This issue tracks potential changes that would be made in a 1.0.0 release. This should be quite far off at this point. Pruning: - see how pruning the kernel in DepthwiseConv2D influences converging accuracy. If not significant, prune by default. (prune_registry.py) - use only tf.math.divide instead of tf.div for 1.XX (pruning_schedule.py > PolynomialDecay > __call__ )
Activity