David Thrower
David Thrower
One related comment is that a more global fix that may be worth considering is replacing https://github.com/kubeflow/katib/blob/master/sdk/python/v1beta1/kubeflow/katib/models/* with Pydantic models representing the same schemas. It could remove a lot of...
Aberrant result in the workflow regression-example-ames-no-preproc-val-set.py (likely 10+ st dev above the mean). This needs investigated.
## TLDR: This may be a structural issue: This is a first pass answer. There's no promise this is the reason in your case, but I found this: The cl_khr_fp16...
## @tilakrayal, I suspect it is possible this may **not** be a bug. ... but I think the error handling message associated with this situation could be improved for clarity,...
Branch for R and D only. Do not merge.
## Do not merge - Counterintuitively, even though the mask is not being passed to the embedding from the tokenizer, setting the parameter zero_mask to False on the Embedding layer...
Proof of concept successful. working on optimizations...
Persistance of models (Model > file > reconstituted model serialization) fixed via #205 #206 Translation to PyTorch: (Made moot by #205 #206 )
# After 3 rounds of optimization and parameter elimination: ## Best parameter ranges for 75 samples at sequence_len 40 (King James Bible) ## Results: - Perplexity: 4.21500301361084
# Do not merge: Proof of concept still being optimized. Downstream work on branch: #220