Loss Landscapes are All You Need: Neural Network Generalization Can Be Explained Without the Implicit Bias of Gradient Descent
This repository trains large number of models in parallel with non-gradient based optimizers.
To set up the environment, you could use conda with conda env create -f environment.yml
All scripts for reproducing the tables in the paper "Loss Landscapes are All You Need: Neural Network Generalization Can Be Explained Without the Implicit Bias of Gradient Descent") - ICLR 2023 can be found in ./scripts.
train_distributed.py trains models in parallel on different host and then save the resulting metrics in a single shared sqllite database.