I have a custom data generator that inherence tf.keras.utils.Sequence class. And I have found that in tensorflow doc, tf.data is more efficient to use and recommend. But re-writing Sequence to tf.data can be costly. There are two option I can do
- one: optimize Sequence generator and integrate some advanced mechanism to boost it speed
- two: rewrite to
tf.datafrom previous data loader
For now, I want to choose one. How can we speed up sequence data loader speed in training time? Not the conventional way but some more smart approach. Thanks.
