Hi, firstly thank you for your wonderful work. I have a question about model structure,
Digraph.gv.pdf
I use two tools to view the structure which backbone is resnet50, in the first one which is constructed by the code(sorry, i forget the name of tool), and the second one is constructed by netron using yolact_resnet50_54_800000.pth. And the other two are the part of the picture. My doubt is that there are two fpn.downsample_layers in Digraph while there are three fpn.downsample_layers in yolact_resnet50_54. Why they are different? (I did not modify the parameters in config)
dlllll-q
All 11 comments
Ah yeah good catch. That's because our original architecture had 3 downsample layers but we found the final downsample layer to be doing absolutely nothing (removing it after training didn't affect the mAP but increased the speed slightly), so we removed it. If you train a new network from scratch it won't have this extra downsample in the state dict, so that's just there as a remnant of older code.
dbolya
on 17 Dec 2019
oh, thank you for your reply!
dlllll-q
on 17 Dec 2019
@dlllll-q can you please share the name of the tool used to view this model architecture , this would be helpful . THanks in advance
abhigoku10
on 18 Dec 2019
@abhigoku10 The first tool to use the pytorch code to construct the architecture is called pytorchviz, and the second one use the weights to view the architecture is called netron. Hope it will be helpful for you.
xuchengggg
on 19 Dec 2019
@abhigoku10 The first tool to use the pytorch code to construct the architecture is called pytorchviz, and the second one use the weights to view the architecture is called netron. Hope it will be helpful for you.
Hi, your works are excellent. Can you tell us the way how to use pytorchviz to visualise the structure or just share the code？So many thanks!
lbq779660843
on 6 May 2020
@lbq779660843 i did not use pytrochviz , i used pyan to obtain the graph . you can check it out https://github.com/davidfraser/pyan
abhigoku10
on 6 May 2020
@abhigoku10
Hi, I've loaded an onnx model by using the way of https://github.com/Ma-Dan/yolact in netron succefully.
I found that we seem have the same goal according to your judgements in other issues, so can u add me in your social network(wechat, QQ, etc.) to discuss some problems?
Thanx！
lbq779660843
on 13 May 2020
@lbq779660843 sure please share your wechat number i shall get connected to u
abhigoku10
on 13 May 2020
@abhigoku10
My wechat:
lbq920506
i am not able to find you on wechat can you please check the number
abhigoku10
on 13 May 2020
@abhigoku10
I'm sure of it, plz check it again!
lbq779660843
on 14 May 2020
@lbq779660843 hi i have sent a request to you yesterday only !! do accept it
abhigoku10
on 14 May 2020
Related issues
sree3333
·
10Comments
explorer359
·
10Comments
Epiphqny
·
6Comments
breznak
·
7Comments
mrtechnoo
·
4Comments
Most helpful comment
@abhigoku10 The first tool to use the pytorch code to construct the architecture is called pytorchviz, and the second one use the weights to view the architecture is called netron. Hope it will be helpful for you.