A lightweight trainer that features callbacks and DDP support written with only 615 lines of core code.
Easily
- log intra batch values to a csv as well as a tqdm progress bar,
- save the weights of a model based on an increase or decrease of a logged metric,
- train on multiple GPUs and multiple nodes with
DataDistributedParallel, - write any custom feature you would like with the use of a callback.