VTAB-1K pytorch

Plug & Play implementation of the Visual Task Adaptation 1K Benchmark (VTAB-1K) in pytorch

Setup

Instructions to setup environment/code and data are provided in SETUP_DATA.md and SETUP_ENV.md.

Instructions for training models to reproduce paper results can be found in SETUP_TRAIN.md. This codebase was used for the following papers:

• MIM-Refiner
• Vision-LSTM
• EVA

Evaluation Protocol

The default evaluation protocol first tunes hyperparameters on the validation set and then trains a model on the union of train and validation set. This model is then evaluated on the testset. Each run can use multiple seeds. As this results in a lot of runs (e.g. by using 5 seeds and 5 learning rates, this results in 30 training runs), a evaluation is defined via a pipeline. The pipeline will use a work directory where it will keep track of what runs are queued/running/finished.

To start an evaluation, copy a pipeline to a work directory (by default ./work) and run python main_runner.py. By default it will use ./work as work_area (change via --work_area <WORK_AREA>) and the GPU with index 0 (change via --devices <GPU_INDEX> e.g. --devices 3).

You can easily parallelize the whole pipeline by simply starting multiple runners on different GPUs with the same work_area.

A pipeline with a small model to test if your setup is correctly setup can be found in pipelines/debug/debug.yaml.

Overview of training pipeline

The main_runner.py script will continuously fetch a new yaml file from the work_area and move it into the subfolder work_area/running. It will then process the yaml (which is either training a model, checking if other runs finished to create a summary or generating new yamls). After the yaml is finished, it will write a result into the file and move it to work_area/finished. In the case where a run needs other runs in order to be considered finished (e.g. if hyperparameters are tuned on the validation set, all hyperparameter search runs need to be finished before the evaluation runs on the testset are started), it will instead be moved into work_area/waiting and wait for the other runs to finish. If a run crashes, it will be moved into work_area/crashed. After all runs are finished, a summary is created in work_area/summarized which can then be easily parsed with main_parse_results.py.

If for whatever reasons runs get stuck in work_area/running or crash and are moved into work_area/crashed, you can simply restart them by moving them to work_area.

Results

The results of each run are written into the corresponding yaml. The script main_parse_results.py summarizes the result of a full VTAB-1K pipeline into a single value for each dataset.

SLURM

• Copy sbatch_config_example.yaml as sbatch_config.yaml and adjust the values to your setup.
• Copy template_sbatch_example.sh as template_sbatch.sh and adjust the values to your setup.
• Submit a job via python main_sbatch_runner.py --time 5:00:00

The main_sbatch_runner.py will do the following steps:

• fetch the sbatch_config.yaml
• include the content of sbatch_config.yaml into template_sbatch.sh
• submit the patched template via sbatch

If you want to parallelize the evaluations, we recommend to submit many jobs with only 1 GPU. While main_sbatch_runner.py supports submitting jobs with many GPUs (e.g. --gpus 16 for 16 GPUs), runners will terminate if they are idle for too long which will lead to GPUs idling. If you submit only single GPU jobs, this will not happen as the job is simply terminated if it has to wait for other runs to finish.

If you heavily parallelize the evaluations, chances are that some runs will crash (hardware failures, ...) and will therefore either be moved into the crashed folder or be stuck in the running folder. In this case, simply move the crashed yamls back into the work folder and start the pipeline again.