Closed bhack closed 4 years ago
Hyperparameter optimization is a sensible way to use a cluster, so it's worth supporting well. Currently, if people load a model in SparkNet from a Caffe .caffemodel
or a TensorFlow .pb
file, then the net architecture will already be specified in the model file. Hyperparameter optimization becomes easier if the users specify the models in Scala, but that seems like a less common use case.
Yes but there are also some emerging approach that pass the boundaries concept of hyperparameters search like what we started to discuss at https://github.com/maxpumperla/hyperas/issues/10
Do you plan to add something similar to https://github.com/maxpumperla/hyperas?