add mnn inference

[syshensyshen]

mnn framework for face detect and landmark

[rlczddl]

为什么mnn版本无法用opencl或者cuda后端推理 程序会直接崩溃 有些层不支持吗

[syshensyshen]

1. mnn op kernel cuda/opencl implementation may be not enough when onnx model has dynamic input shape,
2. you can try long-side or short-side alignment resize input image, fixed pytorch export onnx model input size，onnx-simplifer model, delete some dynamic shape slice infer operations like permute/gather/scatter

[rlczddl]

感谢回复,我去试下。我注意到了导出的onnx模型输入是320,然后推理用的256。

[rlczddl]

你好,我找到原因了,与mnn无关,原因在于https://github.com/syshensyshen/Face-Detector-1MB-with-landmark/blob/ec64e16f41ef25e9fcfc9e576dc9fdbd8c5c24db/Face_Detector_mnn/reatinaface.cpp#L175, 这边应该传入的拷贝回来的host变量,因为你用的是cpu推理所以getSessionOutput回来的也是cpu所以传入没错,换成opencl后端自然就出错了

[syshensyshen]

it is my fault，thank you！