Implementation of the IJRR paper:

Locally Connected Interrelated Network: A Forward Propagation Primitive

The code implements the LCI-Net module embedded in VIN and QMDP-net architectures.

Python 3.6 or newer (tested up to Python 3.9) Tensorflow 1.14 or newer (tested up to Tensorflow 2.7) Python packages: numpy, scipy, pillow

To install these packages using pip:

pip install tensorflow
pip install numpy scipy pillow

python train.py [input_file] --network_type [network_type] --logpath [dir_to_save_models]

The given input_file must be a file produced by env_generator.py (or an empty file of the same format, e.g. for imagegrid large scale realistic envs).

If the input file env type is fully observable (i.e. name contains 'full_obs'), then network type must be 'vin_baseline' or 'vin_lcinet'. Otherwise, network type must be 'qmdpnet_baseline' or 'qmdpnet_lcinet').

e.g.

python train.py inputs/env_grid_nondet_10x10_example.txt --network_type qmdpnet_lcinet --logpath outputs/example

To continue training an already existing model, a loadmodel path can be specified:

python train.py [input_file] --network_type [network_type] --loadmodel [dir_of_saved_model] --logpath [dir_to_save_models]

The loadmodel path should be the same as the logpath used when the existing model was first trained, e.g.

python train.py inputs/env_grid_nondet_10x10_example.txt --network_type qmdpnet_lcinet --loadmodel outputs/example --logpath outputs/continued

An existing model can be evaluated individually by setting epochs to zero, e.g.

python train.py [input_file] --network_type [network_type] --loadmodel [dir_of_saved_model] --epochs 0

If networks have been trained for both network types (LCI-net and original), a comparison between both types on the same set of maps and trajectories can be performed via:

python evaluate.py [input_file] --loadmodel [dir_of_saved_models]

The number of maps, trajectories and repetitions can be manually specified - refer to arguments.py to see available arguments.

Environment files can be generated via the following command:

python env_generator.py [env_type] [num_maps] [trajs_per_map] [name_suffix]

e.g. to generate an input file with 2000 dynamic maze v2 maps with 5 trajectories per map with suffix 'example':

python env_generator.py dynmaze_v2_nondet 2000 5 example

Refer to env_generator.py to find the supported env_type values.