Logan Hallee

Control tags not in tokenizer

1

Curious how we can prepend control tags for conditional generation. I noticed the control tags are not in the tokenizer, so how could we input them to the model?

Create paper_algorithms.py

2

First pass at the supplemental Algorithms. Looking to comment the tensor shapes at every step.

Hello, Great work! Is it okay to say it is just a standard vanilla MLP block? According to the huggingface implementation there is an additional third linear layer and added...

Hello! I see that this implementation of T5 has ~115 extra ids for masking. Is it possible to tile or pattern these in someway to get the model to generate...

Add MedImg for streaming

8

### Feature request Host the MedImg dataset (similar to Imagenet but for biomedical images). ### Motivation There is a clear need for biomedical image foundation models and large scale biomedical...

Diff Attention out_proj dimension fix

1

Subtle potential issue with `self.out_proj = nn.Linear(embed_dim, embed_dim)` working on `attn.reshape(bsz, tgt_len, self.num_heads * 2 * self.head_dim` in the rare case that `embed_dim != self.num_heads * 2 * self.head_dim` Small...

Sequence head not working as intended?

14

Hello, ESMC obviously has very good general protein representations, but it seems to fail the most basic reconstruction task of returning the input sequence. Is this expected? ```python from esm.models.esmc...

Hello, This is really amazing work. Certain fields, like protein language modeling, is quite reliant on BERT-like models. Is there any interest in a BERT implementation / speed run similar...

Hello, How is PPI inference performed with this package? Best, Logan

Predict without MSAs

1

Hello @jwohlwend and @gcorso, Great work with boltz1x, really impressive stuff. I was curious if it was possible to conduct inference without MSAs, and if you have explored any performance...