search

rksltnl / Deep-Metric-Learning-CVPR16

Main repository for Deep Metric Learning via Lifted Structured Feature Embedding
MIT License
341 stars 122 forks source link

solver files for CUB-200-2011 #11

Closed vadimkantorov closed 7 years ago

vadimkantorov commented 7 years ago

Hi,

I noticed that the pretrained models for CUB have baselr_1E-6 in their names. Have you used learning rate 1e-6 for this dataset? If possible, could you post the parameters (or whole solver files) used for training on CUB?

Thanks, Vadim

rksltnl commented 7 years ago

Here's the solver I used for the birds dataset. net: "/INSTALL_PATH/Deep-Lifting-for-Metric-Learning-CVPR/code/birds/model/train_val_googlenet_finetune_liftedstructsim_softmax_pair_m128_multilabel_embed64.prototxt" test_initialization: false # lr for fine-tuning should be lower than when starting from scratch base_lr: 0.000001 lr_policy: "step" gamma: 0.96 # stepsize should also be lower, as we're closer to being done stepsize: 32000 display: 5 max_iter: 40000 momentum: 0.9 weight_decay: 0.0002 snapshot: 10000 snapshot_prefix: "snapshot_birds_googlenet_finetune_liftedstructsim_softmax_pair_m128_multilabel_embed64_baselr1E6" solver_mode: GPU

Here's the solver I used for the cars dataset. net: "/INSTALL_PATH/Deep-Lifting-for-Metric-Learning-CVPR/code/cars196/model/train_val_googlenet_finetune_liftedstructsim_softmax_pair_m128_multilabel_embed64.prototxt" test_initialization: false # lr for fine-tuning should be lower than when starting from scratch base_lr: 0.0001 lr_policy: "step" gamma: 0.96 # stepsize should also be lower, as we're closer to being done stepsize: 32000 display: 5 max_iter: 40000 momentum: 0.9 weight_decay: 0.0002 snapshot: 10000 snapshot_prefix: "snapshot_cars_googlenet_finetune_liftedstructsim_softmax_pair_m128_multilabel_embed128" solver_mode: GPU