pre-train model

[linchaozhang]

Have you successfully trained the model?
can you sent me a pre-train model, i want to train a shufflenet-ssd with a pre-train model.

thanks!

[farmingyard]

@linchaozhang The model is still training right now.

[KeyKy]

@farmingyard what is your current accuracy of val?

[leochli]

Hi @farmingyard ,
I'm sort of confused about your stride case in your deploy prototxt. Why you don't shuffle the channel here?
thanks!

[farmingyard]

@LeoLee96 Do you mean the stride=2 case? Yeah, its my mistake, i have fixed it. Thanks.

[ananddb90]

+1 for pre trained model
@farmingyard thank you for your share :)
is it possible for you to share solver file also ?

[suzhenghang]

Linear lr seems to work well

[farmingyard]

@linchaozhang @KeyKy @LeoLee96 @ananddb90
The pre-trained model is released, please check ReadMe file.

[KeyKy]

THX!

[linchaozhang]

@farmingyard
Thanks a lot.
But i found some differences between prototxt and caffemodel, in 'resx5_conv1', num_output : 60 in your prototxt, and 240 in your model. i think it should be 60 for 'resx5_conv1'(because the number of
bottleneck channels to 1/4 of the output channels for each ShuffleNet unit, said in paper)

[farmingyard]

@linchaozhang
I think you may had a mistake，cause the model is trained with the deploy prototxt file.
BTW，how can you get the model's layer output？

[linchaozhang]

@farmingyard
I have found the real difference, missing the param "group" in resx5_conv1 and resx13_conv1 in your prototxt, and missing the shuffle_layer, i think you forget to update the prototxt.

[farmingyard]

@linchaozhang
Sorry for that I updated the deploy file 9 days ago, you may not follow it. Please check the update info.
fdb6e83

[123chengbo]

i use your model and deploy.prototxt for imagenet classification， but result seems to be totally wrong（result labels are all wrong）， i compare it to caffenet model、deploy.prototxt（result labels are right）。

[xchani]

Hi, @farmingyard
Did you do any data augmentation in training stage?

[farmingyard]

@xchani
Just crop and mirror.

[ghost]

@farmingyard Hi, thanks for your sharing, great job. And i just wonder where is your pre-trained model, you have mentioned that have released, but i just can not see anything~~I hope i have not missed anything

[yunren]

@farmingyard
Hi, thanks for sharing your great job.
And I just wonder where is your pre-trained model that you have mentioned above, but i just can not find it after having checked ReadMe file

[farmingyard]

@fujihuang @yunren
Sorry for some reasons that I can't release the pre-trained model any more.

[CrazySssst]

any one can send the pre-trained model to me ? i really need it, thanks for your help @linchaozhang @farmingyard @KeyKy @LeoLee96 @ananddb90

[ananddb90]

can you send me your mail-id

[yonghenglh6]

Hi, would u put up the val accuracy of your model, please?

[dingdashen]

Hi,ananddb90
can you send the pre-trained model to me? I want finetune it to VOC dataset.

[CrazySssst]

@ananddb90 this is my email : push_back_max@163.com , thank you !

[acdreamdpl]

any one can send the pre-trained model to me? I hope to finetune it on faster-rcnn. @farmingyard @linchaozhang @KeyKy @LeoLee96 @ananddb90 . my email acdreamdpl@163.com thanks

[ananddb90]

sorry for late reply .
https://bitbucket.org/ananddb90/shuffle_net/src

could you please share your experience after fine tunning on detection problem as I am also trying the same and yet not able to get good results

[CrazySssst]

thanks for you sharing.
i will share my results on SSD after finish my experiment

[7oud]

@ananddb90
Thanks for your sharing.
Can you put up the accuracy of your sharing ShuffleNet on ImageNet？

[zoufangyu1987]

@farmingyard @ananddb90 @linchaozhang @KeyKy @LeoLee96
Do you have 0.5_shufflenet and 0.75_shufflenet caffe pre-trained model，train.prototxt?
thank you !

[zoufangyu1987]

@farmingyard @ananddb90 @linchaozhang @KeyKy @LeoLee96
I need to run ShuffleNet on "ARM" or "DSP",
so that speed is the most important.
Do you have any suggestion?
Thank you very much !

[felixfuu]

@123chengbo I have the same problem with you , do you find the cause of this problem?Looking forward to your reply!

[zoufangyu1987]

@farmingyard
THANK YOU !

[felixfuu]

@farmingyard @ananddb90 Thank you~

[MissDores]

Dear friend, could you help send your pre-trained models to my email box, 13261520092@163.com? Thank you very much! I am also trying to implement detection task based on ShuffleNet and FRCNN. Have you achieved ideal results? @ananddb90

[zoufangyu1987]

@MissDores
You can download shufflenet.caffemodel from this network.
https://bitbucket.org/ananddb90/shuffle_net/src
Thanks for @ananddb90's sharing!

[MissDores]

@zoufangyu1987 @ananddb90
Thank you so much!

[nnop]

What was performance when you train the model? @ananddb90

[32l]

Hi,ananddb90, can you send the pre-trained model to me? liaoting@whu.edu.cn Thank you very much!@ananddb90

[ryancuicui]

@ananddb90 HI， i opened the website you provided,but i can not access the pre-train caffemodel.Can you provided the caffemodel again? THANKS!!!

[colinwubot]

Hi friend, your works is wonderful. Could you share your pre-trained models with my? Thank you very much! @ananddb90

[tzt101]

@ananddb90 Hi, Could you share your pre-trained models? tzt101@outlook.com Thank you very much!

[dlyldxwl]

@ananddb90 @zoufangyu1987 Thanks for your sharing ! But the link isn't useful, can you share the caffemodel again or send to me ? My email is 844518133@qq.com .
Thanks again！

[mameng1]

@ananddb90 @zoufangyu1987 Thanks for your sharing ! But the link isn't useful, can you share the caffemodel again or send to me ? My email is mameng2908071445@gmail.com .
Thanks again！

[ananddb90]

hello all,
apologies for the late reply. Also I am sorry to say that I am not allowed to make it public yet
I will write down once it will be public again :)

[zoufangyu1987]

@dlyldxwl @mameng1
I'm sorry to have kept you waiting.
You can download it by "https://pan.baidu.com/s/1eS8NOm2", password is "bcj6".

@ananddb90
Without your permission, I've shared it, and if you don't allow sharing, I can delete it.
Thanks to @ananddb90 again!

[dlyldxwl]

OK, i get it。
thanks again ~

[zoufangyu1987]

@dlyldxwl
You are welcome！

[mydear33000]

@zoufangyu1987
Thanks a lot!

[fujiaweigege]

Could anyone share train_val.prototxt and solver.prototxt of ShuffleNet, thanks a lot! 282550434@qq.com

[dlyldxwl]

@fujiaweigege I think the origin deploy.prototxt of shufflenet has some issue in scale, I add pad param in some layers. i retain the shufflenet_ssd train.prototxt,becase of network connection ,I can't upload file,you can send me your email.Of course I don’t even make sure it’s much accurate.

[fujiaweigege]

@dlyldxwl I got your email, besides I want train_val.prototxt and solver.prototxt, could you send me them. 282550434@qq.com. Thank you very much!

[Justin-Yuan]

great work!! I am working on a project on human-tracking robots with microcontroller and would like to try out the pretrained shufflenet. Can you send me the caffemodel and prototxt files. justin.yuan@mail.utoronto.ca. Thank you so much !

[cjerry1243]

The link is now broken.
Could any one send the pretrained model to me?
cjerry1243@gmail.com

[duan348733684]

@dlyldxwl can you send your sufflenet_ssd train.prototxt to my Email 348733684@qq.com ? thanks

[neelakanth]

@zoufangyu1987 I am not able to download "https://pan.baidu.com/s/1eS8NOm2". Link gets reset after every few KBs of download. Can you upload on dropbox or googledrive

If any of you have pre-trained model, corresponding proto, top1/5 error and test image list please upload.

[zoufangyu1987]

@neelakanth ,
How much is your e-mail address? I'll send it to you

[zoufangyu1987]

My email address is 305434863@qq.com

[neelakanth]

@zoufangyu1987 Thanks. Here is my email id neelakanth.shigihalli@gmail.com

[neelakanth]

@zoufangyu1987 I am able to use caffemodel that you sent. I took latest code from https://github.com/BVLC/caffe.git and copied shuffle layer code and depth wise convolution code from (https://github.com/farmingyard/caffe-mobilenet)
I tested for 50000 images of imagenet on NVIDIA GPU. I am seeing Top1 62.3% and Top5 84.2%. What Top1 and Top5 you got?

I see that @msnqqer (https://github.com/msnqqer/ShuffleNet) got 65.45%/86.38%. I haven't tried with msnqqer's caffe model yet.

[zoufangyu1987]

@neelakanth Sorry, I didn't train on imagenet data.
I only trained on my own dataset.
So I have no imagenet result.

[Changkou]

@ananddb90 Could you please share the pre-training model to my email 383457094@qq.com？ As
the Bitbucket declined my request. Thanks!!!