JensenHJS

希望作者的回复

1

量化部分的两个文件夹的代码，如果用单张显卡，每次在一个epoch之后保存模型，显存会暴涨，导致显存不够（单张12g或者6g都不够用），多张显卡才能正常训练。有办法解决这个问题吗

Hello，did you try to implement mobilenet-yolo through tensorRT，I search tensorRT in issues but I didn't find anything about tensorRT.

已经将caffe模型转换为int8的ncnn模型，那么请问除了手机端、海思芯片外，带GPU的硬件平台不能验证有加速吗？ncnn现在是支持tx2这类英伟达嵌入式GPU的，但是通过您的转换方式在英伟达GPU上是否会有加速？

use onnx model to inference

4

I can provide the onnx model. Have you planned to support onnx models with yolov3 currently? it's original model is yolov3.weights, using python API convert it to yolov3-608.onnx,then convert onnx...

whether the sample is the same as the official sample? Have you ever run official sample? I try with different version of onnx=1.2.1 1.2.2 1.4.1. But all of them raise...

about inference time

1

I haven‘t look at code detailedly. But is the inference time the same as darknet frame? In official darknet frame,it indeed costs about 300ms to predict one picture.But your unquantized...

模型转换遇到错误 error

23

将自己的sim-onnx模型转caffe模型时遇到如下错误： 另外还想问一下，博主的caffe测试版本是否跟BVLC的caffe版本有所区别？ {'images': 3, '167': 3, '168': 32, '169': 32, '170': 32, '171': 64, '172': 64, '173': 64, '174': 32, '175': 32, '176': 32, '177': 32, '178': 32, '179': 32,...

有没有人讨论下作者给出的量化论文DoReFa-Net？感觉看了不是很好理解。

Is it accelerate the inference