Keras main concern tf.keras, while continuing to support Theano / CNTK
This version comes with many API changes so that the back-end multi Keras API and TensorFlow advanced API tf.keras "synchronization." However, some TensorFlow 2.0 features are not supported. This is the reason the team recommends that developers in their Keras codes will switch to tf.keras 2.0 TensorFlow.
Migrating to tf.keras will enable developers to quickly perform such access, TPU training as well as better integration between the lower TensorFlow advanced concepts such as Layer and the Model.
After this release, the team plans to focus on further development of tf.keras. "Development will focus on the future. We will continue to maintain multiple back-end Keras in the next six months, but we will only merge bug fixes. API changes will not be transplanted," the team wrote.
To make it easier for the community to contribute to the development of Keras, the team will develop tf.keras in keras-team / keras independent GitHub repository.
The following are some of the API Keras 2.3.0 update:
add_metric method is added to Layer / Model, which is similar to add_loss method, but for the indexes.
Keras 2.3.0 introduces several missing class-based, including MeanSquaredError, MeanAbsoluteError, BinaryCrossentropy, Hinge like. With this update, you can be parameterized constructor argument is missing.
We added a lot of class-based metrics, including Accuracy, MeanSquaredError, Hinge, FalsePositives, BinaryAccuracy and so on. This metric can be updated so that stateful and parameterizing the constructor argument.
train_on_batch and test_on_batch method now has a new parameter named resent_metrics of. You can set this parameter to True, in order to maintain the state of different batches of metrics in the training or assessment cycle written in lower level.
model.reset_metrics () method is added to the Model, to clear state metric of the epoch in the preparation of lower-level training or evaluation cycle.
With the API changes, Keras 2.3.0 contains some significant changes. In this version, not recommended batch_size, write_grads, embeddings_freq and embeddings_layer_names, and therefore will be ignored when used in conjunction with TensorFlow 2.0. It will now be reporting metrics and losses depending on the exact name of the user specified. In addition, repeated activation of the default change from sigmoid RNN hard_sigmoid all layers.
For more information please see linux www.linuxprobe.com
View the official announcement , understand Keras 2.3.0 What else is news.