Train a neural network to predict cropped out rectanglular area of an image.
This repo has been dug out from earlier days for showcasing. The original data is not mine, and the showcased results in exploration.ipynb are from the FiftyOne Dataset Zoo.
├── LICENSE
├── README.md <- Toplevel README
├── data
│ ├── external <- Testset and example-testset.
│ ├── interim <- Selected data.
│ ├── processed <- Images resized s.t. the largest dimension is 100.
│ └── raw <- The original, immutable data dump.
│
├── models <- Trained and serialized models.
│
├── architectures.py <- Architectures implemented in PyTorch.
├── datasets.py <- Dataset classes.
├── functional.py <- Useful functions used in the project.
├── image_resizer.py <- Script used to resize images.
├── filter_dataset.py <- Script to filter data/raw to data/interrim
├── config.json <- Network- and training-configuration.
├── train.py <- Training script.
To reproduce results, run generate_predictions.py, with the final model being models/final.pt (only works if data/external/challenge_testset.pkl) is available.