bclarkson-code / Tricycle

Autograd to GPT-2 completely from scratch

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Code versions as a hyperparameter

bclarkson-code opened this issue · comments

It is common practice to record model hyperparameters when experimenting. But I think this misses a lot of extra information. To experiment properly, I think that one should also record specific implementations of different components. you could simply record the git commit that was used, but I think it would be more helpful to record the specific version of each component used.

I think that a good place to start would be recording a hash (or similar) for each layer in the model.

Closing for now as it is off topic