REPRODUCTION.md

Train

All the training details (hyper-parameters, trained layers, backbones, etc.) strictly follow the original MatConvNet version of NetVLAD and SARE.

the following models are not the original implementation from the paper, in order to accommodate small computation, we use channel dimensions of 64*3 and feature sizes of 128/2 X 128/2. We encourage you to change the channel dimensions (L_DIM, M_DIM, H_DIM) in train.sh to increase the performance.

bash train.sh

The default scripts adopt 4 Tesla V100 GPUs (require ~32G per GPU) for training, where each GPU loads one tuple (anchor, positive(s), negatives).

• In case you want to fasten training, enlarge GPUS for more GPUs, or enlarge the --tuple-size for more tuples on one GPU;
• In case your GPU does not have enough memory (e.g. <32G), reduce --neg-num for fewer positives or negatives in one tuple.

Test

During testing, the python scripts will automatically compute the PCA weights from Pitts30k-train or directly load from local files. Generally, model_best.pth which is selected by validation in the training performs the best.

Set the channel dimensions (L_DIM, M_DIM, H_DIM) in test.sh and run the test script with <MODEL PATH>, e.g. 'logs/saved_models/pitts30k-vgg16/conv5-sare_joint-lr0.001-tuple4/model_best.pth':

bash test.sh <MODEL PATH>

The default scripts adopt 4 Tesla V100 GPUs (require ~32G per GPU) for testing.

• In case you want to fasten training, enlarge GPUS for more GPUs, or enlarge the --test-batch-size for larger batch size on one GPU;
• In case your GPU does not have enough memory (e.g. <32G), reduce --test-batch-size for smaller batch size on one GPU.