Simple_VPR_codebase

This repository serves as a starting point to implement a VPR pipeline. It allows you to train a simple ResNet-18 on the GSV dataset. It relies on the pytorch_metric_learning library.

Run an experiment

You can choose to validate/test on sf_xs or tokyo_xs.

python main.py --exp_name example_run --train_path /path/to/datasets/gsv_xs --val_path /path/to/datasets/sf_xs/val --test_path /path/to/datasets/sf_xs/test

Test on a pretrained model

In every experiment the best model, chosen by evaluating the R@1, and the last model are saved in the ./logs/lightning_logs/<exp_name>/checkpoints/ folder in 2 distinct .ckpt files.

We can use these checkpoint files to test the model against a new test dataset.

To test the pretrained model on new test dataset you can simply add the --ckpt_path arg, indicating the path to a .ckpt file, when executing the main.
This will not trigger the training and the validation phases but will only test the model on the dataset indicated in the --test_path arg.

python main.py --ckpt_path ./logs/lightning_logs/\<exp_name\>/checkpoints/last.ckpt \
  --train_path /content/gsv_xs --val_path /content/sf_xs/val --test_path /content/tokyo_xs/test

Continue a model training

We can use the checkpoints files to resume the training from the last checkpoint:

python main_ckpt.py --ckpt_path ./logs/lightning_logs/<exp_name>/checkpoints/last.ckpt \
  --train_path /content/gsv_xs --val_path /content/sf_xs/val --test_path /content/tokyo_xs/test

Take care about selecting all the same parameters of the initial training in case you want to continue the training.

Select Aggregator Layer and Loss

It is possible to pass specific args to select the Aggregator Layer, the Miner and the Loss we want to use to train the model.

python main.py --train_path /content/gsv_xs --val_path /content/sf_xs/val --test_path /content/sf_xs/test \
  --num_workers 2 --loss_name MultiSimilarityLoss --agg_arch ConvAP

It is possible to choose among multiple types of Aggregator Layers, Miners and Losses, for a complete list check:

• the file ./utils/losses.py for Miners and Losses
• the folder ./models/aggregators/ for the Aggregator Layers

Usage on Colab

We provide the notebook colab_example.ipynb. It shows you how to attach your GDrive file system to Colab, unzip the datasets, install packages and run your first experiment.