Uncertainty

[AlexDuvalinho]

In this PR, we continue the work done on depfaenet and deup-depfaenet, integrating it into the active learning pipeline.

Aims:
(1) have a trained DepFAENet model fine-tuned on selected adsorbates, to predict the adsorption energies for catalyst-adsorbate pairs generated by GFlowNet
(2) have a deup-dataset which the intermediate representations of the above model, to train a Gaussian Process that measures uncertainty
(3) have a (v0) deup-depfaenet model that predicts uncertainty of the core DepFAENet model

Main changes of this PR are:

• Fix cluster issues ✅
• Update all associated configs and checkpoints relative to deup / depfaenet ✅
• Train a DepFAENet model, without concatenation, with tag0 and with dropout ✅
• Obtain a deup dataset with above model ✅
• Train deup-depfaenet model that predicts the uncertainty ✅
• Ensure that it can be integrated to the active learning repo ✅
• Add a density estimator (q) for deup-faenet
• Add GP or DKL (replacing mc dropout ?)

Information on best checkpoints are included in https://www.notion.so/Coding-Probabilistic-Surrogate-Model-328aece28af948f686fa91a2623d8635?pvs=4

---

From the uncertainty-depfaenet branch:

• Import DepFAENet from disconnected_gnns Alvaro ✅
• Adapt DepFAENet (dropout, etc.) and pre-train it on full data ✅
• Re-train DepFAENet on selected adsorbates ✅
• Create deup-dataset using MC dropout depfaenet ✅
• Create Deup-DepFAENet ✅
• Train deup-depfaenet on deup-dataset ✅
• Active learning script to use deup-dataset ✅
• Fix imports issue ✅
• Allow FAENet Wrapper to keep track of hidden representations for active-learning part ✅

CAREFUL: only store graph-level rep in deup-dataset from now on. Need to change it if we want to perform MC dropout again (related to deup-faenet)

[vict0rsch]

Anyone knows why this disappeared?