Bolt NLP Files

Added

• Support x86 compilation and cross-compialtion for ios/android on MacOs
• Support x86 compilation and cross-compilation for android on Windows
• Support MTK armv7 cross compilation toolchains on Linux by using linux-armv7_blank target
• Add Gitbook for user reference
• Support image nearest Resize and align_corners Resize
• Support more graph optimizations : Transpose+Concat+Transpose, Padding+Transpose, HardSwish-Fusion, Relu6-Fusion, Resize-Fusion, SwapTransposeEltwise, SwapPadTranspose, Convolution+Eltwise, Transpose+Matmul
• Support more operators: 3D-convolution, Where, SoftPlus, Exp, Split, Tdnn, Dropout, TopK, SpaceToBatchNd, BatchToSpaceNd, Abs, Equal, Sign, Resize(more mode)
• Support more networks on CPU: Reactnet, Tdnn, ShuffleNet, DenseNet, Hrnet, Efficientnet, Noah KWS2.0
• Support more networks on mali GPU : TinyBert, nmt
• Add more kit android/iOS demos : Simple-Image-Classification, Image-SuperResolution, Image-Classification
• Support float16, int8 model storage on any hardware
• Add Flow Java API

Changed

• Change install, GPU library process shell script
• Optimize TfSlice with 75%+ speed-up on cpu
• Optimize Concat with 50%+ speed-up on cpu
• Optimize Deconvolution with 10%+ speed-up on cpu
• Optimize YoloDetection network with 15%+ speed-up on cpu
• Optimize resnet50 from 90ms+ to 70ms+ on x86, faster than openvino
• Optimize mobilenet v1/v2 with 10%+ speed-up on x86
• Optimize tts-melgan network from 200ms+ to 160ms on x86
• Optimize model read time
• Change Java API package name and use com.huawei.noah, split single API file to 6 files.

Fixed

• Fix length of op/tensor name > 128 not-supporting bug
• Fix Caffe input dims extraction bug
• Fix Concat with single input in onnx converter
• Fix padding(nhwc) not-supporting bug
• Fix relu6 insertion in tflite converter
• Fix GRU, LSTM LBR_GRU model converter and inference bug
• Fix X86 convolution, fully connected operators inference bug

Removed

• Remove third party library FFTW and using FFTS for ASR example