rockyzhengwu / libtorch-yolov4

yolov4 implement by lobtorch
MIT License
10 stars 2 forks source link

Make failed #1

Open ly0303521 opened 4 years ago

ly0303521 commented 4 years ago

你好, 在Make 的时候出错了, libtorch 应该没有问题,因为之前我用来调用yolov3是成功的 Screenshot from 2020-07-14 17-14-47

jackft2 commented 4 years ago

I met the same problem. maybe, it's libtorch version problem. can you offer your environment? @rockyzhengwu

rockyzhengwu commented 4 years ago

@ly0303521 @jackft2 I use libtorch 1.5.1+cu101

ly0303521 commented 4 years ago

@rockyzhengwu @jackft2 problem solved, use libtorch 1.5.1+cu101, tks

ly0303521 commented 4 years ago

I met the same problem. maybe, it's libtorch version problem. can you offer your environment? @rockyzhengwu

have you solved the problem, just use the right version of libtorch, but when it comes to yolov4-tiny, new problem come out, what(): shape '[256, 384, 3, 3]' is invalid for input of size 695826, do you have any idea?

rockyzhengwu commented 4 years ago

I haven't test on yolov4-tiny, there is some difference between yolov4 and yolov4-tiny. such like route layer in yolov4 groups isn't exist , but in yolov4-tiny groups=2 at somewhere .
I will fix this as soon as i can. by the way i think this implementation is too simple or ugly , it's just parse the yolo config file and load the weight to tensor . may be exhibit define every layer in yolov4 is more clear just like Python version : https://github.com/Tianxiaomo/pytorch-YOLOv4

ly0303521 commented 4 years ago

I haven't test on yolov4-tiny, there is some difference between yolov4 and yolov4-tiny. such like route layer in yolov4 groups isn't exist , but in yolov4-tiny groups=2 at somewhere . I will fix this as soon as i can. by the way i think this implementation is too simple or ugly , it's just parse the yolo config file and load the weight to tensor . may be exhibit define every layer in yolov4 is more clear just like Python version : https://github.com/Tianxiaomo/pytorch-YOLOv4

OK, tks very mush

rockyzhengwu commented 4 years ago

@ly0303521 I have fix this , but maybe some about the yolo layer had to be update , because i just copy from the yolov3 .

ly0303521 commented 4 years ago

Tks for your reply, it works. And realy a great job

ly0303521 commented 4 years ago

Hi, I find a bug in darknet.cc, because the index of layer is different with the darknet, for example, when you get the index, which is postive, it is the absolute index, so here should be index + 1 Screenshot from 2020-07-16 17-26-48

rockyzhengwu commented 4 years ago

i'm not sure about this , i test on the result It seems ok

ly0303521 commented 4 years ago

i'm not sure about this , i test on the result It seems ok

it has problem when it comes to yolov3-tiny, this will fix it