MindSpore Insight Release Notes

查看中文

MindSpore Insight 2.2.1 Release Notes

Bug fixes

• [I88AN5] MindSpore Insight adapts to Numpy version higher than 1.20.0.

Contributors

Thanks goes to these wonderful people:

Ning Ma, Jiaxing Zhu, Jiarong Ji, Yanming Miao, Nan Wang, XiaoXian Jin, Qingxiang Zang, Yang Luo, TianCi Xiao, DaWei Fan.

Contributions of any kind are welcome!

MindSpore Insight 2.2.0 Release Notes

Major Features and Improvements

Profiler

• [STABLE] Profiler supports collecting custom AICPU operator time consumption.
• [Beta] Supports timeline data merging capabilities of multiple cards.

Dump

• [Beta] Provides statistical capabilities for overflow operators.

Bug fixes

• [I7J1LF] Fixed the IndexError problem in Profiler parsing datagrams.
• [I82CGQ] Fixed overflow detection reporting core dump issue.

Contributors

Thanks goes to these wonderful people:

Ning Ma, Jiaxing Zhu, Jiarong Ji, Yanming Miao, Nan Wang, XiaoXian Jin, Qingxiang Zang, Yang Luo, TianCi Xiao, DaWei Fan.

Contributions of any kind are welcome!

MindSpore Insight 2.1.0 Release Notes

Major Features and Improvements

Profiler

• [STABLE] Profiler supports the collection of time-consuming data at each stage of the Host.
• [Beta] Profiler supports the collection of memory data at each stage of the Host.
• [Beta] Profiler supports the execution time consumption of collecting data processing operators.

Contributors

Thanks goes to these wonderful people:

Ning Ma, Jiaxing Zhu, Jiarong Ji, Yanming Miao, Nan Wang, XiaoXian Jin, Qingxiang Zang, Yang Luo, TianCi Xiao, DaWei Fan.

Contributions of any kind are welcome!

MindSpore Insight 2.0.0 Release Notes

Bug fixes

• [I7BIKO] Fix the inaccurate Flops problem in the mixed scene of cube and vector

Contributors

Thanks goes to these wonderful people:

Ning Ma, Jiaxing Zhu, Jiarong Ji, Yanming Miao, Nan Wang, XiaoXian Jin, Chuting Liu, Han Gao, Qingxiang Zang.

Contributions of any kind are welcome!

MindSpore Insight 2.0.0-rc1 Release Notes

Major Features and Improvements

MindSpore Insight

• [STABLE] MindSpore Insight and Mindspore version matching verification.

Debugger

• [STABLE] The debugger shows that the upper limit of the number of nodes in the calculation graph is configurable.

Profiler

• [STABLE] The time-consuming ratio of the Profiler operator is calculated using total time.

Bug fixes

• Fix some page display issues.

Contributors

Thanks goes to these wonderful people:

Ning Ma, Chuting Liu, Jiaxing Zhu, Qingxiang Zang, Yaomin Mao.

Contributions of any kind are welcome!

MindInsight 2.0.0-alpha Release Notes

Major Features and Improvements

Profiling

• [STABLE] Profiler supports enabling through environment variables
• [STABLE] Provides an interface for generating PMU performance data (Ascend)
• [BETA] PyNative Mode, Accuracy optimization of Profiler operator performance data (Ascend)
• [BETA] Profiler supports PyNative mode basics (GPU)
• [STABLE] Support Msprof binary tool to pull up Mindspore Profiling (Ascend)
• [BETA] Profiling supports Dynamic shape network (GPU)

Dump

• [STABLE] Dump supports dynamic shape

Contributors

Thanks goes to these wonderful people:

Ning Ma, Chuting Liu, Jiaxing Zhu, Qingxiang Zang, Yaomin Mao.

Contributions of any kind are welcome!

MindInsight 1.9.0 Release Notes

Major Features and Improvements

Profiling

• [BETA] Parallel Training Execution Analysis (Ascend)
• [BETA] Msadvisor function integration into Mindspore / MindStudio (Ascend)

Summary

• [STABLE] Automatic identification of precision reduction operator,In Ascend scenario, some operators only support float16 with the highest accuracy, which will cause the accuracy of this type of operator to automatically decline. This function is used to help users identify this type of precision reducing operator.

API Change

Backwards Compatible Change

Python API

Contributors

Thanks goes to these wonderful people:

Kai Wen, Yue Wang, Ximiao Yu, Ning Ma, Haitao Yang, Han Gao, Chuting Liu, Jiaxing Zhu, Qingxiang Zang.

Special thanks to Zhongwei Wang, Rongchen Zhu, Jiaying Lu, Zhiyong Wang, Yating Wei, Yong Dai, Luoxuan Weng, etc., from State Key Lab of CAD&CG, Zhejiang University led by Prof. Wei Chen, for their contributions of innovative frontend and interaction technology to support parallel training execution analysis module, collective communication analysis module, etc.

Contributions of any kind are welcome!

MindInsight 1.8.0 Release Notes

Major Features and Improvements

Profiling

• [STABLE] Profiler supports dynamic shape operator (Ascend)
• [STABLE] The profiler sample code is adjusted according to the import specification

Debugger

• [STABLE] Dump, fixed randomness document optimization

API Change

Backwards Compatible Change

Python API

• [stable] Profiler adds new operator performance query interface

Contributors

Thanks goes to these wonderful people:

Congli Gao, Longfei Li, Yongxiong Liang, Chongming Liu, Pengting Luo, Yanming Miao, Gongchang Ou, Kai Wen, Yue Wang, Lihua Ye, Ximiao Yu, Yunshu Zhang, Ning Ma, Yihui Zhang, Hong Sheng, Ran Mo, Zhaohong Guo, Tianshu Liang, Shuqiang Jiang, Yanjun Peng, Haitao Yang, Jiabin Liu, Han Gao, Xiaohui Li, Ngaifai Ng, Hui Pan, Weifeng Huang, Yifan Xia, Xuefeng Feng, Yanxi Wei, Yufeng Lv, Maohua He, Chuting Liu, Jiaxing Zhu, Yuanwei Song.

Special thanks to Zhiyong Wang, Zhongwei Wang, Rusheng Pan, Yating Wei, Luoxuan Weng, Rongchen Zhu, Jingli Xu, Qinxian Liu, Haozhe Feng, Tong Xu, etc., from State Key Lab of CAD&CG, Zhejiang University led by Prof. Wei Chen, for their contributions of innovative frontend and interaction technology to support strategy perception including Computational Graph Exploration module, Parallel Strategy Analysis module, etc.

Contributions of any kind are welcome!

MindInsight 1.7.0 Release Notes

Major Features and Improvements

Profiling

• [STABLE] Profiler supports analyzing iterative trajectories (GPU)
• [BETA] Profiler supports PyNative format (Ascend)

Debugger

• [STABLE] Summary API provides instructions in Chinese
• [STABLE] Add Summary sample code
• [STABLE] Debugger checking watchpoint optimization, performance improvement

Contributors

Thanks goes to these wonderful people:

Congli Gao, Longfei Li, Yongxiong Liang, Chongming Liu, Pengting Luo, Yanming Miao, Gongchang Ou, Kai Wen, Yue Wang, Lihua Ye, Ximiao Yu, Yunshu Zhang, Ning Ma, Yihui Zhang, Hong Sheng, Ran Mo, Zhaohong Guo, Tianshu Liang, Shuqiang Jiang, Yanjun Peng, Haitao Yang, Jiabin Liu, Han Gao, Xiaohui Li, Ngaifai Ng, Hui Pan, Weifeng Huang, Yifan Xia, Xuefeng Feng, Yanxi Wei, Yufeng Lv, Maohua He, Chuting Liu, Jiaxing Zhu, Yuanwei Song.

Special thanks to Zhiyong Wang, Zhongwei Wang, Rusheng Pan, Yating Wei, Luoxuan Weng, Rongchen Zhu, Jingli Xu, Qinxian Liu, Haozhe Feng, Tong Xu, etc., from State Key Lab of CAD&CG, Zhejiang University led by Prof. Wei Chen, for their contributions of innovative frontend and interaction technology to support strategy perception including Computational Graph Exploration module, Parallel Strategy Analysis module, etc.

Contributions of any kind are welcome!

MindInsight 1.6.0

MindInsight 1.6.0 Release Notes

Major Features and Improvements

Profiling

• [STABLE] Support starting Profiler in the process of training.(Ascend)
• [STABLE] Support strategy perception includes Computational Graph Exploration module, Parallel Strategy Analysis module, etc.(Ascend)
• [STABLE] Support cluster performance helper to give some prompts to users.(Ascend)

MindConverter

• [STABLE] Support migrating definition scripts and trained weights by TorchScript.(Ascend/GPU)

Debugger

• [STABLE] Support Python API for offline debugger analysis.(Ascend/GPU)

Landscape

• [BETA] Support analyzing model training optimization process by computing and visualizing Loss landscape.(Ascend/GPU)

API Change

Backwards Compatible Change

Python API

• Add mindconverter.pytorch2mindspore() interface for converting models from PyTorch into MindSpore.
• Add mindinsight.debugger Python API for offline debugger analysis.

Contributors

Thanks goes to these wonderful people:

Congli Gao, Longfei Li, Yongxiong Liang, Chongming Liu, Pengting Luo, Yanming Miao, Gongchang Ou, Kai Wen, Yue Wang, Lihua Ye, Ximiao Yu, Yunshu Zhang, Ning Ma, Yihui Zhang, Hong Sheng, Ran Mo, Zhaohong Guo, Tianshu Liang, Shuqiang Jiang, Yanjun Peng, Haitao Yang, Jiabin Liu, Han Gao, Xiaohui Li, Ngaifai Ng, Hui Pan, Weifeng Huang, Yifan Xia, Xuefeng Feng, Yanxi Wei, Yufeng Lv, Maohua He, Chuting Liu, Jiaxing Zhu, Yuanwei Song.

Special thanks to Zhiyong Wang, Zhongwei Wang, Rusheng Pan, Yating Wei, Luoxuan Weng, Rongchen Zhu, Jingli Xu, Qinxian Liu, Haozhe Feng, Tong Xu, etc., from State Key Lab of CAD&CG, Zhejiang University led by Prof. Wei Chen, for their contributions of innovative frontend and interaction technology to support strategy perception including Computational Graph Exploration module, Parallel Strategy Analysis module, etc.

Contributions of any kind are welcome!

MindInsight 1.5.0

MindInsight 1.5.0 Release Notes

Major Features and Improvements

Profiling

• [STABLE] Unify performance data output path.
• [STABLE] Analyse overlap time between communication operators and compution operators.

MindConverter

• [STABLE] Support migrating definition scripts and trained weights for object detection model(YOLOv5s), face detection model(RetinaFace), NLP model(BigBird) and document image understanding model(LayoutLM).(Ascend/GPU)
• [STABLE] Optimize and improve the usability of the official documentation, describe the migration procedure in detail and supplement FAQs.(Ascend/GPU)

Model Explanation

Debugger

• [STABLE] Add tensor memory control for offline debugger.(Ascend/GPU)
• [STABLE] Support search on watchpoint hit nodes.(Ascend/GPU)
• [STABLE] Guidance document for model precision problem locating, guidance document for model precision optimization.(Ascend/GPU)

Build & Installation

API Change

Backwards Compatible Change

Python API

Command Line Interface

reviously, we don't set memory limit for offline debugger. In order to use offline debugger in limited environment, we provide with memory limit options when start MindInsight server. View the Offline Debugger Tutorial.

New start command options:

Name	Attribute	Description	Type	Default Value	Range
-offline-debugger-mem-limit	Optional	Specifies the maximum memory limit of a single offline debugger session. When the offline debugger cannot be executed due to insufficient memory, set it according to the device memory.	Integer	16*1024	6*1024~The upper limit of int32
--max-offline-debugger-session-num	Optional	Specifies the maximum session number of the offline debugger. The session number refers to the amount of training jobs that can be debugged at the same time.	Integer	2	1~2

Bug fixes

• Wrong sorting of cards displayed on the single page and cluster.!11801

Contributors

Thanks goes to these wonderful people:

Congli Gao, Longfei Li, Yongxiong Liang, Chongming Liu, Pengting Luo, Yanming Miao, Gongchang Ou, Kai Wen, Yue Wang, Lihua Ye, Ximiao Yu, Yunshu Zhang, Ning Ma, Yihui Zhang, Hong Sheng, Ran Mo, Zhaohong Guo, Tianshu Liang, Shuqiang Jiang, Yanjun Peng, Haitao Yang, Jiabin Liu, Han Gao, Xiaohui Li, Ngaifai Ng, Hui Pan, Weifeng Huang, Yifan Xia, Xuefeng Feng, Yanxi Wei, Yufeng Lv, Maohua He, Chuting Liu.

Contributions of any kind are welcome!

MindInsight 1.4.0

MindInsight 1.4.0 Release Notes

Major Features and Improvements

Build & Installation

API Change

Backwards Compatible Change

Python API

Command Line Interface

NA

Bug fixes

Contributors

Thanks goes to these wonderful people:

Congli Gao, Longfei Li, Yongxiong Liang, Chongming Liu, Pengting Luo, Yanming Miao, Gongchang Ou, Kai Wen, Yue Wang, Lihua Ye, Ximiao Yu, Yunshu Zhang, Ning Ma, Yihui Zhang, Hong Sheng, Ran Mo, Zhaohong Guo, Tianshu Liang, Shuqiang Jiang, Yanjun Peng, Haitao Yang, Jiabin Liu, Han Gao, Xiaohui Li, Ngaifai Ng, Hui Pan, Weifeng Huang, Yifan Xia, Xuefeng Feng, Yanxi Wei.

Contributions of any kind are welcome!

MindInsight 1.3.0

MindInsight 1.3.0 Release Notes

Major Features and Improvements

Profiling

• [STABLE] Support memory analysis using heat map in cluster profiling ui page.(Ascend)
• [STABLE] Support show scope information of operations in timeline.(Ascend/GPU)
• [STABLE] Support FLOPs statistics in single machine and cluster profiling ui page.(Ascend)
• [STABLE] Support show link bandwidth, waiting and communication time of communication promitives including allreduce,allgather,etc in cluster profiling ui page.(Ascend)

MindConverter

• [STABLE] Support both recommend model(wide&deep, deepfm) and NLP model(albert, bert, bert_nezha, LSTM) definition script and trained weights migration from TensorFlow or PyTorch.
• [STABLE] Support convert ONNX model whose size is larger than 2GB.
• [STABLE] Support adjust readability using Fix CheckPoint file Tool.

Model Explanation

• [STABLE] Support counterfactual explanation for image classification.

Debugger

• [STABLE] Support offline debugger.(Ascend/GPU)
• [STABLE] Support source code mapping.(Ascend/GPU)
• [STABLE] Support download tensor from UI.(Ascend/GPU)

Build & Installation

• [STABLE] Unified MindInsight installation package, supporting multiple Linux distributions, CPU architectures(x86/ARM), and Python versions(3.7/3.8/3.9).

API Change

Backwards Compatible Change

Python API

Add parameter 'profile_memory' to Profiler.(!17742)

Determine whether collect memory information while profiling. Default is False.

Add parameter 'profile_communication' to Profiler.(!17558)

Determine whether collect communication performance information while profiling. Default is False.

Command Line Interface

NA

Bug fixes

• Error information missing when running on an unsupported device (e.g, cpu).(!11801)

Contributors

Thanks goes to these wonderful people:

Congli Gao, Longfei Li, Yongxiong Liang, Chongming Liu, Pengting Luo, Yanming Miao, Gongchang Ou, Kai Wen, Yue Wang, Lihua Ye, Ximiao Yu, Yunshu Zhang, Ning Ma, Yihui Zhang, Hong Sheng, Ran Mo, Zhaohong Guo, Tianshu Liang, Shuqiang Jiang, Yanjun Peng, Haitao Yang, Jiabin Liu, Han Gao, Xiaohui Li, Ngaifai Ng, Hui Pan, Weifeng Huang, Yifan Xia, Xuefeng Feng, Yanxi Wei.

Contributions of any kind are welcome!

MindInsight 1.2.0

MindInsight 1.2.0 Release Notes

Major Features and Improvements

Profiling

• [STABLE] Support memory profiling.(Ascend)
• [STABLE] Support host cpu utilization profiling.(Ascend/GPU)
• [STABLE] Support timeline for Host&Device Hybrid Training.(Ascend/GPU)
• [STABLE] Support show step breakdown information(Step Interval, Forward and Backward Propagation, and Step Tail) of each device in cluster profiling ui page.(Ascend)

MindConverter

• [STABLE] Support both classic computer vision and bert model definition script and trained weights migration from TensorFlow or PyTorch.
• [STABLE] Support ONNX model migration to improve the usability of PyTorch model migration.

Model Explanation

• [STABLE] Support counterfactual explanation for image classification.

API Change

Backwards Compatible Change

Python API

add parameter export_options for SummaryCollector and SummaryRecord(!10881)

Perform custom operations on the export data. You can customize the export data with a dictionary. For example, you can set {'tensor_format': 'npy'} to export tensor as npy file.

add parameter raise_exception for SummaryRecord(!10436)

The parameter raise_exception determines whether to throw an exception when an exception occurs.

add API register_uncertainty for explainer.ImageClassificationRunner(!11309)

register_uncertainty helps register uncertainty instance to compute the epistemic uncertainty base on the Bayes’ theorem.

add API register_hierarchical_occlusion for explainer.ImageClassificationRunner(!11309)

register_hierarchical_occlusion helps register hierarchical occlusion instances.

Command Line Interface

MindConverter removes support for pth format model, --project_path deleted(!1253)

The pth format model is not supported anymore, please use ONNX to migrate.

Bug fixes

• Error information missing when running on an unsupported device (e.g, cpu). !11801

Contributors

Thanks goes to these wonderful people:

Congli Gao, Longfei Li, Yongxiong Liang, Chongming Liu, Pengting Luo, Yanming Miao, Gongchang Ou, Kai Wen, Yue Wang, Lihua Ye, Ximiao Yu, Yunshu Zhang, Ning Ma, Yihui Zhang, Hong Sheng, Ran Mo, Zhaohong Guo, Tianshu Liang, Shuqiang Jiang, Yanjun Peng, Haitao Yang, Jiabin Liu, Han Gao, Xiaohui Li, Ngaifai Ng, Hui Pan, Weifeng Huang, Yifan Xia, Xuefeng Feng, Yanxi Wei.

Contributions of any kind are welcome!

MindInsight 1.1.0

MindInsight 1.1.0 Release Notes

Major Features and Improvements

Precision tuning framework

• Support useful checks on weights, activations, gradients and tensors, such as:
  • check unchanged weight
  • check weight change above threshold
  • check activation range
  • check gradient vanishing
  • check tensor overflow
• Support rechecking with new watch points on the same data.
• Newly designed tensor view with fix suggestions and tensor context to quickly locate root cause of problems.
• Support recommending watch points to find common precision problems.
• Support debugger on multigraph network.

Profiler

• Support GPU step trace profiling.
• Support GPU minddata profiling.

MindConverter

• Support TensorFlow model definition script to MindSpore for CV field.
• Conversion capability of PyTorch is enhanced.

Model Explanation

Provide explanations and their benchmarks for image classification deep CNN models.

• Support 6 explanation methods: Gradient, Deconvolution, GuidedBackprop, GradCAM, RISE, Occlusion
• Support 4 benchmark methods: Localization, Faithfulness, Class Sensitivity, Robustness
• Provide a high-level API (ImageClassificationRunner) for users to execute explanation methods and benchmark methods and store the results easily.

API Change

Improvements

Command Line Interface

• --enable_debugger: Support both 1 and True (!1051)
• ENABLE_MS_DEBUGGER: Support both 1 and True (!10199)
• parse_summary: Add parse_summary function to convert summary file to image file and csv file (!774)

Bugfixes

Profiler

• Fix parser framework file error if the profiling data of one op is saved separately to two files.(!7824)

Model Explanation

• Add reset_offset when CRCLengthError and CRCError happen(!955)
• FIx the bug which ignore the sample_event when sample_id == 0.(!968)

Thanks to our Contributors

Thanks goes to these wonderful people:

Congli Gao, Jianfeng Zhu, Zhenzhong Kou, Longfei Li, Yongxiong Liang, Chongming Liu, Pengting Luo, Yanming Miao, Gongchang Ou, Yongxiu Qu, Luyu Qiu, Kai Wen, Yue Wang, Lihua Ye, Ximiao Yu, Yunshu Zhang, Ning Ma, Yihui Zhang, Shuide Wang, Hong Sheng, Ran Mo, Zhaohong Guo, Hui Pan, Weining Wang, Weifeng Huang, Yifan Xia, Chen Cao, Ngaifai Ng, Xiaohui Li, Yi Yang, Luyu Qiu, Yunpeng Wang, Yuhan Shi, Yanxi Wei.

Contributions of any kind are welcome!

MindInsight 1.0.0

MindInsight 1.0.0 Release Notes

Major Features and Improvements

• Release MindSpore Debugger.
• MindConverter ability is enhanced, supporting scripts generation based on PyTorch model.
• Support training hyper-parameter importance visualization.
• Support GPU timeline.

Bugfixes

• Optimize aicpu display method. (!595)
• Add the summary loading switch mechanism. (!601)
• Detect a summary dir having summary files or not. (!632)

Thanks to our Contributors

Thanks goes to these wonderful people:

Congli Gao, Jianfeng Zhu, Zhenzhong Kou, Hongzhang Li, Longfei Li, Yongxiong Liang, Chongming Liu, Pengting Luo, Yanming Miao, Gongchang Ou, Yongxiu Qu, Luyu Qiu, Kai Wen, Yue Wang, Lihua Ye, Ximiao Yu, Yunshu Zhang, Ning Ma, Yihui Zhang, Shuide Wang, Hong Sheng, Ran Mo, Zhaohong Guo, Hui Pan, Junyan Qin, Weining Wang, Weifeng Huang, Yifan Xia.

Contributions of any kind are welcome!

MindInsight 0.7.0-beta

MindInsight 0.7.0 Release Notes

Major Features and Improvements

• Optimize node name display in computation graph.
• MindSpore Profiler supports network training with GPU operators.
• MindWizard generates classic network scripts according to user preference.
• Web UI supports language internationalization, including both Chinese and English.

Bugfixes

• Optimize UI page initialization to handle timeout requests. (!503)
• Fix the line break problem when the profiling file number is too long. (!532)

Thanks to our Contributors

Thanks goes to these wonderful people:

Congli Gao, Weifeng Huang, Zhenzhong Kou, Hongzhang Li, Longfei Li, Yongxiong Liang, Chongming Liu, Pengting Luo, Yanming Miao, Gongchang Ou, Yongxiu Qu, Hui Pan, Luyu Qiu, Junyan Qin, Kai Wen, Weining Wang, Yue Wang, Zhuanke Wu, Yifan Xia, Lihua Ye, Weibiao Yu, Ximiao Yu, Yunshu Zhang, Ting Zhao, Jianfeng Zhu, Ning Ma, Yihui Zhang, Shuide Wang, Hong Sheng, Lin Pan, Ran Mo.

Contributions of any kind are welcome!

MindInsight 0.6.0-beta

MindInsight 0.6.0 Release Notes

Major Features and Improvements

• Provide monitoring capabilities for each of Ascend AI processor and other hardware resources, including CPU and memory.
• Visualization of weight, gradient and other tensor data in model training.
  • Provide tabular from presentation of tensor data.
  • Provide histogram to show the distribution of tensor data and its change over time.

Bugfixes

• UI fix for the error message display mode of the tensor during real-time training. (!465)
• The summary file size is larger than max_file_size. (!3481)
• Fix real-time training error when disk is full. (!3058)

Thanks to our Contributors

Thanks goes to these wonderful people:

Congli Gao, Weifeng Huang, Zhenzhong Kou, Hongzhang Li, Longfei Li, Yongxiong Liang, Chongming Liu, Pengting Luo, Yanming Miao, Gongchang Ou, Yongxiu Qu, Hui Pan, Luyu Qiu, Junyan Qin, Kai Wen, Weining Wang, Yue Wang, Zhuanke Wu, Yifan Xia, Lihua Ye, Weibiao Yu, Ximiao Yu, Yunshu Zhang, Ting Zhao, Jianfeng Zhu, Ning Ma, Yihui Zhang, Shuide Wang.

Contributions of any kind are welcome!

MindInsight 0.5.0-beta

MindInsight 0.5.0 Release Notes

Major Features and Improvements

• MindSpore Profiler
  • Provide performance analyse tool for the input data pipeline.
  • Provide timeline analyse tool, which can show the detail of the streams/tasks.
  • Provide a tool to visualize the step trace information, which can be used to analyse the general performance of the neural network in each phase.
  • Provide profiling guides for the users to find the performance bottlenecks quickly.
• CPU summary operations support for CPU summary data.
• Over threshold warn support in scalar training dashboard.
• Provide more user-friendly callback function for visualization
  • Provide unified callback SummaryCollector to log most commonly visualization event.
  • Discard the original visualization callback SummaryStep, TrainLineage and EvalLineage.
  • SummaryRecord provide new API add_value to collect data into cache for summary persistence.
  • SummaryRecord provide new API set_mode to distinguish summary persistence mode at different stages.
• MindConverter supports conversion of more operators and networks, and improves its ease of use.

Bugfixes

• Fix FileNotFound exception by adding robust check for summary watcher (!281).
• UI fix operator table sort jump problem (!283).
• Dataset serializer return schema json str when schema type is mindspore.dataset.engine.Schema (!2185).

Thanks to our Contributors

Thanks goes to these wonderful people:

Chao Chen, Congli Gao, Ye Huang, Weifeng Huang, Zhenzhong Kou, Hongzhang Li, Longfei Li, Yongxiong Liang, Chongming Liu, Pengting Luo, Yanming Miao, Gongchang Ou, Yongxiu Qu, Hui Pan, Luyu Qiu, Junyan Qin, Kai Wen, Weining Wang, Yue Wang, Zhuanke Wu, Yifan Xia, Lihua Ye, Weibiao Yu, Ximiao Yu, Yunshu Zhang, Ting Zhao, Jianfeng Zhu.

Contributions of any kind are welcome!

MindInsight 0.3.0-alpha

MindInsight 0.3.0 Release Notes

Major Features and Improvements

• Profiling
  • Provide easy to use apis for profiling start/stop and profiling data analyse (on Ascend only).
  • Provide operators performance display and analysis on MindInsight UI.
• Large scale network computation graph visualization.
• Optimize summary record implementation and improve its performance.
• Improve lineage usability
  • Optimize lineage display and enrich tabular operation.
  • Decouple lineage callback from SummaryRecord.
• Support scalar compare of multiple runs.
• Scripts conversion from other frameworks
  • Support for converting PyTorch scripts within TorchVision to MindSpore scripts automatically.

Bugfixes

• Fix pb files loaded problem when files are modified at the same time (!53).
• Fix load data thread stuck in LineageCacheItemUpdater (!114).
• Fix samples from previous steps erased due to tags size too large problem (!86).
• Fix image and histogram event package error (!1143).
• Equally distribute histogram ignoring actual step number to avoid large white space (!66).

Thanks to our Contributors

Thanks goes to these wonderful people:

Chao Chen, Congli Gao, Ye Huang, Weifeng Huang, Zhenzhong Kou, Hongzhang Li, Longfei Li, Yongxiong Liang, Pengting Luo, Yanming Miao, Gongchang Ou, Yongxiu Qu, Hui Pan, Luyu Qiu, Junyan Qin, Kai Wen, Weining Wang, Yue Wang, Zhuanke Wu, Yifan Xia, Weibiao Yu, Ximiao Yu, Ting Zhao, Jianfeng Zhu.

Contributions of any kind are welcome!

MindInsight 0.2.0-alpha

MindInsight 0.2.0 Release Notes

Major Features and Improvements

• Parameter distribution graph (Histogram).

  Now you can use HistogramSummary and MindInsight to record and visualize distribution info of tensors. See our tutorial.

• Lineage support Custom information

• GPU support

• Model and dataset tracking linkage support

Bugfixes

• Reduce cyclomatic complexity of list_summary_directories (!11).
• Fix unsafe functions and duplication files and redundant codes (!14).
• Fix sha256 checksum missing bug (!24).
• Fix graph bug when node name is empty (!34).
• Fix start/stop command error code incorrect (!44).

Thanks to our Contributors

Thanks goes to these wonderful people:

Ye Huang, Weifeng Huang, Zhenzhong Kou, Pengting Luo, Hongzhang Li, Yongxiong Liang, Gongchang Ou, Hui Pan, Luyu Qiu, Junyan Qin, Kai Wen, Weining Wang, Yifan Xia, Yunshu Zhang, Ting Zhao

Contributions of any kind are welcome!

MindInsight 0.1.0-alpha

MindInsight 0.1.0 Release Notes

• Training process observation

  • Provides and displays training process information, including computational graphs and training process indicators.

• Training result tracing

  • Provides functions of tracing and visualizing model training parameter information, including filtering and sorting of training data, model accuracy and training hyperparameters.