This code is built based on https://github.com/Nanne/pytorch-NetVlad
run
python main.py --mode=train --beta=0.0 --lr=0.00001 --batchSize=2 --atten --rect_atten --p_margin=0.5 --loss=impr_triplet --pooling=atten_vlad --mul=3mul --relu=softplus --add_relu --nEpochs=60 --cacheRefreshRate=1000 --margin=0.1 --evalEvery=1 --optim=ADAM --arch=self_define --num_clusters=64 --nGPU=1 --random_crop --dataset=mapillary
run
python main.py --mode=test --ckpt=best --split=test --resume=./runs/Jun24_05-55-35_-1_3mul --batchSize=2 --atten --rect_atten --p_margin=0.5 --loss=impr_triplet --pooling=atten_vlad --mul=3mul --relu=softplus --add_relu --margin=0.1 --arch=self_define --num_clusters=64 --dataset=mapillary
to evaluate on beeldbank dataset, modify --split=beelbank
run
python mkmmd_main.py --mode=train --loss=impr_triplet --beta=0.0 --nGPU=1 --lr=0.00001 --alpha=0.99 --batchSize=2 --mul=3muk --DA --atten --rect_atten --pooling=atten_vlad --relu=softplus --add_relu --nEpochs=60 --cacheRefreshRate=1000 --margin=0.1 --evalEvery=1 --optim=ADAM --arch=self_define --num_clusters=64 --random_crop --dataset=mapillary
the labels for cross domain evalutaion is in positives.npy
for each line(each beelbank img), it contains the index of the positives in mapillary domain.
The root_dir to dataset is ../data