tensorrt c++ And mask decoder can be fp16 without too big error?
Mediumcore opened this issue · comments
Hello, Thank you for sharing the great work. Do you have a plan for tensorrt c++ version? That will be awsome too,Thank you very much.
I see mask decoder used with fp32, does fp16 tensorrt will produce big accuracy loss or not? Best regards.
Hi @Mediumcore ,
Thanks for reaching out!
We don't have a plan currently to add C++ support. But I'll update this thread if this changes.
I don't recall the exact decision to use FP32 with the mask decoder, but I think accuracy was the reason. I'm not 100% sure. To check, the accuracy difference could be verified by running the evaluation instructions.
John
Thank you for the great work you have done and look forward to your C++ initiatives
嗨,
感谢您伸出援手!
我们目前没有添加 C++ 支持的计划。但是,如果情况发生变化,我会更新此线程。
我不记得将 FP32 与掩模解码器一起使用的确切决定,但我认为准确性是原因。我不是100%确定。为了进行检查,可以通过运行评估指令来验证精度差异。
John