Running Evaluation - Details

This document describes how to perform evaluation on our DEVIL splits. It covers various details such as using .tar files, environment variables, and Slurm.

Computing Evaluation Features

Note: Evaluation features take up a lot of space (~50GB per DEVIL split x 5 DEVIL splits). Before running the following commands, make sure that <project root>/eval-data points to a disk with sufficient capacity (you can use a symbolic link if necessary).

First, make sure that the desired DEVIL splits corresponding to real videos exist as normal directories in datasets/devil. If not, download and extract them:

for split in ( all bsm-h bsm-l cm-h cm-l  ); do
  python -m src.main.download_devil_splits flickr-$split -e;
done

Then, generate the evaluation features:

for split in ( all bsm-h bsm-l cm-h cm-l  ); do
  source scripts/compute-devil-evaluation-features.sh flickr-$split;
done

OPTIONAL: Archive evaluation features

The evaluation features can be stored in .tar files, which are slightly more NFS-friendly than normal directories. To create them, run the following script:

./scripts/archive-evaluation-features.sh

Downloading Inpainting Results

Previously generated inpainting results can downloaded with download_inpainting_results.py. For instance, the following example downloads CPNet predictions on the fvi-fgd-h split:

python -m src.main.download_inpainting_results cpnet flickr-all fvi-fgd-h

For the list of supported models and DEVIL splits, refer to src/main/download_inpainting_results.py. Results will be saved under the inpainting-results/devil folder, and will overwrite any existing results.

Running Evaluation

NOTE: The commands in this section should be run from the slurm directory.

The submit-evaluate-inpainting.py script runs evaluation metrics on our data. Below is an example of running this script on CPNet predictions:

python submit-evaluate-inpainting.py cpnet flickr-all fvi-fgd-h -m local

More generally, given a call of the following form:

python submit-evaluate-inpainting.py <model> <source split> <mask split> -m local

the script will look for results in <project root>/inpainting-results/devil/<source split>_<mask split>/<model> and compare them to the ground-truth data at <project root>/datasets/devil/<source split> and <project root>/datasets/devil/<mask split>.

If you have predictions for your own model (e.g., at <project root>/inpainting-results/devil/flickr-all_fvi-fgd-h/mymodel), your can use our evaluation script on them.

FID

For some reason, the evaluation script usually fails when FID is computed alongside the other metrics; for this reason, FID is excluded from the default configuration of submit-evaluate-inpainting.py and should be computed in a separate run. The following example computes FID and adds it to the existing CPNet evaluation results:

python submit-evaluate-inpainting.py cpnet flickr-all fvi-fgd-h -m local -a ::include fid ::update

OPTIONAL: Advanced features

If you want to use the .tar files that were optionally generated (i.e., for the DEVIL datasets and evaluation features), use the --use_tar flag:

python submit-evaluate-inpainting.py cpnet flickr-all fvi-fgd-h -m local --use_tar

Environment variables can be passed in as string pairs with the -e flag. The following example runs evaluation with a custom scratch storage location (which is used to temporally store data during evaluation):

python submit-evaluate-inpainting.py cpnet flickr-all fvi-fgd-h -m local -e SCRATCH_ROOT /tmp/$USER

This script also supports Slurm job allocation, which can be done by removing the -m local flag:

python submit-evaluate-inpainting.py cpnet flickr-all fvi-fgd-h

sbatch arguments can be passed in as string pairs with the -s flag. This illustrative example runs evaluation on a custom partition:

python submit-evaluate-inpainting.py cpnet flickr-all fvi-fgd-h -s partition my-partition

If necessary, -s and -e can be used at the same time.

Printing Evaluation Results

The following script writes all quantitative results to inpainting-results-quantitative-summary.tsv as a tab-separated text table:

./scripts/print-quant-results-table.sh