Uploaded image for project: 'SystemDS'
  1. SystemDS
  2. SYSTEMDS-1159

Enable Remote Hyperparameter Tuning



    • Type: Improvement
    • Status: Open
    • Priority: Major
    • Resolution: Unresolved
    • Affects Version/s: SystemML 1.1
    • Fix Version/s: None
    • Component/s: None
    • Labels:


      Training a parameterized machine learning model (such as a large neural net in deep learning) requires learning a set of ideal model parameters from the data, as well as determining appropriate hyperparameters (or "settings") for the training process itself. In the latter case, the hyperparameters (i.e. learning rate, regularization strength, dropout percentage, model architecture, etc.) can not be learned from the data, and instead are determined via a search across a space for each hyperparameter. For large numbers of hyperparameters (such as in deep learning models), the current literature points to performing staged, randomized grid searches over the space to produce distributions of performance, narrowing the space after each search [1]. Thus, for efficient hyperparameter optimization, it is desirable to train several models in parallel, with each model trained over the full dataset. For deep learning models, a mini-batch training approach is currently state-of-the-art, and thus separate models with different hyperparameters could, conceivably, be easily trained on each of the nodes in a cluster.

      In order to allow for the training of deep learning models, SystemML needs to determine a solution to enable this scenario with the Spark backend. Specifically, if the user has a train function that takes a set of hyperparameters and trains a model with a mini-batch approach (and thus is only making use of single-node instructions within the function), the user should be able to wrap this function with, for example, a remote parfor construct that samples hyperparameters and calls the train function on each machine in parallel.

      To be clear, each model would need access to the entire dataset, and each model would be trained independently.

      [1]: http://www.jmlr.org/papers/volume13/bergstra12a/bergstra12a.pdf


          Issue Links



              • Assignee:
                return_01 Janardhan
                dusenberrymw Mike Dusenberry
              • Votes:
                0 Vote for this issue
                1 Start watching this issue


                • Created: