This project adapts a ResNet50 model architecture to perform pose estimation on several series of satellite images (both real and synthetic).
For more information, please see the SPARK Challenge ( https://cvi2.uni.lu/spark2022/ ) organized as part of the AI4Space workshop, in conjunction with the European Conference on Computer Vision (ECCV 2022).
See requirements.txt and make sure to also install cudatoolkit if you plan to run with a GPU.
If running on the UniLux HPC, please see notes in hpc_setup.sh
To train the model:
python run_train_model.py
To train the model and output all print statements to a local file:
python run_train_model.py > LOGS.txt
To load and test the saved, pre-trained model on the test_real dataset:
python run_train_model.py -l -tr
To load and test the saved, pre-trained model on the test_synthetic dataset:
python run_train_model.py -l -ts
After training, the trained model and optimizer are stored in results/model+optimizer.pth
Model predictions are stored as CSV files in predictions/
Loss function plots are stored in figures/
PyTorch Resnet implementation taken from:
https://github.com/aladdinpersson/Machine-Learning-Collection/blob/master/ML/Pytorch/CNN_architectures/pytorch_resnet.py
Helpful tutorials:
https://pytorch.org/tutorials/beginner/blitz/neural_networks_tutorial.html
https://pytorch.org/tutorials/beginner/blitz/cifar10_tutorial.html
Creating and validating a custom PyTorch Dataloader:
https://pytorch.org/tutorials/recipes/recipes/custom_dataset_transforms_loader.html
https://glassboxmedicine.com/2022/01/21/building-custom-image-data-sets-in-pytorch-tutorial-with-code/
Reference: PyTorch complete example - CNN with MNIST dataset:
https://nextjournal.com/gkoehler/pytorch-mnist
Loss Functions:
https://heartbeat.comet.ml/5-regression-loss-functions-all-machine-learners-should-know-4fb140e9d4b0
https://neptune.ai/blog/pytorch-loss-functions#:~:text=Broadly%20speaking