Welcome to the Model Zoo for MindSpore

The MindSpore models repository provides different task domains, classic SOTA model implementations and end-to-end solutions. The purpose is to make it easier for MindSpore users to use MindSpore for research and product development.

In order to facilitate developers to enjoy the benefits of MindSpore framework, we will continue to add typical networks and some of the related pre-trained models. If you have needs for the model zoo, you can file an issue on gitee or MindSpore, We will consider it in time.

WHAT IS NEW

• We've done code refactoring for classic SOTA models,modularized data processing, model definition&creation, training process and other common components with new lanched MindSpore CV/NLP/Audio/Yolo/OCR Series toolbox. link.

• Old models were implemented by original MindSpore API with tricks for model training speedup.

Disclaimers

Mindspore only provides scripts that downloads and preprocesses public datasets. We do not own these datasets and are not responsible for their quality or maintenance. Please make sure you have permission to use the dataset under the dataset’s license. The models trained on these dataset are for non-commercial research and educational purpose only.

To dataset owners: we will remove or update all public content upon request if you don’t want your dataset included on Mindspore, or wish to update it in any way. Please contact us through a Github/Gitee issue. Your understanding and contribution to this community is greatly appreciated.

MindSpore is Apache 2.0 licensed. Please see the LICENSE file.

License

Apache License 2.0

FAQ

For more information about MindSpore framework, please refer to FAQ

• Q: How to resolve the lack of memory while using the model directly under "models" with errors such as Failed to alloc memory pool memory?

  A: The typical reason for insufficient memory when directly using models under "models" is due to differences in operating mode (PYNATIVE_MODE), operating environment configuration, and license control (AI-TOKEN).

  • PYNATIVE_MODE usually uses more memory than GRAPH_MODE , especially in the training graph that needs back propagation calculation, there are two ways to try to solve this problem. Method 1: You can try to use some smaller batch size; Method 2: Add context.set_context(mempool_block_size="XXGB"), where the current maximum effective value of "XX" can be set to "31". If method 1 and method 2 are used in combination, the effect will be better.
  • The operating environment will also cause similar problems due to the different configurations of NPU cores, memory, etc.;
  • Different gears of License control (AI-TOKEN ) will cause different memory overhead during execution. You can also try to use some smaller batch sizes.

• Q: How to resolve the error about the interface are not supported in some network operations, such as cann not import?

  A: Please check the version of MindSpore and the branch you fetch the modelzoo scripts. Some model scripits in latest branch will use new interface in the latest version of MindSpore.

• Q: What is Some RANK_TBAL_FILE which mentioned in many models?

  A: RANK_TABLE_FILE is the config file of cluster on Ascend while running distributed training. For more information, you could refer to the generator hccl_tools and Parallel Distributed Training Example

• Q: How to run the scripts on Windows system?

  A: Most the start-up scripts are written in bash, but we usually can't run bash directly on Windows. You can try start python directly without bash scripts. If you really need the start-up bash scripts, we suggest you the following method to get a bash environment on Windows:

  1. Use a virtual system or docker container with linux system. Then run the scripts in the virtual system or container.
  2. Use WSL, you could turn on the Windows Subsystem for Linux on Windows to obtain an linux system which could run the bash scripts.
  3. Use some bash tools on Windows, such as cygwin and git bash.

• Q: How to resolve the compile error point to gflags when infer on ascend310 with errors such as undefined reference to 'google::FlagRegisterer::FlagRegisterer'?

  A: Please check the version of GCC and gflags. You can refer to GCC and gflags to install GCC and gflags. You need to ensure that the components used are ABI compatible, for more information, please refer to _GLIBCXX_USE_CXX11_ABI.

• Q: How to solve the error when loading dataset in mindrecord format on Mac system, such as Invalid file, failed to open files for reading mindrecord files.?

  A: Please check the system limit with ulimit -a, if the number of file descriptors is 256 (default), you need to use ulimit -n 1024 to set it to 1024 (or larger). Then check whether the file is damaged or modified.

• Q: What should I do if I can't reach the accuracy while training with several servers instead of a single server?

  A: Most of the models has only been trained on single server with at most 8 pcs. Because the batch_size used in MindSpore only represent the batch size of single GPU/NPU, the global_batch_size will increase while training with multi-server. Different gloabl_batch_size requires different hyper parameter including learning_rate, etc. So you have to optimize these hyperparameters will training with multi-servers.

• Q: How to solve the out of memory error?

  A: If there is a max_device_memory field in the model's configuration file *.yaml, set this field, for example: max_device_memory: "20GB". If computing splitting is used, the device memory in the 8-cut scenario is 8G, and it is recommended to set it to 6GB. If there is no max_device_memory field in the configuration file *.yaml, 8-cut scenarios recommend adding ms.set_context(max_device_memory="6GB") to the training file.