gematria
gematria copied to clipboard
google
Machine learning for machine code.
We cannot make the dataclass frozen when doing this, but this enables TF to automatically find trainable variables within GraphNetworkLayer objects which means we can get rid of the somewhat...
tf_slim is very much idiomatically TF1, to the point that I'm not sure it works when we are executing eagerly/utilizing tf.function. We should replace usages of it (mostly continuous evaluation...
Migrate `training.partially_restore_from_checkpoint` to the new TF APIs. It may or may not work currently as it has not been thoroughly tested and the new APIs natively support most of the...
Bumps the pip group with 1 update: [werkzeug](https://github.com/pallets/werkzeug). Updates `werkzeug` from 3.0.4 to 3.0.6 Release notes Sourced from werkzeug's releases. 3.0.6 This is the Werkzeug 3.0.6 security fix release, which...
![dependabot[bot] avatar](https://avatars.githubusercontent.com/in/29110?v=4 "dependabot[bot] avatar"){kind=avatar}
We should probably make `GraphNetworkLayer` subclass from `tf.Module` otherwise we have to use the hacky `_get_trainable_variables` that also looks through the graph network layers as the top level `tf.Module` in...
Currently gnn_model_base and a couple other subclasses have a lot of duplicated code between their `init` functions and their _forward` functions. We should be able to figure out a better...
After the conversion to TF2 we currently only support saving by an epoch count rather than saving based on time. Saving based on time can be a lot nicer in...
We need to support saved model export to get tflite models for inference. We should probably just support this similar to the old export graph def functionality in `main_function`
When we were looking through the model_base changes we were trying to find the original motivation for model_base._clip_if_not_none and could not finding anything compelling. We should investigate further to see...