Development Schedule

Release Module Feature
v0.1 Sep 2015 Neural Network Feed forward neural network, including CNN, MLP
    RBM-like model, including RBM
    Recurrent neural network, including standard RNN
  Architecture One worker group on single node (with data partition)
    Multi worker groups on single node using Hogwild
    Distributed Hogwild
    Multi groups across nodes, like Downpour
    All-Reduce training architecture like DeepImage
    Load-balance among servers
  Failure recovery Checkpoint and restore
  Tools Installation with GNU auto Tools
v0.2 Jan 2016 Neural Network Feed forward neural network, including AlexNet, cuDNN layers,Tools
    Recurrent neural network, including GRULayer and BPTT
    Model partition and hybrid partition
  Tools Integration with Mesos for resource management
    Prepare Docker images for deployment
    Visualization of neural net and debug information
  Binding Python binding for major components
  GPU Single node with multiple GPUs
v0.3 April 2016 GPU Multiple nodes, each with multiple GPUs
    Heterogeneous training using both GPU and CPU CcT
    Support cuDNN v4
  Installation Remove dependency on ZeroMQ, CZMQ, Zookeeper for single node training
  Updater Add new SGD updaters including Adam, AdamMax and AdaDelta
  Binding Enhance Python binding for training
v1.0 Sep 2016 Programming abstraction Tensor with linear algebra, neural net and random operations
    Updater for distributed parameter updating
  Hardware Use Cuda and Cudnn for Nvidia GPU
    Use OpenCL for AMD GPU or other devices
  Cross-platform To extend from Linux to MacOS
    Large image models, e.g., VGG and Residual Net
v1.1 Jan 2017 Model Zoo GoogleNet; Health-care models
  Caffe converter Use SINGA to train models configured in caffe proto files
  Model components Add concat and slice layers; accept multiple inputs to the net
  Compilation and installation Windows suppport
    Simplify the installation by compiling protobuf and openblas together with SINGA
    Build python wheel automatically using Jenkins
    Install SINGA from Debian packages
v1.2 Oct 2017 Numpy API Implement functions for the tensor module of PySINGA following numpy API
  Distributed training Migrate distributed training frameworks from V0.3
  Memory optimization Replace CNMEM with new memory pool to reduce memory footprint
  Execution optimization Runtime optimization of execution scheduling