p /Users/<YOU>/tbp/tbp.monty/projects/grid_cells/grid_dataset -objects <0 1 2 3...> -num_points 500

Run the experiment

python tbp.monty/projects/monty_runs/run.py -e <CONFIG NAME>. For example, you can run python run.py -e htm_random_path_base if you ran generate_paths.py or python run.py -e htm_random_points_base if you ran generate_random.py.