Version 1.6 Released ¶
By: Jon on Dec. 28, 2020, 9:41 p.m.
- Exposing Exp, Log, and Negate nodes
- Fix computation of output dimension
- Fix execution crash when a graph component fails to parse
- Support for 4.26
- SmartEngineAssetStore::LoadObject() and SmartEngineAssetStore::SaveObject() no longer set the serialize binary property of the archive by default. This was initially set to skip properties marked with transient or skip serialization; however, it breaks serialization when new properties are added to a class. Instead, the flag is not set and the transient properties are skipped through other means, but as a result, old data will not load. The old behavior is still available by passing in a flag to LoadObject() and SaveObject().
- Disabling stats when sub frame ticking is enabled in a training harness. This is to avoid a rare crash that can occur. The drawback is that once disabled, it cannot be re-enabled in the same session.
- Limiting the precision of float values to 3 decimal places in networked training harness standard output
- Fix bug causing unchanged assets to be reimported if the file is touched. This happens often when changing branches in git. It may happen to reimport one more time as the file hashes are updated. Note that upon file change, the editor will say it failed to import. This is normal. Looking into ways to suppress this message.
- Adding setting to specify how to initialize matrices
- Adding setting to specify where to check for NaNs
- Exposing Zero() method
- Exposing HasNaN() method