search

enazoe / yolo-tensorrt

TensorRT8.Support Yolov5n,s,m,l,x .darknet -> tensorrt. Yolov4 Yolov3 use raw darknet *.weights and *.cfg fils. If the wrapper is useful to you,please Star it.
MIT License
1.19k stars 314 forks source link

batchsize 只能设置为1吗 #6

Open zhangxiaopang88 opened 4 years ago

zhangxiaopang88 commented 4 years ago

您好，我想修改测试的batchsize，只是修改_p_net = std::unique_ptr{ new YoloV3(1, _yolo_info, _infer_param) };这里好像不行，请问还有别的地方需要修改吗 @enazoe

enazoe commented 4 years ago

https://github.com/enazoe/yolo-tensorrt/blob/5ffe5869c50868c24193d317fa0be6b9d8f8e995/modules/class_yolo_detector.hpp#L84
这里和相关的地方

zhangxiaopang88 commented 4 years ago

好像增大batchsize会提高速度，我准备试试，但是还是有一些问题 @enazoe

richjjj commented 4 years ago

我试了下batchsize=2，yolov3可以正常推理，而yolov4只能得到第一张图片的结果，第二张结果为空。耗时为batchsize=1的两倍左右

jstumpin commented 4 years ago

yolov4-kHALF.engine (inputSize= 320x320, batchSize = 2) is working fine as depicted below: yolov4-batch2

yolov3-tiny-kHALF.engine (inputSize= 608x608, batchSize = 2) is also working fine as depicted below: yolov3-tiny-batch2

But yolov4-tiny-kHALF.engine (inputSize= 608x608, batchSize = 2) is working fine for the first image (batchId = 0) and not for the second one (batchId = 1): yolov4-tiny-batch2

enazoe commented 4 years ago

@jstumpin @zhangxiaopang88 @chongzhong sorry, batch inference not support yet ,I will working on it and support soon. And welcom PR at same time.

jstumpin commented 4 years ago

Everything's fixed now, kudos @enazoe for the prompt action! yolov4-tiny-kHALF-batch2.engine: yolov4-tiny

yolov4-tiny-kHALF-batch2.engine (using person.jpg + dog.jpg example as per in result.jpg; slight difference due to yolov4 vs. yolov4-tiny): yolov4-tiny2

yolov4-tiny-kHALF-batch4.engine (add std::string batchSize; after yolo.h#L56, add _yolo_info.batchSize = "batch" + std::to_string(_config.n_max_batch); after class_yolo_detector.hpp#L138, change class_yolo_detector.hpp#L144 to _yolo_info.enginePath = dataPath + "-" + _yolo_info.precision + "-" + _yolo_info.batchSize + ".engine";, and add config_v4_tiny.n_max_batch = 4; after sample_detector.cpp#L37): yolov4-tiny3

enazoe commented 4 years ago

@jstumpin good idea，and welcome pr :)