Localised training and dynamic class targets

General Discussion
,
#1

Tensorflow appears to be designed for ML engineering projects (end-to-end backprop training), however non-backprop learning algorithm research requires a) the ability to specify precisely which parts of a model are being trained at a given time (ie localised training), and b) the precise training objective (class targets), where the class targets may be a function of the output of another part of the network, not just some predefined training/test set. I have encountered a number of exploratory algorithms which appear impossible to implement with the current framework (without copying weights across independent graphs).

#3

@bairesearch Welcome to the TF Forum!

While TensorFlow is widely used for end-to-end backpropagation training, it offers flexibility for non-backprop learning algorithms as well. Here are ways to address your requirements:

  1. Localised Training: Manual Gradient Calculation: For complete control, leverage tf.GradientTape 's gradient method with custom computations. Layer-wise Trainability:** Set trainable=False for layers you want to freeze during training.

  2. Precise Training Objectives: Custom Loss Functions:** Define any loss function using TensorFlow operations, even those dependent on other parts of the network. Modify loss calculation within training loops based on model outputs or other conditions.

  3. Advanced Techniques for Non-Backprop Algorithms: Custom Training Loops:** For full control over training steps, create custom loops using tf.GradientTape and optimizer’s apply_gradients method.
    TensorFlow allows experimenting with different gradient calculation methods, not limited to backpropagation. Explore gradient-free optimization methods (e.g., evolutionary algorithms, reinforcement learning) with TensorFlow’s core operations and functions.

Let us know if this helps!