Mindspore

The above table uses the mobilenet_v2 model from TF official website. Using MindSpore Lite quantization, the precision of A8W8 (8-bit activation value quantization and 8-bit weight quantization) decreases from 0.82% to 0.4% after accuracy loss compensation, for 7-bit quantization, the precision loss is still no more than 1%.

Training on Device

Within MindSpore 1.1 release, the MindSpore Lite provides the following Training-on-Device (ToD) capabilities:

1. Learning from scratch and Transfer Learning strategies are supported
2. MindSpore based models can be converted and used in training on the device. (Third-party models such as TensorFlow and PyTorch for now cannot be directly imported to the framework)
3. Grad operations are supported for more than 30 operators such as Dense layers, Convolutions and Batch Normalizations. Momentum, SGD, and ADAM optimizers are supported.
4. Supports networks such as LeNet, Alexnet, Resnet, MobileNetV1/V2/V3, and EffectiveNet, and provides complete model loading, conversion, and Python training scripts on the device side.

The MindSpore Lite ToD framework is already in use in the newest Huawei Smart TV, providing a unique and personalized user experience as a family entertainment center.

API Change

API Incompatible Change

C++ API

- [Modify] Context now support multi-context configuration.(Context.h)
- [Modify] Callback is move from lite_session.h into ms_tensor.h.
- [Modify] GetInputsByName in lite_session.h is changed into GetInputsByTensorName
- [Add] add static LiteSession *CreateSession(const char*model_buf, size_t size, const lite::Context *context) in lite_session.h
- [Add] add GetErrorInfo interface returning error message in errorcode.h
- [Delete] Remove model_generated.h, ops_generated.h and headers of FlatBuffers library from interfaces

Java API

- [Add] Implement JNI layer and add Java api for CPU and GPU backend

Deprecations

C++ API

Deprecate Interface GetOutputsByNodeName

Bug fixes

- [BUGFIX] Fix the bug in sub-graph segmentation
- [BUGFIX] Fix the bug in Tensor getitem in which the ellipsis matches the wrong dim-size.
- [BUGFIX] Fix the bug that activation modification after defining Dense will not take effect.

Contributors

Thanks goes to these wonderful people:

zhouyifengCode, huqi, JulyAi, damon0626, chenbo116, rmdyh, davidmc, gray0v0, doitH, Gogery, zymaa, xinyunfan

Adel, AGroupofProbiotocs, anthonyaje, anzhengqi, askmiao, baihuawei, baiyangfan, bai-yangfan, bingyaweng, BowenK, buxue, caifubi, CaoJian, caojian05, caozhou, Cathy, changzherui, chenbo116, chenfei, chengxianbin, chenhaozhe, chenjianping, chenzomi, chenzupeng, chujinjin, cj, cjh9368, Corleone, damon0626, danish, Danish, davidmc, dayschan, doitH, eric, Eric, fary86, fuzhiye, Gaoxiong, gengdongjie, Gogery, gongdaguo, gray0v0, gukecai, guoqi, gzhcv, hangq, hanhuifeng2020, Harshvardhan, He, heleiwang, hexia, Hoai, HuangBingjian, huangdongrun, huanghui, huangxinjing, huqi, huzhifeng, hwjiaorui, Jesse, jianghui58, jiangzhiwen, Jiaqi, jin-xiulang, jinyaohui, jjfeing, John, Jonathan, jonyguo, JulyAi, jzg, kai00, kingfo, kingxian, kpy, kswang, laiyongqiang, leonwanghui, Li, liangchenghui, liangzelang, lichen_101010, lichenever, lihongkang, lilei, limingqi107, ling, linqingke, liubuyu, liuwenhao4, liuxiao78, liuxiao93, liuyang_655, liuzhongkai, Lixia, lixian, liyanliu, liyong, lizhenyu, luoyang, lvchangquan, lvliang, lz, mahdi, Mahdi, maning202007, Margaret_wangrui, mayang, mengyuanli, nhussain, ougongchang, panfengfeng, panyifeng, Payne, Peilin, peixu_ren, Pengyongrong, qianlong, r1chardf1d0, riemann_penn, rmdyh, Sheng, shenwei41, simson, Simson, Su, sunsuodong, tao_yunhao, tinazhang, VectorSL, , Wan, wandongdong, wangdongxu, wangmin, wangnan39huawei.com, wangyue01, wangzhe, wanyiming, Wei, wenchunjiang, wilfChen, WilliamLian, wsc, wukesong, wuweikang, wuxuejian, Xiaoda, xiefangqi, xinyunfan, xuanyue, xulei2020, Xun, xuyongfei, yanghaitao, yanghaitao1, yanghaoran, YangLuo, yangruoqi713, yankai, yanzhenxiang2020, yao_yf, yepei6, yeyunpeng, Yi, yoni, yoonlee666, yuchaojie, yujianfeng, yuximiao, zengzitao, Zhang, zhanghaibo5huawei.com, zhanghuiyao, zhangyihui, zhangz0911gm, zhanke, zhanyuan, zhaodezan, zhaojichen, zhaoting, zhaozhenlong, zhengjun10, zhoufeng, zhousiyi, zhouyaqiang, zhouyifengCode, Zichun, Zirui, Ziyan, zjun, ZPaC, zymaa

Contributions of any kind are welcome!

Major Features and Improvements

DataSet

- [STABLE] Support integrity check, encryption and decryption check for MindRecord to protect the integrity and security of user data.
- [STABLE] MindRecord api changes: FileWriter.open_and_set_header is deprecated since it has been integrated into FilterWriter, if the old version code reports an error, delete this call; Add type checking for data in FileWriter to ensure that the data type defined by the Schema matches the real data type; The return value of all methods under Mindrecord are removed, replaced by an exception when processing error is occurred.
- [STABLE] Support Ascend processing backend for the following transforms: ResizedCrop, HorizontalFlip, VerticalFlip, Perspective, Crop, Pad, GaussianBlur, Affine.
- [STABLE] Optimized the content of data processing part in model migration guide, providing more examples to compare with third-party frameworks.
- [STABLE] Optimized the parsing efficiency of TFRecordDataset in multiple data columns scenario, improving the parsing performance by 20%.

PIJIT

- [BETA]PIJit analyzes and adjusts the Python bytecode and performs graph capture and graph optimization on the execution flow. Supported Python codes are executed in static graph mode, and unsupported ones are divided into subgraphs and executed in dynamic graph mode, automatically achieving dynamic and static unification. Users can enable the PIJit function by decorating the function with jit(mode="PIJit", jit_config={options:value}).

Inference

- [DEMO] The integrated architecture of large model inference, upgrade, training, and promotion unifies scripts, distributed policies, and runtime. The period from training to inference deployment of typical large models is reduced to days. Large operators are integrated to reduce the inference latency and effectively improve the network throughput.

AutoParallel

- [STABLE] Add msrun startup method to launch distributed job with single instruction.
- [STABLE] Add to be deprecated hint for RankTable startup method.
- [STABLE] Eliminate redundant constants in graph mode to improve compilation performance and memory overhead.
- [STABLE] The subgraph scenario optimizer parallelizes the first subgraph inline, allowing some computation and communication masking under pipeline parallelism to be performed.
- [STABLE] Communication information export: export model communication information (communication domain, communication volume) during compilation, and input it to the cluster as the basis for communication scheduling.
- [STABLE] Pipeline parallel inference is optimized, eliminates shared weights forwarding between stages, improving execution performance. Supports automatic broadcast of pipeline inference results, improving the usability of autoregressive inference.
- [STABLE] Operator-level parallel sharding supports the configuration of the mapping between the device layout and tensor layout during MatMul/Add/LayerNorm/GeLU/BiasAdd operator sharding.
- [STABLE] Supports gradient communication and backward calculation overlapping in the data parallel dimension.
- [STABLE] Single device simulation compilation, used to simulate the compilation process of a certain device in multi device distributed training, assisting in analyzing the compilation processes and memory usage on the front and back ends.
- [STABLE] Implement ops.Tril sharding to reduce the memory and performance requirements on a single device.
- [BETA] Supports the fusion between communication operators and computing operators, in order to overlap communication overheads with computation and improve network performance.
- [BETA] Load checkpoints and compile graphs in parallel to accelerate fault recovery.

Runtime

- [BETA] Support O0/O1/O2 multi-level compilation to improve static graph debugging and tuning capabilities.

FrontEnd

- [STABLE] The framework supports the bfloat16 data type. dtype=mindspore.bfloat16 can be specified when a tensor is created.
- [STABLE] The syntax support capability of the rewrite component is optimized, syntaxs such as class variables, functions, and control flows can be parsed.
- [STABLE] New context setting: debug_level. User can use mindspore.set_context(debug_level=mindspore.DEBUG) to get more debug information.

Profiler

- [BETA] Dynamically start and stop profiling. Users can collect profiling data in real time according to the training situation, reducing the amount of data collected.
- [BETA] Profiling the communication operator time-consuming matrix. Users can find cluster communication performance bottlenecks by analyzing the communication operator time-consuming matrix.
- [BETA] Improve the performance of Ascend environment in parsing profiling data.
- [BETA] Supports offline analysis of data generated by Profiling. Users can collect data first and then parse the data as needed.
- [BETA] Supports collecting performance data of HBM, PCIe, and l2_cache to enrich performance analysis indicators.

Dump

- [BETA] The statistical information saved by Dump records MD5 values, and users can determine small differences in tensor values through MD5 values.
- [BETA] Dump supports the float16 data type and supports users to locate float16 type operator accuracy issues.

PyNative

- [STABLE] Reconstruct the single operator calling process for dynamic graphs to improve the performance of dynamic graphs.

Ascend

- [BETA] Support set configuration options of CANN, which are divided into two categories: global and session. Users can configure them through mindspore.set_context(Ascend_configuration={"ge_options": {"global": {"global_option": "option_value"}, "session": {"session option": "option_value"}}).

API Change

- Add mindspore.hal API to support stream, event, and device management capabilities.
- Add mindspore.multiprocessing API to provide the capability of creating multiple processes.

Operators

- [BETA] mindspore.ops.TopK now supports the second input k as an int32 type tensor.

Bug Fixes

- [I92H93] Fixed the issue of 'Launch kernel failed' when using the Print operator to print string objects on the Ascend platform.
- [I8S6LY] Fixed RuntimeError: Attribute dyn_input_sizes of Default/AddN-op1 is [const vector]{}, of which size is less than 0 error of variable-length input operator, such as AddN or Concat, for dynamic shape process in graph mode on the Ascend platform.
- [I9ADZS] Fixed the data timeout issue in network training due to inefficient dataset recovery in the fault recovery scenario.

Contributors

Thanks goes to these wonderful people:

AlanCheng511，AlanCheng712，bantao，Bingliang，BJ-WANG，Bokai Li，Brian-K，caifubi，cao1zhg，CaoWenbin，ccsszz，chaiyouheng，changzherui，chenfei_mindspore，chengbin，chengfeng27，chengxb7532，chenjianping，chenkang，chenweifeng，Chong，chuht，chujinjin，Cynthia叶，dairenjie，DavidFFFan，DeshiChen，douzhixing，emmmmtang，Erpim，fangzhou0329，fary86，fengxun，fengyixing，fuhouyu，gaoshuanglong，gaoyong10，GaoZhenlong，gengdongjie，gent1e，Greatpan，GTT，guoqi，guoxiaokang1，GuoZhibin，guozhijian，hangq，hanhuifeng，haozhang，hedongdong，hejianheng，Henry Shi，heyingjiao，HighCloud，Hongxing，huandong1，huangbingjian，HuangLe02，huangxinjing，huangziling，hujiahui8，huoxinyou，jiangchenglin3，jianghui58，jiangshanfeng，jiaorui，jiaxueyu，JichenZhao，jijiarong，jjfeing，JoeyLin，JuiceZ，jxl，kairui_kou，kate，KevinYi，kisnwang，lanzhineng，liangchenghui，LiangZhibo，lianliguang，lichen，ligan，lihao，limingqi107，ling，linqingke，liruyu，liubuyu，liuchao，liuchengji，liujunzhu，liuluobin，liutongtong9，liuzhuoran2333，liyan2022，liyejun，LLLRT，looop5，luochao60，luojianing，luoyang，LV，machenggui，maning202007，Margaret_wangrui，MaZhiming，mengyuanli，MooYeh，moran，Mrtutu，NaCN，nomindcarry，panshaowu，panzhihui，PingqiLi，qinzheng，qiuzhongya，Rice，shaojunsong，Shawny，shenwei41，shenyaxin，shunyuanhan，silver，Songyuanwei，tangdezhi_123，tanghuikang，tan-wei-cheng，TingWang，TronZhang，TuDouNi，VectorSL，WANG Cong，wang_ziqi，wanghenchang，wangpingan，wangshaocong，wangtongyu6，weiyang，WinXPQAQ，wtcheng，wudawei，wujiangming，wujueying，wuweikang，wwwbby，XianglongZeng，xiaosh，xiaotianci，xiaoxin_zhang，xiaoxiongzhu，xiaoyao，XinDu，xingzhongfan，yanghaoran，yangluhang，yangruoqi713，yangzhenzhang，yangzishuo，yanjiaming，Yanzhi_YI，yao_yf，yefeng，yeyunpeng2020，yide12，YijieChen，YingLai Lin，YingtongHu，youshu，yuchaojie，YuJianfeng，zangqx，zby，zhaiyukun，zhangdanyang，zhanghaibo，zhanghanLeo，zhangminli，zhangqinghua，zhangyanhui，zhangyifan，zhangyinxia，zhangyongxian，ZhangZGC，zhanzhan，zhaoting，zhengyafei，zhengzuohe，ZhihaoLi，zhouyaqiang0，zhuguodong，zhumingming，zhupuxu，zichun_ye，zjun，zlq2020，ZPaC，zuochuanyong，zyli2020，陈宇，代宇鑫，狄新凯，范吉斌，冯一航，胡彬，宦晓玲，黄勇，康伟，李良灿，李林杰，刘崇鸣，刘力力，刘勇琪，吕浩宇，没有窗户的小巷，王禹程，吴蕴溥，熊攀，徐安越，徐永飞，许哲纶，俞涵，张峻源，张树仁，张王泽，张栩浩，郑裔，周莉莉，周先琪，朱家兴，邹文祥

Contributions of any kind are welcome!