NNTrainer 0.5.0 Release

We are releasing NNTrainer v0.5.0

RPM Files are for Tizen, built daily at build.tizen.org (https://build.tizen.org/package/show/Tizen:Unified/nntrainer)
and available at download.tizen.org. ( http://download.tizen.org/snapshots/tizen/unified/latest/repos/standard/packages/ )
DEB files are for Ubuntu, built and download from launchpad.net (https:://launchpad.net/~nnstreamer/+archive/ubuntu/ppa )

If you have unresolved dependencies, please download them from Ubuntu universe and nnstreamer PPA

In this release:

Fixes

• Reordering of execution order
• split apply gradient step in execution order
• Fix the memory pool and Tensor Pool bugs
  and more.

New Features

• New Features
  • Support Proactive Swap for less memory consumption
  • Add Cache Pool / Cache Loader / Cache Element
  • Update and add Memory Planner
  • Add Execution Order & Memory Usage Tracing for Debugging
  • Add TaskExecutor for multi-threading
  • Add Swish Activation
  • NNStreamer Training Plugins ( Un-stable )
  • Tensorflow-lite Exporter (Un-Stable)
    and more.
• Provides More C/C++ APIs
• New Applications
  • Add Android Applications ( Kotlin & Java ) for On-Device training of Resnet18
    and more

NNTrainer 0.4.0 Release

We are releasing NNTrainer v0.4.0

RPM Files are for Tizen, built daily at build.tizen.org (https://build.tizen.org/package/show/Tizen:Unified/nntrainer)
and available at download.tizen.org. ( http://download.tizen.org/snapshots/tizen/unified/latest/repos/standard/packages/ )
DEB files are for Ubuntu, built and download from launchpad.net (https:://launchpad.net/~nnstreamer/+archive/ubuntu/ppa )

If you have unresoved dependencies, please download them from Ubuntu universe and nnstreamer PPA

In this release:

Fixes

• Fix Batch Normalization Bugs
• Fix Embedding Layer Bugs
• Fix Grdient Access Bugs
• Add a lot of unit tests to evaluate NNTrainer implementation
  and more.

New Features

• New Layers
  • Attention Layer
  • Eanble Weight / Tensor Sharing
  • Implement Realizer to manipulate the network graph
    • Flatten Realizer, Recurrent Realizer with in/out property, Privious Input Realizer, Attach Activation Layer Realizer
  • Support Conv1D Layer
    • Support Dilation Property
  • Support multi-label/input for model
  • Support reshape Layer
  • Support Batch normalization 1 D
  • Support LSTM Cell Layer
  • Support RNNCell Layer
  • Support GRUCell Layer
  • Support Mol Attention Layer
  • Support Multi-Head Attention Layer
  • Support Gradient Clipping by Global Norm
  • Support Reduce Mean Layer
  • Support Leaky Relu Layer
  • Support Zoneout LSTM Cell Layer
  • Support Learning Rate Scheduling
  • Improve Load/Save Model
  • Support TFLite Export (Experimental)
  • Support Positional Encoding Layer
  • Support Layer Normalization
    and more
• Provides More C/C++ APIs
• New Applications
  • Transformer Applications
    and more

NNTrainer 0.3.0 Release

We are releasing NNTrainer v0.3.0

RPM Files are for Tizen, built daily at build.tizen.org (https://build.tizen.org/package/show/Tizen:Unified/nntrainer)
and available at download.tizen.org. ( http://download.tizen.org/snapshots/tizen/unified/latest/repos/standard/packages/ )
DEB files are for Ubuntu, built and download from launchpad.net (https:://launchpad.net/~nnstreamer/+archive/ubuntu/ppa )

If you have unresoved dependencies, please download them from Ubuntu universe and nnstreamer PPA

In this release:

Fixes

• Fix Batch Normalization Bugs
• Fix Stride and Padding in Conv2D and Pooling2D Layer
• Add a lot of unit tests to evaluate NNTrainer implementation
  and more.

New Features

• New Layers
  • Recurrent Layers : RNN, LSTM, GRU
  • Embedding Layer
  • Distributed Layer
  • KNN Layer
  • L2Norm Layer
• Rewrite DataSet to support element-wise ( not batch-wise ) getter & Better Data Handling
• Interpreter to convert the model into various other framework such as TfLite
• Provides More C/C++ APIs
  • Inferece APIs
  • Save / Load Model APIs
• New Applications
  and more

NNTrainer 0.2.0 Release

We are releasing NNTrainer v0.2.0

RPM Files are for Tizen, built daily at build.tizen.org (https://build.tizen.org/package/show/Tizen:Unified/nntrainer)
and available at download.tizen.org. ( http://download.tizen.org/snapshots/tizen/unified/latest/repos/standard/packages/ )
DEB files are for Ubuntu, built and download from launchpad.net (https:://launchpad.net/~nnstreamer/+archive/ubuntu/ppa )

If you have unresoved dependencies, please download them from Ubuntu universe and nnstreamer PPA

In this release:

Fixes

• Rewrite CONV2D to support Multi-Stride & Padding
• Fix DataSet synchronization problem
• Add a lot of unit tests to evaluate NNTrainer implementation
  and more.

New Features

• New Layers
  • Batch Normalization Layer
  • Addition & Concat Layer
  • Augmentation Layers : Flip / Translate / Permute / Split
  • Backbone Layer
  • Multi-Output Layer
  • Split Layer
• Support Custom Layer with Container (AppContext)
• Introdue Network Graph Structure & Optimization Scheme
• Introduce Techniques to Maximize Buffer Reusability
• Introduce Dynamic Fine-Tuning
• Support In/Out-Place & Lazy Tensor Computation
• Support In/Out Place Layer Calculation to reduce memory consumption
• Introduce Optimizer & Memory Manager for better maintain
• Provides C/C++ APIs
• New Applications
  • VGG
  • ResNet
  • Custom Layers
  • SimpleShot ( Meta-Learning )
    and more

NNTrainer 0.1.1 Release

We are releasing NNTrainer v0.1.1

RPM Files are for Tizen, built daily at build.tizen.org (https://build.tizen.org/package/show/Tizen:Unified/nntrainer)
and available at download.tizen.org. ( http://download.tizen.org/snapshots/tizen/unified/latest/repos/standard/packages/ )
DEB files are for Ubuntu, built and download from launchpad.net (https:://launchpad.net/~nnstreamer/+archive/ubuntu/ppa )

If you have unresoved dependencies, please download them from Ubuntu universe and nnstreamer PPA

In this release:

Fixes

• Fix for Softmax calculation
• Use im2col to compute Convolution Layer
• Update hyper-parameter keywords.
  • Network to Model
  • Weight_Decay to Weight_Regularizer
  • model_path to save_path
  • and others.
• Update Documentation
• Fix undeterministic behavior of databuffer
• Fix race condition of databuffer
• Resolve coverity and save issues
  and more.

New Features

• Added NNStreamer Filter Element for NNTrainer Inference
• Accelerate Tensor Calculation with BLAS Library

NNTrainer 0.1.0.rc1 Release

We are finally release NNtrianer v0.1.0.rc1.

RPM Files are for Tizen, built daily at build.tizen.org (https://build.tizen.org/package/show/Tizen:Unified/nntrainer)
and available at download.tizen.org. ( http://download.tizen.org/snapshots/tizen/unified/latest/repos/standard/packages/ )
DEB files are for Ubuntu, built and download from launchpad.net (https:://launchpad.net/~nnstreamer/+archive/ubuntu/ppa )

If you have unresolved dependencies, please download them from Ubuntu universe and nnstreamer PPA.

In this release:

New Features

• Supported Layers
  • Fully Connected Layer
  • Convolution 2D Layers
  • Pooling 2D Layer
  • Input Layer
  • Flatten Layer
  • Activation Layer
  • Loss Layer
• Supported Optimizers
  • sgd : Stochastic Gradient Descent
  • adam : Adaptive Moment Estimation
• Supported Loss
  • mse : Mean Squared Error
  • Cross Entropy : Sigmoid and Softmax
• Activations
  • tanh
  • sigmoid
  • relu
  • softmax
• Normalization
  • Weight Initialization : Xavier (Normal/Uniform), LeCun(Normal/Uniform), HE (Normal/Uniform)
  • Weight Decay : L2Norm
  • Learning Rate Decay
• Tensor : 4D Tensor ( B, C, H, W) accelerated by open blas
  • add, sub, mul, div
  • sum, average, argmax
  • dot, transpose
  • normalization, standardization
  • save, read
• APIs
  • Tizen Core API Support
• Applications
  • Tizen Application
    • Custom Shortcut Application
  • Full Training
    • mnist example
  • Reinforcement Learning
    • DeepQ Learning : CartPole
  • Transfer Learning
    • Classification of cifar10
    • Sticker Example
    • Sticker Example (KNN)
  • Logistic Regression