Data

We use the trackdat repository to manage data. Refer to seqtrack.data and seqtrack.train.setup_data for more detail.

Note that we store a cached index of each dataset in data_cache_dir. This should be a single location that is shared by all instances of the program. This cache is generated if it does not exist. Caution: This may lead to a race condition if jobs are submitted in parallel (and the cache does not exist).

We support multiple pre-processed versions of each dataset. The default is original. This is the version which is used to produce the cachced index. To use another pre-processed instance of a dataset:

    --preproc=resize_force_360x360 \

It is possible to untar the data at the start of the job. To untar to a fixed location:

    --untar \
    --tmp_data_dir="/tmp/seqtrack/data" \

This is particularly useful for SLURM, to create a local copy of the dataset. It is generally fast to copy a single large tar file over the network, but slow to access each individual image. If you are using SLURM, tmp_data_dir will be overridden in seqtrack/tools/train_work.py.

During development, we often work on a single machine and we do not want to untar the dataset every time we train. Then remove --untar and simply do:

    --data_dir="/tmp/seqtrack/data" \

(Note that you can run the program once with --untar and --tmp_data_dir to set up the dataset.)

Training

To train a model:

python -m seqtrack.tools.train \
    --num_trials=1 \
    --imwidth=360 \
    --imheight=360 \
    --eval_datasets otb_50 \
    --eval_samplers full \
    --max_eval_videos=100 \
    --period_assess=50000 \
    --period_ckpt=10000 \
    --num_steps=100000 \
    --optimizer=adam \
    --lr_init=1e-3 \
    --loglevel=info \
    --verbose_train

Caution: The training results are cached in cache/ by default.

Training with SLURM

Training includes multiple runs with different random seeds. To perform this execution in parallel with SLURM:

python -m seqtrack.tools.train --num_trials=5 --slurm --slurm_flags partition=devel time=1:00:00 cpus-per-task=5 gres=gpu:1

Pre-trained features

To use the siamese model with pre-trained features:

python -m seqtrack.tools.train --model_params '{
        "template_size": 123,
        "search_size": 251,
        "feature_params": {
            "arch": "slim_resnet_v1_50",
            "arch_params": {
                "padding": "VALID",
                "num_blocks": 1
            }
        },
        "feature_model_file": "path/to/resnet_v1_50.ckpt"
    }'

Creating experiments

To create a new experiment directory:

aux/workspace/init-experiment.sh experiments/ name

This will create a directory experiments/yyyy-mm-dd-name/.

This script requires that you have a remote called local. To create this repository:

mkdir ~/projects/seqtrack.git
cd ~/projects/seqtrack.git
git init --bare

cd <repo_dir>
git remote add local ~/projects/seqtrack.git
git push local master

Project structure

Here is one way to structure the project files.

project_root/
    seqtrack/
    dataset_cache/
    experiments/
    seqtrack.git/

torrvision / seqtrack