The code is heavily based on the PyTorch implementation of CycleGAN-pix2pix.
Coming soon...
Take UTKFace for example,
DATASET=UTK
ln -s path/to/dataset/folder/that/contains/all/images $(pwd)/datasets/${DATASET}
ls datasets/${DATASET} > ${DATASET}_all.txt
This will create ${DATASET}_train.txt, ${DATASET}_val.txt, and ${DATASET}_test.txt under ./sourcefiles/, each of which contains 10000, 1000, and 10000 filenames respectively.
Then sample 1000 images from train and val for visualization.
# split dataset
python scripts/split_dataset.py --input sourcefiles/${DATASET}_all.txt --output sourcefiles/${DATASET} --num_samples 10000 1000 10000
# sample a subset for visualization
python scripts/sample_sourcefile.py --input sourcefiles/${DATASET}_train.txt sourcefiles/${DATASET}_val.txt --output sourcefiles/${DATASET}_sample.txt --num_samples 1000
Generate pairs for train and val,
# generate pairs
python scripts/generate_pairs.py --input sourcefiles/${DATASET}_train.txt --output sourcefiles/${DATASET}_pairs_train.txt --num_pairs 20000 --margin 10
python scripts/generate_pairs.py --input sourcefiles/${DATASET}_val.txt --output sourcefiles/${DATASET}_pairs_val.txt --num_pairs 1000 --margin 10
# extract different pairs for training PC-GAN
python scripts/reweight_sourcefile.py --input sourcefiles/${DATASET}_pairs_train.txt --output sourcefiles/${DATASET}_diff_pairs_train.txt --sample_weight 1 0 1
Open a new terminal to start a visdom server with python -m visdom.server.
Train a basic CNN Elo rating network,
python siamese.py \
--dataroot datasets/${DATASET} \
--datafile sourcefiles/${DATASET}_pairs_train.txt \
--dataroot_val datasets/${DATASET} \
--datafile_val sourcefiles/${DATASET}_pairs_val.txt \
--datafile_emb sourcefiles/${DATASET}_sample.txt \
--mode train \
--name elo_${DATASET}_cnn
Train a Bayesian Elo rating network,
python siamese.py \
--dataroot datasets/${DATASET} \
--datafile sourcefiles/${DATASET}_pairs_train.txt \
--dataroot_val datasets/${DATASET} \
--datafile_val sourcefiles/${DATASET}_pairs_val.txt \
--datafile_emb sourcefiles/${DATASET}_sample.txt \
--mode train \
--name elo_${DATASET}_bnn \
--noisy true \
--bayesian true \
--bnn_dropout 0.2
Get the mean (embedding_mean), std (embedding_std), and quantization clusters (embedding_bins) of the ratings (or 'embeddings' called in the code) using get_embedding_cluster.m in ./scripts/. In the following commands, replace the corresponding argument values with the computed values.
To train a basic PC-GAN with a basic encoder,
python train.py \
--model wsgan_emb \
--dataroot datasets/${DATASET} \
--sourcefile_A sourcefiles/${DATASET}_diff_pairs_train.txt \
--name emb_${DATASET}_cnn \
--pretrained_model_path_E checkpoints/elo_${DATASET}_cnn/latest_net.pth \
--embedding_mean 0 \
--embedding_std 1 \
--embedding_bins "[-2, -1, 0, 1, 2]"
To train a PC-GAN with a Bayesian encoder,
python train.py \
--model wsgan_emb \
--dataroot datasets/${DATASET} \
--sourcefile_A sourcefiles/${DATASET}_diff_pairs_train.txt \
--name emb_${DATASET}_bnn \
--noisy_var_type ae \
--noisy true \
--bayesian true \
--bnn_dropout 0.2 \
--pretrained_model_path_E checkpoints/elo_${DATASET}_bnn/latest_net.pth \
--embedding_mean 0 \
--embedding_std 1 \
--embedding_bins "[-2, -1, 0, 1, 2]"
