After the 1.9.0-rc0 release early last month, the TensorFlow team is out with another update 1.9.0-rc2, unveiling major features and updates. The new release includes latest improvements, bug fixes, and other changes.
Let’s have a look at the noteworthy features in TensorFlow 1.9.0-rc2:
Key features and improvements
- Docs for tf.keras namely the new Keras-based ‘get started’ page and the programmers guide page have been updated.
- Layers tf.keras.layers.CuDNNGRU and tf.keras.layers.CuDNNLSTM have been added.
- Python interface for TFLite Optimizing Converter is expanded. Command line interface (AKA: toco, tflite_convert) is added in the standard pip installation again.
- Data loading and text processing has improved with:
Tf.decode_compressed, tf.string_strip, and tf.strings.regex_full_match
- Headers used for custom apps have moved from site-packages/external into site-packages/tensorflow/include/external.
- On opening empty variable scopes; replace variable_scope(”, …) by variable_scope(tf.get_variable_scope(), …).
- tfe.Network has now been deprecated. You can now inherit from tf.keras.Model.
- The layered variable names have changed for the conditions mentioned below:
- using tf.keras.layers with custom variable scopes
- using tf.layers in a subclass of tf.keras.Model class.
- In tf.data, DatasetBase::DebugString() method is now const. The tf.contrib.data.sample_from_datasets() API for randomly sampling from multiple datasets has now been added.
- In tf.contrib, tf.contrib.data.choose_from_datasets() is added and tf.contrib.data.make_csv_dataset() will now support line breaks in quoted strings. From make_csv_dataset, two arguments were removed.The tf.contrib.framework.zero_initializer supports ResourceVariable. “Constrained_optimization” is added to tensorflow/contrib.
- GCS Configuration Ops has been added.
- The signature of Makelterator has been changed to enable propagation of error status.
- The bug in tf.reduce_prod gradient has been fixed for complex dtypes.
- Benchmark for tf.scan has been updated in order to match ranges across eager and graph modes.
- Optional args argument added to Dataset.from_generator().
- Ids in nn.embedding_lookup_sparse have been made unique which helps reduce RPC calls that are made for looking up the embeddings in case there are repeated ids in the batch.
- tf.train.Checkpoint is added for reading/writing object-based checkpoints.
To get more information on the new updates and features in the latest TensorFlow 1.9.0-rc2 release, check out the official release notes.