PartialResidualNetworks
PartialResidualNetworks copied to clipboard
WongKinYiu
Pretrained Model
@WongKinYiu , and I have a question about your paper:
Can you explain this picture for me?
G1 Timestamp t = 1 start at Layer 5 right?
And i don't know why G2 at Layer 2 in t =2. I don't see any connection between L4 and L2.
shortcut/concatenate also will receive the gradient at the same time. so G1 start at layers 1,3,5 (above connections are shortcuts). then G2 will be transmitted to layers 2,4 (5➔4, 3➔2, 1➔).
@WongKinYiu , Can you share darknet-53-prn and darknet53-fprn pretrained model on Imagenet for me ?
@CuongNguyen218 Hello, the model files are too big to upload to github. and i do not have enough cloud drive space.
@WongKinYiu , I have a unlimited google drive account. Please give me email address and i will send you my account. I need your pretrained model on imagenet for my bachelor thesis. Thank!
https://github.com/WongKinYiu/PartialResidualNetworks/blob/master/pdf/iccvw-prn.pdf
the first one is my e-mail address.
I got your pretrain on ImageNet. How can I used it for Yolov3 (convert to darknet53-conv74 format)
Sent from Mailhttps://go.microsoft.com/fwlink/?LinkId=550986 for Windows 10
From: Kin-Yiu, Wongmailto:[email protected] Sent: Monday, December 9, 2019 4:44 PM To: WongKinYiu/PartialResidualNetworksmailto:[email protected] Cc: Nguyen Ngoc Cuong 20150510mailto:[email protected]; Mentionmailto:[email protected] Subject: Re: [WongKinYiu/PartialResidualNetworks] Pretrained Model (#9)
i just upload the pre-trained weights.
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHubhttps://github.com/WongKinYiu/PartialResidualNetworks/issues/9?email_source=notifications&email_token=AKKXRBSNAFJGUZWOX36BAPLQXYHRFA5CNFSM4JXZZLN2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEGIPULA#issuecomment-563149356, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AKKXRBQ7HMVWWEHXB3TBHHTQXYHRFANCNFSM4JXZZLNQ.
https://github.com/AlexeyAB/darknet#how-to-train-tiny-yolo-to-detect-your-custom-objects
using partial.
@WongKinYiu ,
I'm training yolov3 with your darknet53-prn weights and get some -nan on log. is it normal ?
@CuongNguyen218
It seems there are no small objects in your own dataset.
@WongKinYiu ,
Can you explain the image above for me? Somethings wrong with the number of channel in the middle feature map. For example, 96 = 128 - (1 - 1/2)/2 *128 is not (1- gamma) / 2 * c in the picture. and why 1/8 appear in the red line below ?
there r 1/2/8/8/4 layers in 1/2/3/4/5 stages.
and...
so in 3rd stage, c is 256 and γc is 128. the channel number of 8 layers should be 240, 224, 208, 192, 176, 160, 144, 128.
there r 1/2/8/8/4 layers in 1/2/3/4/5 stages. and...
so in 3rd stage, c is 256 and γc is 128. the channel number of 8 layers should be 240, 224, 208, 192, 176, 160, 144, 128.
I know it when read your paper but I think your figure of characteristic 2 lead to misleading.
@WongKinYiu , I try replace darknet53 with darknet 53-prn but my inference time is increase about 20% from 45 FPS to 55 FPS. How did you do to reduce inference time from 22ms to 11ms
55 FPS is faster than 45 FPS. The inference time is decrease, not increase...
and did u check the blops of ur implementation is same as our reported bflops (25.1 BFLOPs)?
@CuongNguyen218 Hello, Can you share this link of darknet53-PRN pretrained weights ?
@CuongNguyen218 Hi, can you share your pretrained weights of darknet53-PRN with your google drive link? @WongKinYiu Or can you send the pretrained weights of darknet53-PRN to my email? My email address is [email protected] Thanks!
