已经按照https://www.yuque.com/mnn/cn/cvrt_windows的步骤 ，安装好了protobuf。
但是在进行编译模型转换工具步骤时报错如下：

步骤为：

cd tools/converter
mkdir build
cd build
cmake -G "Ninja" -DMNN_BUILD_SHARED_LIBS=OFF -DCMAKE_BUILD_TYPE=Release ..

想知道问题是什么，感谢~

Thank you for this great library,

I'm trying to use the pure flow approach, but I have not been able to understand how to do it looking at ExampleCpuSingle....
Could you please provide a sample code showing how to build and train the net in the example image ?

Thank you very much

• - slength
• - maxSize
• - size
• - snull
• - (!?)
• - (!)
• - index
• - index'
• - head'
• - shead'
• - last'
• - indexM
• - headM
• - sheadM
• - lastM
• - unconsM
• - unsnocM
• - slice
• - slice'
• - sliceM
• - sslice
• - sliceAt
• - `slic

In init function of 'ShuffleNet' class, when groups==1 the number of channels should be 576, not 567.

The correct case should be:

if groups == 1:
self.stage_out_channels = [-1, 24, 144, 288, 576]

The fp_reduction_latency benchmarks were the very first benchmark, optimization and primitive code tested in Laser.

Unfortunately it is currently very confusing.

It should be reorganized:

• 1. Multiple accumulators: https://github.com/numforge/laser/blob/af191c086b4a98c49049ecf18f5519dc6856cc77/benchmarks/fp_reduction_latency/reduction_bench.nim and https://github.com/numforge/laser/blo

We should probably use a custom function instead of the matrix warping, so that the last pixel in the original image corresponds to the last pixel in the new image.

It would be good to keep track of the requirements this project has in a machine readable format such as the common requirement.txt output. By using the requirements specifiers from pip you can then specify versions required for this project, this will make it easier for people to install the dependencies and t