It is essential to install install torchdiffeq
git clone https://github.com/rtqichen/torchdiffeq.git
cd torchdiffeq
pip install -e .
In order to calculate Wasserstein Earth Mover's Distance (WEMD), one must compile the WEMD C++ library. We use the same implementation of as the implementation of Overcoming Limitations of Mixture Density Networks: A Sampling and Fitting Framework for Multimodal Future Prediction paper.
Installation steps:
cd wemd
mkdir build
cd build
cmake ..
makeAs a result, wemd/lib/libwemd.so file should be created.
The toy example provided in the paper can be run with the script:
python train_regression_SDD.py --log_name experiment_regression_flow_toy \
--lr 1e-3 --num_blocks 1 --batch_size 1000 --epochs \
1000 --save_freq 5 --viz_freq 1 --log_freq 1 --gpu 0 --hyper_dims \
128-32 --dims 32-32-32 --input_dim 1 --weight_decay 1e-5The script saves the model with --save_freq and visualize the results with viz_freq frequencies.
The model can be loaded and tested using:
python test_toy.py --resume_checkpoint /experiment_regression_flow_toy/experiment_regression_flow_toy/checkpoint-latest.pt
--num_blocks 1 --gpu 0 --hyper_dims 128-32
--dims 32-32-32 --input_dim 1 The train and test data are located at:
https://lmb.informatik.uni-freiburg.de/resources/binaries/Multimodal_Future_Prediction/sdd_train.zip
https://lmb.informatik.uni-freiburg.de/resources/binaries/Multimodal_Future_Prediction/sdd_test.zip
The datasets should be located in data\SDD\train and data\SDD\test locations.
You can run the training procedure with the script train_regression_SDD.py using the following settings:
train_regression_SDD.py --log_name "experiment_regression_flow_SSD" \
--lr 2e-5 --num_blocks 1 --batch_size 20 --epochs 100 --save_freq 1 \
--viz_freq 1 --log_freq 1 --gpu 0 --dims 128-128-128 --input_dim 2
Validation can be run using script test_SDD with parameters:
python test_SDD.py --data_dir /data/SDD \
--resume_checkpoint /experiment_regression_flow_SSD/checkpoint-latest.pt \
--num_blocks 1 --gpu 0 --dims 128-128-128 --input_dim 2
First, the data should be obtained from:
Specifically you will need these files:
US-101:
'0750am-0805am/trajectories-0750am-0805am.txt'
'0805am-0820am/trajectories-0805am-0820am.txt'
'0820am-0835am/trajectories-0820am-0835am.txt'
I-80:
'0400pm-0415pm/trajectories-0400-0415.txt'
'0500pm-0515pm/trajectories-0500-0515.txt'
'0515pm-0530pm/trajectories-0515-0530.txt'The files should be further processed using prepocess_data.m from:
https://github.com/nachiket92/conv-social-pooling
After processing files TrainSet.mat, ValSet.mat, and TestSet.mat should be created.
The model for this dataset can be trained with the following script:
python networks_regression_NGSIM.py
--log_name experiment_regression_flow_NGSIM
--lr 1e-3 --num_blocks 1 --batch_size 10000
--epochs 100 --save_freq 1 --viz_freq 1 --log_freq
1 --val_freq 6000 --gpu 0 --dims 16-16-16 --input_dim
2 --weight_decay 1e-5 --data_dir /data/files/locationThe evaluation for the dataset can be run using:
python test_NGSIM.py --data_dir/data/files/location
--resume_checkpoint
/experiment_regression_flow_NGSIM/checkpoint-latest.pt
--num_blocks 1 --gpu 0 --dims 16-16-16
--input_dim 2Synthetic Car-Pedestrian Interaction dataset introduced here, which can be generated with utilities in cpi_generation directory.
To obtain our setup:
cd cpi_generation/
# train dataset
python CPI-generate.py --output_folder cpi/train --n_scenes 20000 --history 3 --n_gts 20 --dist 20
# test dataset
python CPI-generate.py --output_folder cpi/test --n_scenes 54 --history 3 --n_gts 1000 --dist 20The model for this dataset can be trained with
train_regression_CPI.py --log_name "experiment_regression_flow_CPI" \
--lr 2e-5 --num_blocks 1 --batch_size 20 --epochs 100 --save_freq 1 \
--viz_freq 1 --log_freq 1 --gpu 0 --dims 128-128-128 --input_dim 2 --data_dir cpi_generation/cpi
To test it:
python test_CPI.py --data_dir cpi_generation/cpi \
--resume_checkpoint experiment_regression_flow_CPI/checkpoint-latest.pt \
--num_blocks 1 --gpu 0 --dims 128-128-128 --input_dim 2