philipwastakenwastaken / 02456_project

02456 Deep Learning - Project

Set train or evaluate

python src/main.py session=train python src/main.py session=evaluate

Examples of how to modify other hyperparameters

Set session mode to train. Overwrite learning rate, batch size and number of episodes set in config files. python src/main.py session=train model.learning_rate=0.001 train.batch_size=32 train.num_episodes=100

Change environment python src/main.py session=train environment=mspacman

python src/main.py session=train environment=asterix

Changing wrapper (and model used)

python src/main.py session=train model=crop_model

python src/main.py session=train model=resize_model

python src/main.py session=train model=stretch_mode

NOTE that this changes the input size of the model. This means that models trained with one setting are not compatible with each other.

Weights and Biases

Currently, API key from s174274 is used (set in config file). You can see runs here https://wandb.ai/philipwastaken/02456_project. You can disable WandB if you so chose.

Notes on model_path

src/hparams/model/exp1.yaml contains the field model_path. This is the name of the model you wish to either train or run. ONLY input the name of the model file, NO preceding directories (e.g. models/<model_name_here>. If this is set to an empty string (i.e., ''), a new model is generated and trained from scratch. If you set this parameter to a previously trained model, you can continue training the model (and even use different hyperparameters).

It is important that you set this parameter if you wish to run the model. Otherwise it simply chooses the most recently modified model for convinience.

Changing device

Acceptable device settings are auto, cpu or gpu. auto will prefer gpu if possible and fallback to cpu. If gpu is chosen but not available, the program exits.