Jiko

你好，关于sisnr loss, 我看你在loss文件里通过calloss调用了，我有个小疑问，我们训练的时候 要调用calloss还是sisnr？ 假设是调用calloss的话，我会遇到一个问题就是，我传入的参数是[batch, 1, signal_size]，会经常遇到 loss_total = loss_total / (output.shape[0] - zerocount) 除以0的问题。 假设直接调用sisnr，我传入参数是[batch, signal_size]，会导致 t = torch.sum(x_zm * s_zm) * s_zm / (l2norm(s_zm)**2 + eps) 除以维度不统一。

Another question is that, is there any dev and test set, which I cat get access? Or I have to split the dev from the train myself?

I tried to use the alignment tool, and I realised each audio length is about 15-25s. Is it possible to modify the script to align shorter length(

@lix19937 > What is the diff when you use all fp32 ? It is quite strange that the result of the model (onnxfp16->tensortfp32) is also totally different from the onnx...

@lix19937 > You can use follow to compare the diff between trt and ort. I have run both fp32 and fp16 `polygraphy run fp16.onnx --trt --onnxrt (--fp16) --execution-providers=cuda` fp16 is...

I fixed the problem with polygraphy debug precision Thank you.

> Can you share your model? Testing this workflow with https://github.com/onnx/models/blob/main/validated/vision/classification/resnet/model/resnet50-v1-12.onnx the output is as expected. The model is quite big, it is unable to upload here. any other way...

这个目前有解决吗

I also face a memory management problem that the results from both onnx and trt models are the same when I do htod_async. but the result is different when I...

> cuda.memcpy_dtod_async(self.inputs[i].device, model_input.data_ptr(), model_input.element_size() * model_input.nelement(), self.stream) > to > > cuda.memcpy_dtod_async(inp.device, model_input.data_ptr(), model_input.element_size() * model_input.nelement(), self.stream) Is there any difference? It is because in the original code, the i...