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.
NB: if you are using Colab, skip this section
The following script:
python download_datasets.py
allows you to download GSV_xs, SF_xs, tokyo_xs, which are reduced version of the GSVCities, SF-XL, Tokyo247 datasets respectively.
NB: if you are using Colab, skip this section
You can install the required packages by running
pip install -r requirements.txt
You can choose to validate/test on sf_xs or tokyo_xs.
python main.py --train_path /path/to/datasets/gsv_xs --val_path /path/to/datasets/tokyo_xs/test --test_path /path/to/datasets/tokyo_xs/test --exp_name expname
The code will save the best (according to validation score) and last models. If you your experiment dies and you want to resume from where you left off, you can simply run you experiment passing the argument --checkpoint model_path. You can find the model checkpoints under logs/lightning_logs/exp_name/checkpoints
The code will log everything under the directory logs/lightning_logs/exp_name. You will find the models under checkpoints.
All the textual outputs generated by the code is saved in 2 files, namely logs/lightning_logs/exp_name/debug.log and logs/lightning_logs/exp_name/info.log, where typically info.log contains only relevant info whereas debug.log is a superset of it and contains additional (typically useless) prints.
If you want to add any prints to the code, you can do so by using the functions logging.debug of logging.info.
Finally, in this directory you will see some binary files generated by tensorboard, that you can use with the proper library.
Once you install tensorboard via pip (check the documentation on how to do it), you can download to your local machine the logs directory and inspect the logs using tensorboard --logdir logs/lightning_logs. It will launch a webserver running on localhost:6006, that you can inspect using your browser
Once you have trained your models, you can run an inference only step using the eval.py script, and passing the --checkpoint argument to specify the model checkpoint to load
python eval.py --checkpoint logs/lightning_logs/exp_name/c heckpoints/_epoch(01)_R@1[30.4762]_R@5[49.2063].ckpt --train_path data/gsv_xs --val_path data/tokyo_xs/test --test_path data/tokyo_xs/test --exp_name test_model
We provide you with 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.
NB: BEFORE running this notebook, you must copy the datasets zip into your GDrive. You can use the link that we provided and simply click 'create a copy'. Make sure that you have enough space (roughly 8 GBs)
NB^2: you can ignore the dataset robotcar_one_every_2m.