Updates 2017/07/21

The new semester started last week, so I was busy with my students. Development has mainly happened around deeplearning4j and prediction support for Microsoft's deeplearning library CNTK.

Fixes

  • LoadBalancer: fixed losing of outer variables; uses a Flow control actor now internally for better encapsulation.

Changes

  • Added support for outputting relative paths with the FileSystemSearch source: LocalDirectorySearch, LocalDirectorySearchWithComparator, LocalDirectorySearchWithCustomSort, LocalFileSearch
  • Panels managed by the DisplayPanelManager get re-used via their unique ID now properly (eg when using a variable), not just when mergable. Allows out of order updates of sequence plots now.
  • The Min and Max transformers can return 1-based indices now.
  • Added support for ADAMS_LIBRARY_PATH environment variable to adams.core.management.Launcher: its content gets supplied to the JVM via -Djava.library.path (used for native libraries like CNTK, MKL).
  • adams-dl4j:
    • Added ability to DL4JTrainModel transformer for testing the model on a test set (split off the training data) and output the best model found so far, with associated statistic(s).
    • Added support for criteria to stop training to DL4JTrainModel rather than just having fixed number of epochs.
  • adams-weka: The WekaFilter transformer can make use of storage and source actor now for obtaining the actual filter to use, not just serialized file or the filter specification.

Additions

  • Added the ArrayNormalize array statistic, which normalizes an array to sum up to 1.0.
  • adams-cntk:
    • added support for applying CNTK models: CNTKModelApplier.
    • added spreadsheet writer for CNTK text file format: CNTKSpreadSheetWriter
    • added image feature generator: DefaultCNTK
  • adams-cntk-weka: Added pseudo-classifier that uses a pre-built model: functions.CNTKPrebuiltModel
  • adams-dl4j: Added transformer for randomizing dataset: DL4JRandomizeDataset
  • adams-imaging:
    • Added ScaleReportObjects transformer for scaling objects defined in reports.
    • Added example flow for training an OpenCV Haar cascade from annotated images: adams-imaging-opencv_train_haar.flow
  • adams-imaging-openimaj: added generic object detector class hierarchy, to be used by adams.flow.transformer.locateobjects.OpenIMAJObjectDetector
  • adams-spreadsheet: added class hierarchy for processors that work on the selected rows in a spreadsheet table, e.g., copying files using the filename from the specified column. Functionality available through SpreadSheetDisplay sink.