Project-Qpid

Abstract

This is the homepage of our code repository Project-Qpid. This repository was created to facilitate the construction, training and testing of our trajectory prediction models (using the pytorch backend now). It includes some basic dataset preprocessing and reading as well as training and testing functions to keep our approaches' code repositories compatible with each other and easy to be managed. Click the following buttons to learn how it works.

Getting Started

Dataset Preparation

Please process the original dataset files to produce qpid-readable data files, including all processed dataset files in dataset_processed and their corresponding configs in dataset_configs. See the required data formats in this document.

For datasets used in our paper, you can use the attached dataset_original module to generate processed dataset files.

For example, you can run the following commands line-by-line to generate processed files in ETH-UCY and SDD (Stanford Drone Dataset):

cd dataset_original
python main_ethucysdd.py

Then, make soft links of these folders (or you can simply copy them manually)

cd ..
ln -s ./dataset_original/dataset_configs ./
ln -s ./dataset_original/dataset_processed ./

These processed dataset folders should be organized in the following way:

(ROOT_DIR)
|__qpid
|__dataset_original
|__dataset_configs
|__dataset_processed
|__...

Code Entrance

The qpid package can be used by adding a simple entrance file, like

import sys
import qpid

if __name__ == '__main__':
    qpid.entrance(sys.argv)

This file is usually named as main.py in our trajectory prediction models' repos. You can start training, testing, or visualizing models by adding different terminal args in the following way:

python main.py --ARG_KEY1 ARG_VALUE2 --ARG_KEY2 ARG_VALUE2 -SHORT_ARG_KEY3 ARG_VALUE3 ...

where ARG_KEY is the name of args, and ARG_VALUE is the corresponding value.

You can run the simple linear prediction model to verify if your dataset files are properly put and processed (please follow the instructions in this page, or the README in different models' repos):

python main.py --model linear --split zara1

Its output may like

>>> [Train Manager]: Test Results
    - ADE(Metrics): 0.6033005118370056 (meter).
    - FDE(Metrics): 1.1830337047576904 (meter).
    - Average Inference Time: 5 ms.
    - Fastest Inference Time: 5 ms.

You are totally set if you get similar results! Otherwise, please check the dataset files and where they are placed.

Args

You can use the following command to view all supported args:

python main.py --help

For example, it may print documents like

...

- `--batch_size` (short for `-bs`):
  Batch size when implementation.
    - Type=`int`, argtype=`dynamic`;
    - The default value is `5000`.

- `--compute_loss`:
  Controls whether to compute losses when testing.
    - Type=`int`, argtype=`temporary`;
    - The default value is `0`.

...

Here, we can see several properties of this arg, including

• Data Type

  Type of the arg's accepted value, which could be one of the basic python types like int or float or str. Please note that the type int could also be used as indicators for bool values, like the above arg --compute_loss. Under these cases, you can pass values 0 or 1 to disable or enable their corresponding functions, like --compute_loss 1 or --compute_loss 0, or simply using the --compute_loss alone to enable it, without other values followed. For example,

    python main.py ... --batch_size 1000 --compute_loss
  

• Arg Type

  We have defined three arg types to distinguish the loading priority of different args, including static, dynamic, and temporary.

  • Args with argtype static can not be changed once after the training is started. When testing the model, the code will not parse these args to overwrite all the saved values. These args consist mainly of unchangeable attributes of the trajectory prediction model, such as the dimension of the features.
  • Args with argtype dynamic can be re-loaded from the terminal to overwrite the saved values. The code will try to first parse inputs from the terminal, and then try to load from the saved JSON file (args.json in the weights folder). These args are often changeable but have different default values for different models, such as the batch size.
  • Args with argtype temporary will not be saved into JSON files. The code will parse these args from the terminal at each run.

• Short Name

  Short name is the alias of args, which usually starts with one -. These short names and full names are fully replaceable, but please note the number of -. For example, --batch_size 1000 equals to the short-name-value -bs 1000, while --bs 1000 is not a valid arg and the code will not parse it.

• Default Value

  Default values of these args. It works for each arg when there are no manaul inputs from the terminal or saved values from the args.json file.