Search Articles

Overview #

• Made by Google, open source, useful for machine learning and building neural networks
• Keras is the high-level API for TensorFlow
  • You should always use the Keras API for TensorFlow
    • Very few problems need to be solved at the low-level core TensorFlow APIs

Concepts #

• tf.keras.layers.Layer - This is the fundamental abstraction in keras
  • A layer encapsulates a state (weights) and some computation (the call method)
• tf.keras.Model - Provides the fundamental abstraction for a neural network
  • The call function of the model defines the network’s forward pass
  • The summary method can be used to print a summary of the model architecture
• Sequential API - Allows you to build a model by stacking layers on top of each other
  • This is the simplest way to build a model
• Automatic differentiation - TensorFlow automatically computes gradients for you for backpropagation training
  • This is done by the GradientTape API
  • GradientTape records operations for automatic differentiation
    • So the tape knows which operations to record in the forward pass so they can be optimized in the backward pass
    • The most common use case is to compute the gradient of the loss with respect to the models trainable variables1
      • tape.gradient(loss, model.trainable_variables) is the most common use case to abstract away the backpropagation

Layer Types #

• Tensor Flow provides layer abstractions like a Dense layer