Building a custom seq2seq sentence boundary detection system using Transformers to replace Stanza. This reduces the size of dependencies that need to be packaged in the Snap archive, meaning less to decompress at start. This can reduce Argos Translate start time from 10.8s to 6.8s.
10secs by default are quite slow, do you already have
compression: lzo in your snapcraft.yaml ? that should reduce the default decompression time a lot as well …
I enabled LZO compression but then the Stanza dependency increased in size which made the Snap too large to upload to the Snap store so I had to disable it.
Just out of curiosity, how big is too much for the store?
It doesn’t appear to a hard limit but uploads seem to start failing ~1GB.