megengine / mgeconvert

MegEngine到其他框架的转换器

MegEngine

Documentation | 中文文档

MegEngine is a fast, scalable, and user friendly deep learning framework with 3 key features.

• Unified framework for both training and inference
  • Quantization, dynamic shape/image pre-processing, and even derivation with a single model.
  • After training, put everything into your model to inference on any platform with speed and precision. Check here for a quick guide.
• The lowest hardware requirements
  • The memory usage of the GPU can be reduced to one-third of the original memory usage when DTR algorithm is enabled.
  • Inference models with the lowest memory usage by leveraging our Pushdown memory planner.
• Inference efficiently on all platforms
  • Inference with speed and high-precision on x86, Arm, CUDA, and RoCM.
  • Supports Linux, Windows, iOS, Android, TEE, etc.
  • Optimize performance and memory usage by leveraging our advanced features.

---

Installation

NOTE: MegEngine now supports Python installation on Linux-64bit/Windows-64bit/MacOS(CPU-Only)-10.14+/Android 7+(CPU-Only) platforms with Python from 3.6 to 3.9. On Windows 10 you can either install the Linux distribution through Windows Subsystem for Linux (WSL) or install the Windows distribution directly. Many other platforms are supported for inference.

Binaries

To install the pre-built binaries via pip wheels:

python3 -m pip install --upgrade pip
python3 -m pip install megengine -f https://megengine.org.cn/whl/mge.html

Building from Source

• CMake build details. please refer to BUILD_README.md
• Python binding build details, Please refer to BUILD_PYTHON_WHL_README.md

How to Contribute

• MegEngine adopts Contributor Covenant as a guideline to run our community. Please read the Code of Conduct.
• Every contributor of MegEngine must sign a Contributor License Agreement (CLA) to clarify the intellectual property license granted with the contributions.
• You can help to improve MegEngine in many ways:
  • Write code.
  • Improve documentation.
  • Answer questions on MegEngine Forum, or Stack Overflow.
  • Contribute new models in MegEngine Model Hub.
  • Try a new idea on MegStudio.
  • Report or investigate bugs and issues.
  • Review Pull Requests.
  • Star MegEngine repo.
  • Cite MegEngine in your papers and articles.
  • Recommend MegEngine to your friends.
  • Any other form of contribution is welcomed.

We strive to build an open and friendly community. We aim to power humanity with AI.

How to Contact Us

• Issue: github.com/MegEngine/MegEngine/issues
• Email: [email protected]
• Forum: discuss.megengine.org.cn
• QQ Group: 1029741705

Resources

• MegEngine
• MegStudio
• mirror repo
  • OPENI: openi.org.cn/MegEngine
  • Gitee: gitee.com/MegEngine/MegEngine

License

MegEngine is licensed under the Apache License, Version 2.0

Citation

If you use MegEngine in your publication,please cite it by using the following BibTeX entry.

@Misc{MegEngine,
  institution = {megvii},
  title =  {MegEngine:A fast, scalable and easy-to-use deep learning framework},
  howpublished = {\url{https://github.com/MegEngine/MegEngine}},
  year = {2020}
}

Copyright (c) 2014-2021 Megvii Inc. All rights reserved.

mgeconvert's Issues

带bias的linear层转onnx，为什么转换成Gemm+Add？

测试函数为：
test/test_onnx.py/test_linear()

onnx结果：

带bias的linear层转换成Gemm+Add，不带bias的linear层转换成Gemm。其中Gemm的bias都是为1维。

是否应该改为：
带bias的linear层转换成Gemm，其中gemm的bias为oc维；
不带bias的linear层转换成Gemm， bias为空？

OP Split is not supported

onnx->mge 需要支持算子

1.需要支持下reducemax\tile；
2.另外Unsqueeze是有问题的；
3.onnx和onnxsim建议安装的版本好像没有，有一些低版本会有bug

AssertionError: Forward Result of ONNX and Mge Mismatch -3.6064906120300293(ONNX) vs -3.5924134254455566(Mge) with model at index 3

RuntimeError: access invalid Maybe value

Hello,

When trying to convert the CREStereo MegEngine model to ONNX, I am facing the below error.

Below is the system configuration used
ubuntu 20.04
MegEngine==1.13.0
mgeconvert==0.7.0.dev0
cuda 12.3

torch->onnx->mge过程中的后半段（onnx->mge）的问题

在onnx模型转mge模型时候无法正常运行，卡在下面这个位置点不动了。

from mgeconvert.converters.onnx_to_mge import onnx_to_mge
onnx_to_mge( 'onnx2.onnx', output="mge.mge", )
ONNX Model Producer : pytorch
ONNX Model Producer Version: 1.8
ONNX Model IR Version : 6
ONNX Model OpSet : 11
`

tm convert 到 caffe 参数问题

master源码安装，TracedModule 转换到 caffe 的时候由于参数问题会报错，没有outspec

Traceback (most recent call last):
  File "/home/lvhaoran/.local/bin/convert", line 187, in <module>
    main()
  File "/home/lvhaoran/.local/bin/convert", line 180, in main
    args.func(args)
  File "/home/lvhaoran/.local/bin/convert", line 31, in to_caffe
    args.input, prototxt=args.prototxt, caffemodel=args.caffemodel, outspec=outspec
TypeError: tracedmodule_to_caffe() got an unexpected keyword argument 'outspec'

执行convert mge_to_onnx ... 报错

你好！
如题，报错如下：
convert: error: invalid choice: 'mge_to_onnx' (choose from 'onnx')
但是convert onnx ... 却可以， 请问知道是什么原因吗 ？
谢谢

无法转换一个简单的 UNet

环境

1.系统环境：ubuntu 18.04
2.MegEngine版本：MegEngine 1.1
3.python版本：3.6.9

复现步骤

见代码

请提供关键的代码片段便于追查问题

模型见
https://github.com/megvii-research/PMRID/blob/main/models/net_mge.py

先 dump 一个模型出来

from megengine import jit
net = Network()
net.eval()
img = mge.tensor(np.random.randn(1, 4, 64, 64).astype(np.float32))

@jit.trace(capture_as_const=True)
def f(x):
    out = net(x)
    return out

f(img)
f.dump(
    "net.mge",
    arg_names=['img'],
    optimize_for_inference=False,
    output_names=['pred'],
)

请提供完整的日志及报错信息

在将 #4 自己修掉的情况下运行 megconverter:

python3 -m mgeconvert.utils.convert_onnx  -i net.mge -o net.onnx

Traceback (most recent call last):
  File "/usr/lib/python3.6/runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "/usr/lib/python3.6/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/home/wangyuzhi/.local/lib/python3.6/site-packages/mgeconvert/utils/convert_onnx.py", line 29, in <module>
    main()
  File "/home/wangyuzhi/.local/lib/python3.6/site-packages/mgeconvert/utils/convert_onnx.py", line 25, in main
    convert_to_onnx(args.input, args.output, graph_name=args.graph, opset=args.opset)
  File "/home/wangyuzhi/.local/lib/python3.6/site-packages/mgeconvert/onnx_converter/onnx_converter.py", line 120, in convert_to_onnx
    model = converter.convert()
  File "/home/wangyuzhi/.local/lib/python3.6/site-packages/mgeconvert/onnx_converter/onnx_converter.py", line 70, in convert
    unsupported_oprs
AssertionError: Operators {<class 'mgeconvert.mge_context.mge_op.ConvolutionBackwardDataOpr'>} are not supported yet

我不了解 ConvolutionBackwardDataOpr 这个 opr 的作用，凭我朴素的理解，我只需要得到这个网络的前向部分，跟 backward 应该没关系……

希望有大佬能贡献一个转换到 NCNN 模型的例子

这样我就可以在手机上折腾这些东西了

onnx 模型转换问题

WARNING: The argument dynamic_input_shape=True is not needed any more, onnxsim
can now support dynamic input shapes natively, please refer to the latest
documentation. An error will be raised in the future.

找不到cuda_fp16.h

mgeconvert1.0.2版本，onnx模型转mge，报错：

05 09:32:46[mgb] ERR error occurred in computing sequence; synchronizing all comp nodes and releasing vars now ...
RuntimeError: nvrtc compile error: default_program(2): catastrophic error: cannot open source file "cuda_fp16.h"

系统里cuda_fp16.h在/usr/local/cuda-10.1/targets/x86_64-linux/include 这个目录下，请问要如何设置能让mgeconvert找到这个头文件？

ConvolutionBackwardDataOpr 初始化错误

https://github.com/MegEngine/mgeconvert/blob/master/mgeconvert/mge_context/mge_op.py#L302

class ConvolutionBackwardDataOpr(MgeOpr):
    name = "ConvolutionBackwardData"

    def __init__(self, opr):
        super().__init__(opr)
        self.kernel_shape = get_shape(opr.inputs[1])
        self.data_format = opr.params["format"]
        self.dilation_w = opr.params["dilate_w"]
        self.dilation_h = opr.params["dilate_h"]
        self.pad_w = opr.params["pad_w"]
        self.pad_h = opr.params["pad_h"]
        self.stride_w = opr.params["stride_w"]
        self.stride_h = opr.params["stride_h"]
        self.sparse = opr.params["sparse"]

这里应该使用 self.params

为什么命名是以 mge 开头？

meg 和 mge 的区别是啥呢

是否支持多输入多输出的mge到onnx转化

是否支持多输入多输出的mge到onnx转化

是否有计划提供其他框架到MegEngine的转换器

如题，方便使用MegEngine实现推理部署端的加速

.pkl转.onnx有问题

想将CVPR2021的NBNet训练的模型转为.onnx，报错，是否目前仅支持部分网络？

Megengine 1.7.0+cu112