onnx-preview1

[Blinue]

This discussion was created from the release onnx-preview1.

[hooke007]

之前测试的时候，TRT构建引擎的时候，这可以明显缩短时间
--skipInference
builder_optimization_level=0

对紧凑架构似乎优化没有太大效率影响，我在2080上观察不到差别，可能需要弱点的gpu

关掉这些额外优化好像也可以
use_cuda_graph
use_cublas
use_cudnn

[Blinue]

我测试的时候 3 级和 5 级优化有明显区别。现在是 5 级优化，以后会提供选项来修改优化等级。我认为对于 TensorRT，性能应放在第一位，构建时间长点也是可以接受的，而且引擎会被缓存，一个模型通常只需要构建一次。

以后会在构建引擎/编译着色器时展示一个 UI，防止用户以为缩放失败了。

[BOYA1999]

牛

[beike623]

I am using the onnx build for real-time antialiasing. Some models I have created only allow for static onnx builds instead of the dynamic required for magpie. These models have a better image quality with acceptable speed. Would it be possible to add support for static models in the future?

[animeojisan]

Thanks for the great app! Are you considering an update to allow swapping or multiple combinations of onnx filters like shaders? I love this app so much, so please consider it.

[Blinue]

Thank you for your interest, I'm glad to hear this project has been helpful to you. These features will be included once the functionality is officially implemented, though it may take some time. The project has been on pause for a while, as maintaining it has consumed much of my free time, but I plan to resume updates soon.

[animeojisan]

Thank you for your time! I am very, very excited to hear that you are planning to update the application!