The fact that we're able to do everything as part of the TF graph is really nice — Python slowness is never an issue. There's no need for us to rewrite anything in, like, Cython or Rust.
@fchollet
-
Keras Preprocessing Layers: High-Performance In-Graph Implementation
By
–
All the work is done in Keras preprocessing layers, which are implemented in TF ops (everything is 100% in-graph!) so it's highly performant. During training (presumably on GPU/TPU) you'd use async preprocessing in TF data to avoid CPU preprocessing being a bottleneck.