Starcraft Super Resolution

Project to upscale Starcraft images using Super Resolution techniques.

Results

In each image we see the results of applying baseline bilinear interpolation, the trained super resolution work at different epochs, the networks final result at the final epoch, and the ground truth high resolution image. The input to all these models was the ground truth image scaled down 3x. The model here was created using a layers param of "128,9 256,1 19".

Tensorboard Visualizations

When using run_aws.py we can use the hosted Tensorboard app to observe the error plots and see how results evolve over epochs.

---

Requirements

• Python 3.8, install requirements via

pip install -r starcraft_super_resolution/requirements.txt
pip install -e starcraft_super_resolution/

• AWS CLI tool (see https://docs.aws.amazon.com/cli/latest/userguide/install-cliv2.html)

---

Model Settings

The design of the model used is specified via the command line python3 /starcraft_super_resolution/run.py --name {name} --dataset {dataset} --layers {layers} --scaling {scaling} --epochs {epochs} --batch_size {batch_size} --epochs_per_save {epochs_per_save} --overwrite {overwrite} where

• name: The name of your model, used when saving data and displaying via tensorboard.
• dataset: Which dataset to use in training, "units" or "frames".
• layers: String that specifies the model architecture. Example: "layer_1_kernel_size,layer_1_num_filters * layer_2_kernel_size,layer_2_num_filters final_layer_kernel_size". The number of filters is not specifies for * the final layer since this is specified by whether the final image should be greyscale or not.
• scaling: To what scale do we want to upres an image.
• epochs: Number of epochs to train on.
• batch_size: Batch Size used in training.
• epochs_per_save: How many epochs to train on between saving images/models.
• greyscale: True or False, specifies if color images will be used or not.
• overwrite: Whether to overwrite previous model data.

Execute

Local

• Decide on model settings and run

  python3 /starcraft_super_resolution/run.py --name {name} --dataset {dataset} --layers {layers} --scaling {scaling} --epochs {epochs} --batch_size {batch_size} --epochs_per_save {epochs_per_save} --overwrite {overwrite}

  for a specific simple example

  python3 ./starcraft_super_resolution/run.py --name test_model --dataset units --layers 12,3 5 --scaling 3 --epochs 10 --batch_size 32 --epochs_per_save 2 --overwrite False

• To view results launch tensorboard --logdir starcraft_super_resolution/upres/data/output --max_reload_threads 1.

EC2

• Set your AWS variables in a .env file in the base folder. See .env_sample.
• Configure your boto access keys. https://aws.amazon.com/developers/getting-started/python/.
• Edit the Model Settings variables in run_aws.py.
• python3 starcraft_super_resolution/launch.py.
• Output gives SSH command & link to hosted tensorboard instance.

---

Sources

https://arxiv.org/pdf/1501.00092.pdf
https://arxiv.org/pdf/1808.03344.pdf