Working example of Huggingface style hyper parameter search with ray-tune and Trainer
.
The algorithm selected is Population Based Training
.
Existing resources on integrating hyper-parameter search
with Huggingface
is limited and buggy. PR
s are welcome.
conda create -n hyper python=3.8.13 -y
conda activate hyper
pip install -r requirements.txt
- Inadequate inspection of restart process. Undesired behavior might occur with
optimizer
,scheduler
anddata_loader
. ray
does not keep the best performing checkpoint over the training course. This implies that only checkpoints that complete their whole designated training schedule would be available.