Tweets
2022-11-06T03:26:17+00:00 | 🔗
RT @gregeganSF: I’m just waiting for someone to train an LLM exclusively on a religion’s sacred texts, and then market it as a way to conve…
2022-11-06T03:25:26+00:00 | 🔗
@kulesatony Find a good review article and then look at everything citing it and then pay the authors a consulting fee? Or also search the target on ChEMBL for good entry points for lit review
2022-11-06T03:11:40+00:00 | 🔗
RT @ritholtz: The collection at the Met is an endless embarrassment of riches https://t.co/yRtZnKWjRf
2022-11-05T21:20:10+00:00 | 🔗
Loss graph! Blue line is after adding much more data, the second set is 2D coordinates computed by various software and the first set is 2D coordinates of hand-drawn/curated compounds. https://t.co/GwBDCFSOx0
2022-11-04T22:18:42+00:00 | 🔗
@vctr_machine NGL, it was just a random sample from the USPTO Dataset of MOL files… but it looks nice and is really pretty for a symmetric long-ranged graph transformer test!
2022-11-04T20:49:22+00:00 | 🔗
@andrewwhite01 Hmmm well it doesn’t work yet, but I think so. I think you can handle it like how image generation handles inpainting by fixing all atoms except one like painting in only a small region with the image fixed everywhere else
2022-11-04T20:44:59+00:00 | 🔗
@andrewwhite01 Still working it out! Currently I’m adding rotational data augmentation and scaling coordinates to -1,1 only. One interesting challenge is that bond distances may need to be larger or smaller if we want to fit it to a fixed canvas.
2022-11-04T17:56:51+00:00 | 🔗
Fixed the sampling algorithm (I think), so now they only ~ slightly explode ~ but that might just be because of geometric properties or something https://t.co/W4QmpYdRtK
2022-11-04T14:54:46+00:00 | 🔗
Getting started on a side project: diffusion models for 2D depiction of compounds! I got a small script going that does the forward process and then applies the reverse process using the true epsilon! #AI #Chemistry #diffusion #RDKit https://t.co/VFDQXZOU7v
