v3.0.0

Upgrade actions

Read this before upgrading. Details are in the linked entries below.

1. Retrain any model whose project defines a "No Damage" class. That class
   previously received an output channel and could be predicted at inference;
   it no longer carries one. Predictions from earlier models are not
   meaningful for "No Damage" areas. See The "No Damage" constraint loss let
   the model predict "No Damage".
2. Rename the AZURE_BATCH_REGISTRY_SERVER_URL application setting to
   AZURE_BATCH_REGISTRY_SERVER on the api and queues Function Apps
   (value: the bare login server, e.g. myacr.azurecr.io), and confirm the
   training/imageryprep pool ids are identical across both apps. See Batch
   application settings were out of sync with the code that reads them.
3. Re-run any image layer that failed with NodeNotReady. The log upload
   that makes those failures diagnosable applies only to newly submitted
   tasks. See Losing the Batch compute node failed image layers that had
   already succeeded.
4. "No Damage" must now be the last entry in a project's class list for
   the constraint loss to engage; where it isn't, the loss stays off and the
   reason is logged.

Full release notes for v3.0.0 →

Artifacts

hastegeo wheel	1.0.42 — download

Hosted on haste-binaries, the append-only asset store every requirements.txt pins.