0.2.1.7

• Vulkan winograd bug Fix

• Fix Segment Fault issue crash when converting TFLite Models

• Workaround Metal Softmax Op

• Lower required OpenCL version to 1.10 from 2.0 to 1.10

• vulkan winograd bug 修复

• 修正 tflite 转换时段错误问题

• metal softmax bug 暂时规避

• OpenCL 所需要版本由 2.0 降到 1.10 ，提升兼容性

0.2.1.6

Op Support

Added support for:

• Over 30 TFLite Ops
• Over 20 Onnx Ops
• 9 Caffe Ops
• 24 Tensorflow Ops

Project Layout and Engineering Improvements

• CMake Build System Rewrite
• Header Layout Standardization. Public headers are now under include/MNN and is installed as <MNN/>

Inferencing Improvements and Bug Fixes

• OpenCL BinaryOp Bugs
• CPU MatMul Bugs
• Added Unit Testing for over 30 ops
• ImageProcess now supports NV12 input and NV21 / NV12 stride

Training (Experimental Feature)

• Optimize Express Module's Dynamic Graph Execution Policy
• Improve single machine training with a working demo

Op 补全

• 新增 TFlite 30 + op 支持
• 新增 Onnx 20 + op 支持
• 新增 Caffe 9 个 op 支持
• 新增 Tensorflow 24 个 op 支持

工程优化

• 完善CMake相关编译配置
• 头文件目录规范化，原 include 目录下的头文件移到 include/MNN 下

推理相关功能完善与Bug修复

• OpenCL BinaryOp 相关 Bug 修复
• CPU MatMul Bug 修复
• 完善单元测试，添加 30 + op 单元测试用例
• ImageProcess 支持 NV12 输入以及 NV21 / NV12 stride 支持

训练能力(整体仍处调试阶段)

• 优化 Express 模块动态图运行机制
• MNN 单机训练功能完善，Demo完成

0.2.1.5

0.2.1.5

integration

• add travis CI
• fix building parameters for python

converter

• add half storage option for MNN converter
• fix op name lost in converter
• fix converter bug for print input output, identity remove output

ops

• add quantized Convolution & Deconvolution support on OpenCL
• add more expression supports
• add DetectionPostProcess Op for TensorFlow Lite (ssd is supported directly now)
• add supports for LSTM & ELU for ONNX
• add support for Convolution that weights is not constant for ONNX
• fix Unary Op compile error on Linux
• fix Metal backend buffer reuse after resize
• fix Metal raw memory access after model releasing
• fix redundant transpose in Winograd generater

0.2.1.2

build

• unify schema building in core and converter;
• add more build script for android;
• add linux build script for python;

ops impl

• add floor mod support in binary;
• use eltwise impl in add/max/sub/mul binary for optimization;
• remove fake double support in cast;
• fix 5d support for concat;
• add adjX and adjY support for batch matmul;
• optimize conv2d back prop filter;
• add pad mode support for conv3d;
• fix bug in conv2d & conv depthwise with very small feature map;
• optimize binary without broacast;
• add data types support for gather;
• add gather ND support;
• use uint8 data type in gather v2;
• add transpose support for matmul;
• add matrix band part;
• add dim != 4 support for padding, reshape & tensor convert;
• add pad type support for pool3d;
• make ops based on TensorFlow Lite quantization optional;
• add all & any support for reduction;
• use type in parameter as output type in reduction;
• add int support for unary;
• add variable weight support for conv2d;
• fix conv2d depthwise weights initialization;
• fix type support for transpose;
• fix grad outputs count for reduce grad and reshape grad;
• fix priorbox & detection output;
• fix metal softmax error;

python

• add runSessionWithCallBackInfo interface;
• add max nodes limit (1400) for visualization tool;
• fix save error in python3;
• align default dim;

convert

• add extra design for optimization;
• add more post converting optimizers;
• add caffe v1 weights blob support;
• add cast, unary, conv transpose support for onnx model;
• optimize batchnorm, conv with variable weights, prelu, reshape, slice, upsample for onnx model;
• add cos/sin/atan/tan support for unary for tensorflow model;
• add any/all support for reduction for tensorflow model;
• add elu, conv3d, pool3d support for tensorflow model;
• optimize argmax, batchnorm, concat, batch to space, conv with variable weights, prelu, slice for tensorflow model;

others

• fix size computer lock;
• fix thread pool deadlock;
• add express & parameters in express;
• rewrite blitter chooser without static map;
• add tests for expr;

0.2.1.0

0.2.1.0

• dynamic computation graph (beta)
  • add supports (/express)
  • add tests
  • add benchmarks with it (/benchmark/exprModels)
• Python
  • MNN engine and tools were submitted to pip
  • available on Windows/macOS/Linux
• Engine/Converter
  • add supports for each op benchmarking
  • refactor optimizer by separating steps
• CPU
  • add supports for Conv3D, Pool3D, ELU, ReverseSequence
  • fix ArgMax, Permute, Scale, BinaryOp, Slice, SliceTf
• OpenCL
  • add half transform in CPU
  • add broadcast supports for binary
  • optimize Conv2D, Reshape, Eltwise, Gemm, etc.
• OpenGL
  • add sub, real div supports for binary
  • add supports for unary
  • optimize Conv2D, Reshape
• Vulkan
  • add max supports for eltwise
• Metal
  • fix metallib missing problem
• Train/Quantization
  • use express to refactor training codes

beta 0.2.0.9

beta 0.2.0.9

• fix quantization tool compiling on Windows
• fix converter compiling on Windows
• fix eltwise optimization on Windows
• separate sse & avx for Windows
• add LeakyReLU support for TensorFlow
• fix reshape, const for TensorFlow
• fix dimension format error for ONNX ops
• optimize winograd, ReLU for OpenCL
• add fp16 availability & dimensions size check-up for OpenCL
• optimize GEMM for arm32
• fix ExpandDims shape calculation when inputs size == 1

beta 0.2.0.8

beta 0.2.0.8

• add NaN check-up
• add quantification support for ScaleAdd Op
• add binary to eltwise optimization
• add console logs for quantization tool
• better document for quantization tool
• replace redundant dimension flags with dimension format
• optimize performance of TensorFlow Lite Quantized Convolution
• fix axis support for ONNX softmax converting
• fix getPerformance tool compiling error on Windows

beta 0.2.0.7

• move docs to http://www.yuque.com/mnn
• fix bugs for CPU ops TopKV2 and quantized convolution
• add enqueue map buffer error handle for OpenCL
• add nullptr protection for extra tensor desc
• add failure protection for memory acquirement
• fix slice shape calculation
• refactor binary shape calculation

release 0.2.0.6

• fix bugs in quantization
• add evaluating tool for quantization
• add ADMM support in quantization
• fix lock in thread pool
• fix fusing for deconv
• fix reshape converting from ONNX to MNN
• turn off blob size checking by default

beta 0.2.0.5

beta 0.2.0.5

• CPU
• add support for DepthToSpace & SpaceToDepth ops
• OpenGL
• add Android demo
• add half / float runtime option
• add support for ROIPooling, Squeeze
• fix bugs in conv im2col
• OpenCL
• fix Concat, Eltwise, Reshape bugs
• Tools
• add KL threshold method in quantization tool
• support optimization for graph with multiple rnn