Struct ProstT5Translator 

pub struct ProstT5Translator { /* private fields */ }

ProstT5, wrapped for greedy translation.

Why the model sits behind a Mutex, and why it matters more here

T5Runner also holds its model in a Mutex, but there it guards nothing: T5Attention::load sets use_cache: cfg.use_cache && decoder, and an encoder stack loads with decoder = false, so the encoder never caches.

The decoder does. Every decode step appends to a per-layer KV cache, which is what makes incremental generation fast — and what makes a stale cache dangerous: the next translation would concatenate its keys onto the previous one’s and produce a fluent, wrong answer with no error. So translate clears the cache before it starts, and test_repeated_translation_is_deterministic fails if that is ever removed.

Implementations

impl ProstT5Translator

pub fn from_pretrained(model: ProstT5Models, device: Device) -> Result<Self>

Load from the hub at the checkpoint’s published precision (F16).

pub fn from_pretrained_with( model: ProstT5Models, opts: &LoadOptions, ) -> Result<Self>

Load with an explicit device and dtype.

pub fn translate(&self, sequence: &str, direction: Direction) -> Result<String>

Translate sequence between the amino-acid and 3Di alphabets.

Greedy decoding, which is what ProstT5’s own usage does: the mapping is close to deterministic and sampling would only add noise to a structural annotation.

ProstT5’s translation is length-preserving: one output token per input residue. Length, not </s>, is therefore the stopping criterion, and the EOS logit is suppressed at every step — which is exactly what Rostlab’s own inference does by passing min_length == max_length to generate, and what HuggingFace implements as MinLengthLogitsProcessor.

Breaking on </s> instead would silently return a truncated structure for any sequence where the model wanted to stop early, and would disagree with the reference on exactly those cases.

pub fn to_saprot_input(&self, sequence: &str) -> Result<String>

Translate sequence and weave the result back into SaProt’s input form (ferritin-goh.13).

This is the pair that removes the Foldseek C++ dependency from the structure-aware path: ProstT5 predicts the 3Di states that SaProt would otherwise need Foldseek to compute from a real structure.

"MQIFVKTLTGK"  ->  3Di "dvvvvcvvvvd"  ->  "MdQvIvFvVvKcTvLvTvGvKd"

Note what this is and is not: the 3Di states are predicted from sequence, not measured from a structure. That is the point — it works without a structure at all — but a SaProt embedding built this way is not interchangeable with one built from an experimental backbone.

Residues outside SaProt’s 20 become #; see interleave.