Download datasets

• Download imdb and wiki "face only" datasets in: https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/

• md5sum

1. imdb_crop.tar 44b7548f288c14397cb7a7bab35ebe14
2. wiki_crop.tar f536eb7f5eae229ae8f286184364b42b

• Download morph2 (MORPH Academic Set) dataset: https://uncw.edu/myuncw/research/innovation-commercialization/technology-portfolio/morph

Create datasets

• create imdb datasets and save as imdb.npz file in ./datasets folder python create_datasets/create_imdbwiki.py --db imdb --output datasets/imdb.npz

• create wiki datasets and save as wiki.npz file in ./datasets folder python create_datasets/create_imdbwiki.py --db imdb --output datasets/wiki.npz

• create morph2 datasets and save as morph2.npz file in ./datasets folder python create_datasets/create_morph.py --output datasets/morph2.npz

• create mutli resolution Morph2 dataset. python create_datasets/create_morph-three_resolution.py --output datasets/morph2_context.npz

Train HSRNet

Using the bash script 'bash_train_hsr_template.sh' to train from scratch and record the logs. This template script trains SepHSR(30,10) with morph2 dataset from scratch with 50 batchsize and 160 epochs and records the log in './records/model_logs' folder.

The arguments in the bash script can be modified:

• nb_kernels = 30 (integer)
• out_channels = 10 (integer)
• hsr_compress = sep (string), other valid value: None, sep, bsep
• db = morph2 (string), other valid value: imdb, wiki, morph2
• batch_size = 50 (integer), [imdb: 128, wiki: 50, morph2: 50] for our experiments.
• nb_epochs = 160 (integer)

$ bash bash_train_hsr_template.sh

Train HSRNet with IMDB > WIKI > MORPH2 (pipeline)

Using the bash script 'bash_train_hsr_pipeline_template.sh' to first train with imdb, wiki, and then morph2.

$ bash bash_train_hsr_pipeline_template.sh

Train HSRNetContext with morph2_context (3-resolutions)

This requires pipeline model's checkpoint after training the wiki dataset.

$ bash bash_train_hsr_context_template.sh