search

fabio-sim / LightGlue-ONNX

ONNX-compatible LightGlue: Local Feature Matching at Light Speed. Supports TensorRT, OpenVINO
Apache License 2.0
352 stars 31 forks source link

support for different sizes of inputs #83

Open GreatZeZuo opened 2 months ago

GreatZeZuo commented 2 months ago

Hi, I'm a newbie and have this question: lightGlue's python model can support image inputs of different sizes, but how come the converted onnx model only supports inputs like 512, 1024? What should I do if I want to achieve support for different sizes of inputs. I would appreciate it if you give a quick reply!

fabio-sim commented 2 months ago

Hi @GreatZeZuo, thank you for your interest in LightGlue-ONNX.

Dynamic shapes are supported. The default options in python dynamo.py export already assume dynamic image sizes.

If you're exporting with a static input shape, then the only restriction imposed is that both H and W be integer multiples of 8. Technically this isn't a required condition, it's just something I noticed (that SuperPoint actually truncates to the nearest smaller multiple of 8 internally, so I figured it'd be better to expose this restriction to the user). Plus, generally certain operations like conv have better performance when inputs are clean multiples

GreatZeZuo commented 2 months ago

Appreciate your quick reply. BTW, I'm trying C++ compilation right now and I'm having problems：ONNXRuntime environment created failed : Could not find an implementation for MultiHeadAttention(1) node with name '/matcher/transformers.0/self_attn/MultiHeadAttention. Do you know how to solve it? Thanks, again.

fabio-sim commented 2 months ago

MultiHeadAttention is a contrib op available for ORT CPU & CUDA Execution providers. I think it's an issue with ORT version/platform