Releases: KhawajaAbaid/teras
Releases · KhawajaAbaid/teras
teras v0.3.1
- Add hidden layer to
Classification
andRegression
heads to make training easy out of the box as currently using some backbones directly withClassification
orRegression
head was resulting in weird behavior. - Update
SAINTEmbedding
layer to make it serializable - Some bugs/typos fixes
teras v0.3.0
- Re-written from scratch to fully base it on Keras 3
- Now supports all popular ML backends, namely TensorFlow, JAX and PyTorch
- Docs re-written from scratch (A complete new look, totally didn't yoink from JAX)
- Removed almost all of janky/hacky code
- Removed
LayerFlow API
- Add task independent
Backbone
models (idea taken from KerasCV) - Loads of cleaning and much more!
Teras v0.2.0
- Redesign the whole library
- All the models for classification and regression are now made up of
Keras Functional API
LayerFlow API
now serves as the parent to the defaultparametric API
LayerFlow API
now requires all the sub-layers needed for a model or non-atomic layer during instantiation- This point above get rids of any discrepancies and assumptions made by the
LayerFlow API
by trying to plug customized and default layers together, often resulting in errors - Make models saving and reloading compatible with the
Keras V3
or.keras
format - This redesign is also a preparation step to make this library fully compatible with
Keras Core
and hence backend agnostic
Teras v0.1.1
Bugfixes with some LayerFlow models instantiations
Update dataframe_to_tf_dataset
utility function to handle multi-label datasets
Teras v0.1.0
Hello world! It's the first Teras release!! Lets goooo!!