-
- [ ] documenter les responsabilités
- [ ] signatures, règles de collaborations, mentions des sources
- [ ] procédures et politiques
- [ ] documentation
-
As it stands there's no parameter finetuning in #85 or #84.
1. Parameters to be tuned (all models):
- `autoencoder_lr`, `autoencoder_wd`
- `batch_size`
- All adversary parameters. Since we crea…
-
ghost updated
2 years ago
-
A new LINCS submission experienced timeouts in prod and also shows very slow ingest timings under test in dev environment.
The core fact keywords ETL query seems to be the standout query which is r…
-
Trying to record vocab usage for a submission with large synonym lists, for example a compound in recent LINCS submissions, hits an implementation limit and fails.
To remove this limit, we need to …
-
- Gene-gene interactions (GGIs) from mRNA co-expression:
GGIs are computed from thousands of randomly selected RNA-seq samples processed uniformly with the ARCHS4 pipeline. The top 10 genes …
-
see issue #15 for details on schema
ghost updated
2 years ago
-
The Vanilla Model with ~6M parameters is outperforming most of our other model, however these other models have maximum ~2M parameters.
It might be worth scheduling some runs with very large models…
-
Generate 3 new Trapnell splits, one for each of the 3 pathways (epigenetic, cell cycle, tyrosine kinase).
In each split, 10 drugs are left out.
For LINCS we need just one new split, where all 30 …
-
Should go into the `analyze_lincs_all_embeddings_hparam.ipynb`. We use this to show that the counterfactual prediction is working correctly and that the information coming from the embeddings is actua…