In general, the project is managed through Makefile
.
You can use it to build (with command make docker-build
) and run Docker images and do other things. See the comments of individual Makefile goals in the file for details.
Also, at the top of the Makefile
, you can find which environment variables are required by the Makefile
.
torch == 1.3.1 torchvision == 0.4.2
For new dataset prepare annotation files in data/captions
and data/image_splits
.
Then build vocabulary by running:
python3 build_vocab.py --data_set shoes
Pre-train image and text encoders:
CUDA_VISIBLE_DEVICES=2 python3.7 train.py --data_set=shoes --img_root=/home/datasets/style_search-dialog/images
Train SynthTriplet GAN with triplet loss (
CUDA_VISIBLE_DEVICES=2 python3.7 train-c-w.py --loss_type=cgan --img_root=../datasets/images --data_set=shoes
Train SynthTriplet GAN with triplet loss and gradient penalty (
CUDA_VISIBLE_DEVICES=2 python3.7 train-c-w.py --loss_type=wgan --img_root=../datasets/images --data_set=shoes
Train SynthTriplet GAN with triplet loss, text-adaptive discriminator and reconstruction loss (
python3.7 train-t.py --img_root=../datasets/images --data_set=shoes --trainclasses_file=shoes.txt --a_t=10 --ranker=encoder --freeze=gan