Machine learning professionals interested in hardware acceleration are expected to bring exciting developments to the test bench. This is possible through the recently released TensorFlow framework version 2.8.
While the release does not include a host of new capabilities, it develops some of the more recent performance-related tools, as seen by the number of improvements in the still embedding APIs and experimental TensorRT.
Incorporating TensorFlow with the program in v2.8 gives a deeper understanding of the TF-TRT module’s analysis. It lets users prohibit the framework from storing TRT-specific algorithms, minimizing resource utilization.
In the meantime, the embeddings APIs, which have been part of the aspect that enables collaboration with Google’s tensor processing elements, have seen the introduction of a new assertion for defining the form of a feature’s production activation. Moreover, it will coordinate behavior for serving embedding lookup andTPUEmbedding.
Beginning with the current version, TensorFlow’s input processing unit tf.data will parallelize the duplicating of batch items. They will have a more straightforward way with file data, as TensorSliceDataset now understands how to recognize and manage it.
Aside from the more fundamental features, developers will notice a few substantial modifications. For example, Tf.keras, a data augmentation API that was moved to its repo in August 2021, features a new random variable generation for Keras initializers, all RND code, and an output mode option for the Hashing and Discretization layers.
Therefore, consumers will find it easier to lowercase inputs, remove punctuation, and divide text on Unicode sequences using the new TextVectorization options.
Mobile-focused TFLite has GPU assignment functionality for serialization to the Java API and working with certain random value producers, the when, and the raw ops. Moreover, operators should be bucketed. Developers who previously utilized Interpreter::SetNumThreads must update their software to use InterpreterBuilder::SetNumThreads, as the latter has already been discontinued with the update.
Advice from the TensorFlow team
The TensorFlow group also recommends replacing any boosted tree code with TensorFlow Decision Forests, as boosted tree code created specific security vulnerabilities and was removed before version 2.9.
2 minutes ago
