Skip to content

Latest commit

 

History

History
168 lines (150 loc) · 8.41 KB

History.md

File metadata and controls

168 lines (150 loc) · 8.41 KB
Raw

3.1.0

  • Move to VS 2019
  • Fix a codegen error that was resulting in incorrect functional behavior
  • Fix regressions in audio training tutorial (#232)
  • Add importing of Sum nodes to ONNX importer
  • Fix crash in LLVMContext::SetName
  • Improved performance of CNN models on Pi3 with new implementations of spatial, pointwise and regular convolutions
  • Improved performance of reorder node
  • New nodes: ReorderDataCodeNode, SpatialConvolutionNode, MatrixMatrixMultiplyCodeNode
  • Implement parallelization strategies for matrix multiplication nodes.
  • Only enable new MatrixMatrixMultipleCodeNode path for select ARM targets like Pi, and not Intel/AMD CPUs
  • Add the flag --skip_ellcode to compile and wrap.py tools to use OpenBLAS for linear algebra computations.

3.0.3

  • Fix VS 2019 build.

3.0.2

  • Add FastGRNN node to ELL, train_classifier.py and the onnx_importer.

3.0.1

  • Fix make_vad.py so the VAD model works properly.
  • Fix smoothing of classifier output.
  • Improvements to noise mixing in audio training scripts
  • Use featurizer metadata to get correct defaults for some audio script parameters
  • Improvements and code clean up in view_audio.py
  • make_dataset will fill the featurizer internal buffer before returning features so it doesn't featurize zeros.
  • onnx_importer: fixes to ConstantFill, Reshape, Slice, Transpose, and support for constant folding involving Cast, and Mul and fix handling of nodes with multiple outputs.
  • Fix bug causing OptimizeReorderDataNodesTransformation to fail caused by input port and memory layout mismatch
  • move to SWIG 4.0.0.

3.0.0

  • Improve Python API so Map and CompiledMap have a way of calling compute on models with multiple inputs and outputs (see ComputeList)
  • Improve Python Callback handling so Python apps can handle models with any number of and type of callbacks.
  • Simplify datasetFromImages.py so it can create training, testing and validation set from one set of images on disk.
  • Update repurposing tutorial setup to take advantage of dataFromImages.py changes
  • Update index.md to fix bug in Linux instructions
  • Fix FFTNode so that behavior is identical to NumPy behavior
  • Update make_featurizer.py to calculate features correctly

2.5.4

  • Add support for ONNX GlobalAveragePooling
  • Publish 13 new audio models trained using Azure ML
  • Add ONNX importer support for ConstantOfShape
  • Add a -mean option to audio test runs to output mean prediction score
  • Fix InputNode unarchival to correctly set the memory layout

2.5.3

  • Fixed build break in C++ tutorial on Windows

2.5.2

  • Moved to LLVM 8.0
  • Added Asparagus and Pear models to the gallery, removed Elderflower
  • Added support for image preprocessing metadata and image normalization defaults for pytorch + imagenet
  • Updated find_ell.py (and thus other python tools) to work if your working directory contains ELL's build output

2.5.1

  • Fixed API wrapper for PortMemoryLayout to produce correct extent
  • Fixed processing of ConcatenationNode in importer engine
  • Added four build target aliases that build groups of targets: bindings, libraries, tests, and tools
  • Add support for ONNX Cast operator (thanks to Ehsan Azar PR #218)
  • Fixed build error on macOS (GitHub issue #216)
  • Added model per-layer optimizing scripts

2.5.0

  • Improve audio python library and update audio training tutorial
  • Add support for non-unrolled loops in the Value library
  • Add new methods to Python API to support FastGRNN models
  • Add easy node-creation functions
  • Fix bugs in ReinterpretLayoutNode
  • Fix importing of DenseNet models
  • Improve FFT and MelFilterBank nodes so that non-power of 2 sized input buffers can be processed
  • Fix darknet importer
  • Add shuffling of test and validation sets to audio train_classifier
  • Update OpenBLASSetup.cmake so it can find a locally built version of OpenBLAS
  • Fix compiling for cortex-m4 targets
  • Add a link back to the original model file name in our emitted header file
  • Clean up Python API, removing "TensorShape" specific API in favor of the more general PortMemoryLayout
  • Fix bug in Microphone and WavReader classes so they properly handle auto-resampling of audio with multiple channels
  • Add an -auto_scale option to audio scripts to control whether audio is scaled or not before featurization
  • Add ability to mix noise with audio during make_dataset process
  • Add support for shuffling and cleaning the audio training list
  • Add many new options to audio train_classifier, including ability to control learning rate schedulers
  • Fix link error when building on a Mac
  • Fix exception in DetectLowPrecisionConvolutionTransformation when running Compile in Debug mode

2.4.0

  • Initial release of the finetune utility
  • Remove stale references to VS 2015
  • Add logging of build tool options
  • Improve error reporting of compile failures
  • Suppress alias analysis on emitted IR code
  • Improve logging in ELL pythonlibs
  • Remove volatile attribute from memcpy, memmove, and memset LLVM intrinsic calls
  • Allow compilation of models with more than one input or output
  • Fix parsing of boolean command-line arguments
  • Add a RAM requirement for audio tutorial
  • Add support for optimized models
  • Fix ONNX import of no-op concat blocks
  • Gallery updates for Pi3 specific ELL models
  • Added documentation detailing ELL's versioning strategy

v2.3.7

  • Move to LLVM 6.0.1
  • Move ELL to C++17
  • Add support for python 3.7
  • Add support for more ONNX operations
  • Added instructions for Raspbian Lite (https://microsoft.github.io/ELL/tutorials/Raspberry-Pi-setup/)
  • Add new OpenBLAS include path for Ubuntu 18.04
  • Fix reorder data node optimization pass
  • Fix debugCompiler to work with gallery models
  • Improve reordering of data for simple and Winograd depthwise convolutions
  • Fixes for "Getting started with image classification on Raspberry Pi using C++" tutorial
  • Fix importing of depthwise separable models
  • Fix "Repurposing a pretrained image classifier" tutorial

v2.3.6

  • Initial version of the ONNX importer
  • Tools now have a -v/--verbose option that will control the output of additional helpful logging - data
  • Demo helper script has been modified to print labels to STDOUT instead of drawing on the image. - This is useful for some training sets like CIFAR
  • Couple of new options for wrap.py:
    • --convolution_method: this lets you specify the convolution method (it just passes the value on to the compile executable)
    • --llvm-format gets a new option: obj. This, along with --no-llc-tool and --no-opt-tool let you skip the time-consuming llc and opt steps. To retain compatibility with compile, the object file suffix is now .o.

v2.3.5

  • Fixed typos in some tutorials (thanks @lisaong )
  • Added support for emitting relocatable object code
  • Added Travis-CI support (thanks @lisaong )
  • Updated profiler to work with models from the ELL gallery
  • Fix build break for VS2017.7.3 in StlContainerReferenceIterator
  • Fix python interface implementation for LSTM layer nodes
  • Fix crash in retarget trainer (issue #141)
  • Fix docs for opencv setup on raspberry pi
  • Improve support for running models on Raspberry Pi Zero
  • make_profiler.sh can now be run from anywhere
  • Various bug fixes (#161, #157, #149)

v2.3.4

  • Fix importing CNTK models with a Softmax layer
  • Add a Reset function to LSTM and GRU models to clear the state

v2.3.3

  • Fixes issue #140
  • Helps issue #138 by fixing importing of Darknet models with Fully Connected layers
  • Simplified and improved LLVM IR optimization in emitted code
  • Fix importer for next round of separable darknet models in CNTK
  • Improved Winograd convolution implementation
  • Fix building the profiler tool
  • Fix running build.sh without any command line args
  • Miscellaneous bug fixes

v2.3.2

  • Fix issue #137
  • Added resnet models to gallery
  • Add support for naive depthwise separable convolution layers

v2.3.1

  • Added more fine grained profiling support for emitted code
  • Make find_package for LLVM more resilient to finding the version that ELL needs
  • Updated Setting up your Raspberry Pi device for tutorials to move to Raspbian Stretch
  • Updated links to YOLOv2 configuration and weights in Getting started with object detection using - region of interest networks
  • Added better support for importing models from older versions of CNTK

v2.3.0

  • New tutorials available:
  • Getting started with object detection using region of interest networks
  • Getting started with audio keyword classification on the Raspberry Pi
  • Updated importer to handle the importing of resnet models
  • Miscellaneous bug fixes.