Closed hermancollin closed 8 months ago
@hermancollin I think I will just extend the existing aggregation script in this repository to accept additional (optional) arguments that allows us to create generalist models with only a subset of datasets. But let me know what you think!
Sounds great.
@hermancollin Here is the PR addressing this issue.
To study the effect of aggregating data within a single modality, a model will be trained on all 3 BF datasets and compared to the dedicated BF models. We can see this as an "ablation study" for the aggregation process.
Not sure if we should add this pipeline inside this repo or create another one - @ArthurBoschet thoughts on this?