HCNAF - Hyper-conditioned Neural Autoregressive Flow

• Pytorch implementation of HCNAF by Geunseob (GS) Oh, Jean-Sebastien Valois
• Link to the paper - CVPR 2020

Requirements

• Python 3
• PyTorch
• NumPy

HCNAF_Gaussians experiments

Training

EXP: Gaussian 1

python train_hcnaf_gaussians.py --dataset gaussians_exp1 --batch_dim 64 --clip_gradnorm_max 1.0 --n_layers_flow 2 --dim_h_flow 64 --hypernet_layers 2 --norm_HW modified_weightnorm --patience 20

EXP: Gaussian 2

python train_hcnaf_gaussians.py --dataset gaussians_exp2 --batch_dim 4 --clip_gradnorm_max 0.1 --n_layers_flow 3 --dim_h_flow 200 --hypernet_layers 2 --norm_HW scaled_frobenius --patience 50

Testing

Task: plot

python test_hcnaf_gaussians.py --task plot --loadpath $PATH_TO_MODEL_FOLDER --loadfilename $MODEL_FILENAME

Task: NLL computation

python test_hcnaf_gaussians.py --task NLL --loadpath $PATH_TO_MODEL_FOLDER --loadfilename $MODEL_FILENAME

HCNAF_PRECOG_Carla experiments

• HCNAF models are trained & evaluated on CARLA Town01 data
• NOTE: The default path to the data folder is set to data/precog_carla/town1. Make sure to create data/precog_carla/town1 folder and that the town1 folder contains data folders named as "train", "val", "test".

Training

Model: with lidar (a large model)

python train_PRECOG_Carla.py --dataset PRECOG_Carla --ablation_mode All_faster_temporal --batch_size 8 --n_layers_flow 3 --dim_h_flow 100 --norm_HW modified_weightnorm --loss PNLL_output --temporal 1

Model: with lidar (a small model)

python train_PRECOG_Carla.py --dataset PRECOG_Carla --ablation_mode All_faster_temporal --batch_size 4 --n_layers_flow 3 --dim_h_flow 20 --norm_HW scaled_frobenius --loss PNLL_output --temporal 1

Model: without lidar

python train_PRECOG_Carla.py --dataset PRECOG_Carla --ablation_mode No_lidar_faster_temporal --batch_size 4 --n_layers_flow 3 --dim_h_flow 100 --norm_HW modified_weightnorm --loss PNLL_output --temporal 1 --learning_rate 0.0002

Testing

Task: plot

python test_PRECOG_Carla.py --task plot --loadpath $PATH_TO_MODEL_FOLDER --loadfilename $MODEL_FILENAME

Task: extra PNLL computation

python test_PRECOG_Carla.py --task extra_PNLL --loadpath $PATH_TO_MODEL_FOLDER --loadfilename $MODEL_FILENAME

Citation (BibTeX)

@InProceedings{Oh_2020_CVPR,
author = {Oh, Geunseob and Valois, Jean-Sebastien},
title = {HCNAF: Hyper-Conditioned Neural Autoregressive Flow and its Application for Probabilistic Occupancy Map Forecasting},
booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}

License

Licensed under the Apache License 2.0