1. Congratulations!
Congratulations! You've now completed this course on the fundamentals of the TensorFlow API in Python. In this final video, we'll review what you've learned, talk about two useful TensorFlow extensions, and then wrap-up with a discussion of the transition to TensorFlow two point zero.
2. What you learned
In chapter 1, you learned low-level, basic, and advanced operations in TensorFlow.
You learned how to define and manipulate variables and constants.
You also learned the graph-based computational model that underlies TensorFlow and how it can be used to compute gradients and solve arbitrary optimization problems.
In chapter 2, you learned how to load and transform data for use in your TensorFlow projects.
You also saw how to use predefined and custom loss functions.
We ended with a discussion of how to train models, and when and how to divide the training into batches.
3. What you learned
In chapter 3, we moved on to training neural networks.
You learned how to define neural network architecture in TensorFlow, both using low-level linear algebra operations and high-level Keras API operations.
We talked about how to select activation functions and optimizers, and, ultimately, how to train models.
In chapter 4, you learned how to make full use of the Keras API to train models in TensorFlow.
We discussed the training and validation process and also introduced the high-level Estimators API, which can be used to streamline the production process.
4. TensorFlow extensions
In addition to what we covered, there are also a two important TensorFlow extensions that did not fit into the course, but may be worthwhile to explore on your own.
The first is TensorFlow Hub, which allows users to import pretrained models that can then be used to perform transfer learning. This will be particularly useful when you want to train an image classifier with a small number of images, but want to make use of a feature-extractor trained on a much larger set of different images.
TensorFlow Probability is another exciting extension, which is also currently available as a standalone module.
One benefit of using TensorFlow Probability is that it provides additional statistical distributions that can be used for random number generation.
It also enables you to incorporate trainable statistical distributions into your models.
Finally, TensorFlow Probability provides an extended set of optimizers that are commonly used in statistical research. This gives you additional tools beyond what the core TensorFlow module provides.
5. TensorFlow 2.0
Finally, I will say a few words about the difference between TensorFlow 2 and TensorFlow 1.
If you primarily develop in 1, you may have noticed that you do not need to define static graphs or enable eager execution. This is done automatically in 2.
Furthermore, TensorFlow 2 has substantially tighter integration with Keras. In fact, the core functionality of the TensorFlow 1 train module is handled by tf.Keras operations in 2.
In addition to the centrality of Keras, the Estimators API also plays a more important role in TensorFlow 2.
Finally, TensorFlow 2 also allows you to use static graphs, but they are available through the tf.function operation.
6. Congratulations!
Congratulations! You've now completed the course and are ready to begin training your own models in TensorFlow.