An online, BPTT-free method for learning and using synthetic gradients.
Beyond the standard python libraries, you will need Pytorch (we use version 1.7.0, but later should work) to define the neural network models and as well as ignite (version 0.4.2) which wraps the training regime. Sacred (version 0.7.4) is used to record experiment details.
Scripts to run the tasks presented in the paper can be found in the /scripts folder. For a given script:
- Define model/training/dataset hyperparameters*. These can be changed in the respective file in the /configs folder.
- Define the OBSERVE variable (preset to False). If true, then experiment results will be saved in a generated /sims folder.
- Run the script.