AutoML Translation is now generally available. Translations are now faster and quality should be improved or on par with previous versions; individual results will vary depending on your datasets
Share
Services
## Feature
AutoML Translation is now generally available. Translations are now faster and quality should be improved or on par with previous versions; individual results will vary depending on your datasets. In addition, quality improvements can vary across language pairs.
## Change
**Model lifetime** AutoML Models run on a managed platform that is regularly upgraded. AutoML Models need to be re-trained at least every 18 months to ensure consistent performance and stability in translations over time and to get any benefits of the upgrades since original training. Models created before AutoML GA (for example, models created during service Alpha /Beta) need to be re-trained within 18 months of GA to move all assets to the newest version of the platform.
## Feature
AutoML Translation has upgraded its Google NMT base models onto a new training architecture so that it now runs end-to-end on upgraded[TPUs](https://cloud.google.com/tpu/docs/tpus). Going forward, all training using Google NMT as the base model trains and serves on TPUs.
## Change
Customers already using AutoML might experience additional differences between v1 and previous versions of the service:
* Data validation now provides warnings on data import to indicate certain data is not usable for training jobs.
* Data import takes longer in order to support data validation.
* To achieve higher quality on the new architecture, training durations are longer.