Open Tower0823 opened 1 year ago
trans the C++ op MultiScaleDeformableAttnFunction to pytorch code, then solved the problem. But the exported output .onnx/.pt only support the fixed batch size like 1,2,4,8... Is there anyone know how to support dynamic input and output batch?
Can you please share your code to export grounding-dino to onnx? I guess this fixed batchSize is according to gpu nums?
老哥,可以分享一下你导出onnx模型的代码不?万分感谢
trans the C++ op MultiScaleDeformableAttnFunction to pytorch code, then solved the problem. But the exported output .onnx/.pt only support the fixed batch size like 1,2,4,8... Is there anyone know how to support dynamic input and output batch?