Making sense of TensorFlow 2.0

Thoughts

The process becomes syntactically similar to PyTorch, no more usage of sessions and explicitly defining graphs. I like the addition of the ability to record metrics using a dedicated metrics class, which also provides methods for calculating things like KL-Divergence, etc.

Tensorflow 2.0 Quick Start

• Import required libraries.
• Load data tensors and pre-process them.
• Create the Dataset objects for training and testing:
  • Initialize with from_tensor_slices().
  • Shuffle and batch with shuffle(num_samples) and batch(batch_size).
• Define model, loss function, and optimizer.
• Define metrics for accumulation.
• Define training and test loops:
  • Use @tf.function decorator to compile to graph.
  • Training loop:
    • Use tf.GradientTape to record operations.
    • Pass data through model, record loss and weights.
    • Apply gradient update using optimizer.
    • Accumulate metrics.
  • Test loop:
    • Pass data through model, record loss.
    • Accumulate metrics.
• Execute the training and test loops.
• Reset states of metrics with reset_states().