You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
We are consolidating KerasNLP and KerasCV into a new KerasHub package.
We will be renaming this keras-nlp GitHub repository to keras-hub.
All existing usages will continue to work!
You can keep running pip install keras-nlp & import keras_nlp—nothing will break. Same with keras_cv.
What's happening
Popular pretrained models are frequently becoming multi-modal. In the text domain, chat LLMs are adding support for image and audio inputs and outputs. In the vision domain, using text encoders is common for a wide range of tasks from image segmentation to image generation.
We do not believe any particular division of pretrained models will help the Keras ecosystem going forward. Distinctions between model architectures (transformers vs convnets vs diffision models) become fuzzy, as do divisions between modalities.
By consolidating to a single library that focuses on easy-to-use, pretrained architectures and weights, we can better deliver a set of features that apply to all pretrained models—easy model publishing and sharing, PEFT and quantization support, scaled up muli-host training.
To that end, we are consolidating KerasNLP and KerasCV into a KerasHub package.
The plan
We will start start by simply renaming the KerasNLP library to KerasHub. KerasNLP already supports multiple modalities with models like PaliGemma, as well as many features (easy model publishing), we would like to bring to CV models.
This repo will be renamed to keras-hub. Library symbols will be renamed from keras_nlp to keras_hub.
We will keep a keras_nlp package with all the old imports. This change will be fully backward compatible.
To move from keras_nlp to keras_hub, you should be able to simply find and replace all instances of keras_nlp with keras_hub in code.
keras_cv models will move to keras_hub on an ongoing bases. You can expect some usage changes to CV models, as we make sure we deliver a consistent UX across modalities.
When is this happening
The repository is now a preview for the upcoming KerasHub release, but we have not yet made an official release of the keras-hub package. If you would like to try things out as we build, you can do so by trying our nightly package: pip install keras-hub-nightly.
We are tentatively aiming for a mid October release of KerasHub.
Feedback and help
We would appreciate any feedback from the community on this! Please feel free to use this issue to send us thoughts.
Update: we have now renamed the repository and code to keras-hub, but we have not yet released the KerasHub package.
If you would like to try things out as we build, use the nightly package: pip install keras-hub-nightly. If you want the backwards compatible keras_nlp imports, you can try that on nightly too: pip install keras-nlp-nightly.
Lastly, if you were building the package from source, and want to keep things working with the old imports, you will need to install as follows:
tl;dr
keras-nlp
GitHub repository tokeras-hub
.pip install keras-nlp
&import keras_nlp
—nothing will break. Same withkeras_cv
.What's happening
Popular pretrained models are frequently becoming multi-modal. In the text domain, chat LLMs are adding support for image and audio inputs and outputs. In the vision domain, using text encoders is common for a wide range of tasks from image segmentation to image generation.
We do not believe any particular division of pretrained models will help the Keras ecosystem going forward. Distinctions between model architectures (transformers vs convnets vs diffision models) become fuzzy, as do divisions between modalities.
By consolidating to a single library that focuses on easy-to-use, pretrained architectures and weights, we can better deliver a set of features that apply to all pretrained models—easy model publishing and sharing, PEFT and quantization support, scaled up muli-host training.
To that end, we are consolidating KerasNLP and KerasCV into a KerasHub package.
The plan
We will start start by simply renaming the KerasNLP library to KerasHub. KerasNLP already supports multiple modalities with models like
PaliGemma
, as well as many features (easy model publishing), we would like to bring to CV models.keras-hub
. Library symbols will be renamed fromkeras_nlp
tokeras_hub
.keras_nlp
package with all the old imports. This change will be fully backward compatible.keras_nlp
tokeras_hub
, you should be able to simply find and replace all instances ofkeras_nlp
withkeras_hub
in code.keras_cv
models will move tokeras_hub
on an ongoing bases. You can expect some usage changes to CV models, as we make sure we deliver a consistent UX across modalities.When is this happening
The repository is now a preview for the upcoming KerasHub release, but we have not yet made an official release of the
keras-hub
package. If you would like to try things out as we build, you can do so by trying our nightly package:pip install keras-hub-nightly
.We are tentatively aiming for a mid October release of KerasHub.
Feedback and help
We would appreciate any feedback from the community on this! Please feel free to use this issue to send us thoughts.
We have a number of issues related to the port with the contributions welcome tag.
The text was updated successfully, but these errors were encountered: