anklebreaker

anklebreaker

Member Since 3 years ago

Experience Points
1
follower
Lessons Completed
0
follow
Lessons Completed
10
stars
Best Reply Awards
1
repos

7 contributions in the last year

Pinned
Activity
Jan
15
4 days ago
Activity icon
issue

anklebreaker issue comment alibaba/MNN

anklebreaker
anklebreaker

转换tflite2模型提示SPLIT_V OP不支持

平台(如果交叉编译请再附上交叉编译目标平台):

MacOS:

Github版本:

1.2.1

报错:

Start to Convert Other Model Format To MNN Model...
[16:07:05] /Users/megatronking/Workspace/MNN/tools/converter/source/tflite/liteConverter.cpp:239: Check failed: creator ==> NOT_SUPPORTED_OP: [ SPLIT_V ]
[1]    79070 segmentation fault
anklebreaker
anklebreaker

I had the same issue and was able to convert from TFLite -> ONNX -> MNN. However, the converted model has several permute operations to keep the split axis the same. I hope in a future release that the conversion can convert to a TFLite split_v function along the proper axis in MNN's NCHW format.

Jan
14
5 days ago
Activity icon
issue

anklebreaker issue comment alibaba/MNN

anklebreaker
anklebreaker

About MNN Input. If my model needs a 4 channels Image like this (1, h, w, 4), 32 bit float data, what should I do to deliver it to the input tensor in C++?

anklebreaker
anklebreaker

@MingmChen I have same error, did you ever find a solution?

Activity icon
issue

anklebreaker issue alibaba/MNN

anklebreaker
anklebreaker

Noisy output with low precision inference (--fp16 ARM82)

Hello, I am getting very noisy output results when using low precision backend on cpu compiled with ARM82 (See below for heatmap prediction result). The issue is not there when switching to either Metal backend or using CPU with normal precision.

Version: 1.2.0 Platform: Mac OSX Compile Command: cmake -DMNN_METAL=ON -DMNN_COREML=ON -DMNN_ARM82=ON -DMNN_BUILD_CONVERTER=true ..

Screen Shot 2022-01-14 at 3 18 10 PM
Dec
9
1 month ago
Activity icon
issue

anklebreaker issue moble/quaternion

anklebreaker
anklebreaker

osx-arm64 conda

Would it be possible to release a build for osx-arm64 for conda? Currently cannot install via conda using existing channels on a Mac M1 Apple Silicon

Oct
29
2 months ago
started
started time in 2 months ago