Awesome

pretrain_vg : Model pretraining for Visual Geo-localization

This repository is part of the framework proposed in our work "Deep Visual Geo-localization". This module can be used to pretrain the backbones on:

• Google Landmark v2 (GLDv2)
• Places365

Usage

Training on GLDv2

For GLDv2, the backbone used in our work were trained for 30 epochs on the full GLDv2 training set. To launch the training procedure use :

python3 train_gldv2.py [-h] [--device {cuda,cpu}] [--exp_name EXP_NAME]
                    [--gldv2_csv GLDV2_CSV]
                    [--data_path DATA_PATH]
                    [--train_batch_size TRAIN_BATCH_SIZE]
                    [--seed SEED]
                    [--num_workers NUM_WORKERS]
                    [--resize_shape RESIZE_SHAPE RESIZE_SHAPE]
                    [--loss_module {arcface,}] [--arcface_s ARCFACE_S]
                    [--arcface_margin ARCFACE_MARGIN]
                    [--arcface_ls_eps ARCFACE_LS_EPS]
                    [--epochs_num EPOCHS_NUM] [--patience PATIENCE]
                    [--lr LR] [--optim {adam,sgd}]
                    [--arch {vgg16,r18,r50,r101}] [--resume RESUME]

For more details use python3 train_gldv2.py -h

Training on Places365

For Places365, the backbone used in our work were trained until convergence using early stopping with patience equals to 5.

python3 train_places.py [-h] [--device {cuda,cpu}] [--exp_name EXP_NAME]
                       [--data_path DATA_PATH]
                       [--train_batch_size TRAIN_BATCH_SIZE]
                       [--eval_batch_size EVAL_BATCH_SIZE] [--seed SEED]
                       [--num_workers NUM_WORKERS]
                       [--resize_shape RESIZE_SHAPE RESIZE_SHAPE]
                       [--epochs_num EPOCHS_NUM] [--patience PATIENCE]
                       [--lr LR] [--optim {adam,sgd}]
                       [--arch {vgg16,r18,r50,r101}] [--resume RESUME]  

For more details use python3 train_places.py -h