Ablation studies? Is anyone using google fire or other methods?

@jeremy mentioned that he and @sebastianruder used google fire for their ablation studies in the ULMFiT paper, and I’m planning on doing something similar for a model and paper I’m working on. I’ve taken a look at the repo and the corresponding scripts, but I don’t see the bash scripts for actually running the ablation, storing the results of the training along the way, etc.

I’m wondering if anyone has done ablation in this way (or another way) and has some bash scripts that they can share. I can probably figure it out in the end, but it would be nice to build off of something rather than starting from scratch.

Thanks in advance!
Even

2 Likes

Because I still spent a lot of time tweaking the parameters in the end and trying different settings, for efficiency and ease of use I just used a couple of very simple bash scripts with different configurations that I would modify ad hoc. Not the best way to do things, but useful for experimentation.

A bash script might be as simple as:

python -u kaggle/train_tri_wt.py wikitext-103_2_vanilla 0 --lr=1e-3 --cl=6 --preload False > data/logs/pretrain_wikitext-103_vanilla
3 Likes