Clarification about schematic and Receptive field calculation?

Question

Clarification about schematic and Receptive field calculation?

sagarhukkire opened this issue 8 years ago · 16 comments

I was going through schematics and pototxt file. Please correct me if I am wrong

Input volume 128 x 128 x 64 with 1 channel

VNet uses 16 convolution filter with 5 x 5 x 5 , to get original volume there is padding of zeros with size 2 ,totally fine . Then sub sampling of 2 x 2 x 2 ,then volume size to next stage is 64 x 64 x 32
Here VNet uses 32 channels again same size of kernel in figure you showed two convolution layer are they 32 channels with kernel size 5 x 5 x 5, 2 times convolution layer ?
if it is then for next stage 64 channels for 3 times right?

now important thing how you calculate receptive size , you can explain one of them so I will clear myself remaining. I am confused do I consider kernel size or theoretical size 3 x 3 x 3

Thanks for VNet its indeed great working !!

Sagar

Answer 1 · 2017-04-08T07:30:19.000Z

hi Sagar
how do you make the dataset sed you network?

Answer 2 · 2017-04-08T20:58:26.000Z

in the paper i must have forgotten to update the caption of that table…

…

On Mar 31, 2017, at 3:00 AM, sagarax009 ***@***.***> wrote: Hi @faustomilletari <https://github.com/faustomilletari> I was going through schematics and pototxt file. Please correct me if I am wrong Input volume 12812864 with 1 channel VNet uses 16 convolution filter with 555 , to get original volume there is padding of zeros with size 2 ,totally fine . Then sub sampling of 222 ,then volume size to next stage is 646432 Here VNet uses 32 channels again same size of kernel in figure you showed two convolution layer are they 32 channels with kernel size 555, 2 times convolution layer ? <https://cloud.githubusercontent.com/assets/20017611/24539702/093036ba-15f0-11e7-9a18-b68800b9401a.png> if it is then for next stage 64 channels for 3 times right? now important thing how you calculate receptive size , you can explain one of them so I will clear myself remaining. I am confused do I consider kernel size or theoretical size 333 <https://cloud.githubusercontent.com/assets/20017611/24539757/45a82530-15f0-11e7-8879-f0cc1a26bcac.png> Thanks for VNet its indeed great working !! Sagar — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#24>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AMtsvpA5maBiFROuzHJYCXLQ4JwqqstDks5rrKSAgaJpZM4MvSyF>.

Answer 3 · 2017-04-08T22:36:10.000Z

Hi Fasto Can you please send me the link so I can figure out and send me bibtex for Vnet , so I can cite it in my report. Thanks and Regards Sagar Hukkire On Apr 8, 2017 10:58 PM, "Fausto Milletari" <notifications@github.com> wrote:

…

in the paper i must have forgotten to update the caption of that table… > On Mar 31, 2017, at 3:00 AM, sagarax009 ***@***.***> wrote: > > Hi @faustomilletari <https://github.com/faustomilletari> > I was going through schematics and pototxt file. Please correct me if I am wrong > > Input volume 12812864 with 1 channel > > VNet uses 16 convolution filter with 555 , to get original volume there is padding of zeros with size 2 ,totally fine . Then sub sampling of 222 ,then volume size to next stage is 646432 > > Here VNet uses 32 channels again same size of kernel in figure you showed two convolution layer are they 32 channels with kernel size 555, 2 times convolution layer ? > <https://cloud.githubusercontent.com/assets/20017611/24539702/093036ba- 15f0-11e7-9a18-b68800b9401a.png> > if it is then for next stage 64 channels for 3 times right? > > now important thing how you calculate receptive size , you can explain one of them so I will clear myself remaining. I am confused do I consider kernel size or theoretical size 333 > <https://cloud.githubusercontent.com/assets/20017611/24539757/45a82530- 15f0-11e7-8879-f0cc1a26bcac.png> > Thanks for VNet its indeed great working !! > > Sagar > > — > You are receiving this because you were mentioned. > Reply to this email directly, view it on GitHub <https://github.com/ faustomilletari/VNet#24>, or mute the thread <https://github.com/ notifications/unsubscribe-auth/AMtsvpA5maBiFROuzHJYCXLQ4Jwqqs tDks5rrKSAgaJpZM4MvSyF>. > — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#24 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ATFxyzzqeShHjlyxr-UY5cJCaLd8Yh1Kks5rt_TzgaJpZM4MvSyF> .

Answer 4 · 2017-04-08T23:07:50.000Z

Hello, the link to what?? the bibtex is here: @inproceedings{milletari2016v, title={V-net: Fully convolutional neural networks for volumetric medical image segmentation}, author={Milletari, Fausto and Navab, Nassir and Ahmadi, Seyed-Ahmad}, booktitle={3D Vision (3DV), 2016 Fourth International Conference on}, pages={565--571}, year={2016}, organization={IEEE} } Regards, Fausto

…

On Apr 8, 2017, at 6:36 PM, Sagar Hukkire ***@***.***> wrote: Hi Fasto Can you please send me the link so I can figure out and send me bibtex for Vnet , so I can cite it in my report. Thanks and Regards Sagar Hukkire On Apr 8, 2017 10:58 PM, "Fausto Milletari" ***@***.***> wrote: > in the paper i must have forgotten to update the caption of that table… > > > > On Mar 31, 2017, at 3:00 AM, sagarax009 ***@***.***> > wrote: > > > > Hi @faustomilletari <https://github.com/faustomilletari> > > I was going through schematics and pototxt file. Please correct me if I > am wrong > > > > Input volume 12812864 with 1 channel > > > > VNet uses 16 convolution filter with 555 , to get original volume there > is padding of zeros with size 2 ,totally fine . Then sub sampling of 222 > ,then volume size to next stage is 646432 > > > > Here VNet uses 32 channels again same size of kernel in figure you > showed two convolution layer are they 32 channels with kernel size 555, 2 > times convolution layer ? > > <https://cloud.githubusercontent.com/assets/20017611/24539702/093036ba- > 15f0-11e7-9a18-b68800b9401a.png> > > if it is then for next stage 64 channels for 3 times right? > > > > now important thing how you calculate receptive size , you can explain > one of them so I will clear myself remaining. I am confused do I consider > kernel size or theoretical size 333 > > <https://cloud.githubusercontent.com/assets/20017611/24539757/45a82530- > 15f0-11e7-8879-f0cc1a26bcac.png> > > Thanks for VNet its indeed great working !! > > > > Sagar > > > > — > > You are receiving this because you were mentioned. > > Reply to this email directly, view it on GitHub <https://github.com/ > faustomilletari/VNet#24>, or mute the thread <https://github.com/ > notifications/unsubscribe-auth/AMtsvpA5maBiFROuzHJYCXLQ4Jwqqs > tDks5rrKSAgaJpZM4MvSyF>. > > > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > <#24 (comment)>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/ATFxyzzqeShHjlyxr-UY5cJCaLd8Yh1Kks5rt_TzgaJpZM4MvSyF> > . > — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#24 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AMtsvuEna4gB15Ibpa5ZGJ6t4fZzeL91ks5ruAvagaJpZM4MvSyF>.

Answer 5 · 2017-04-09T00:00:27.000Z

@faustomilletari

link for receptive field calculation ? I followed many articles like dilated convolution and all but no way I can get same numbers as your papers?

So it will be great if you just give me one explanation at any stage then I can figure out remaining calculation for receptive field

Thanks
Sagar

Answer 6 · 2017-04-09T00:03:16.000Z

@wxde

Which kind of images you have ,I mean medical or something else?

If medical then I recommend you to use MITK or any tool(MITK ) is great, save image and label (its image with area of body organ)
Either save it in .mhd format as Fausto said, since VNET uses sitk ,so either .mhd or .nrrd(i work with .nrrd) both are fine

all the best

Answer 7 · 2017-04-09T00:11:21.000Z

you have to use a kernel size 5x5x5. then when there is the downsampling the behavior is similar to a pooling 2x2x2. in the up sampling path, the deconvolutions also contribute to increase the receptive field! I calculated it once by hand and another time with matlab, through the package matconvnet.

…

On Apr 8, 2017, at 8:00 PM, Sagar Hukkire ***@***.***> wrote: @faustomilletari <https://github.com/faustomilletari> link for receptive field calculation ? I followed many articles like dilated convolution and all but no way I can get same numbers as your papers? So it will be great at one layer how you are able to get those dimension for receptive field Thanks Sagar — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#24 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AMtsvomaNmVYZey8Cpk5p5ErAl9Y5CS5ks5ruB-bgaJpZM4MvSyF>.

Answer 8 · 2017-04-09T02:14:43.000Z

@sagarax009
my 3D images is medical dataset , how to sent the .mhd farmat to vnet model?
from xu

Answer 9 · 2017-04-09T07:49:51.000Z

@wxde I guess,I got your problem Follow the steps 1) hope you have installed 3D caffe for Vnet ( do it till make runtest) 2) If you go to main.py you will see path to train,test,result,snapshot words ; just create folders acoording For example :/home / Sagar/ Train 3) then Vnet.py give path to caffe. You can use Import sys Sys.path.insert (0,"your caffe path ") 4) if you have made folder structure above and data properly placed then pass system parameters like train or test as per your need (At end of main.py you can see that) Or simply Google it, its easy you will find it 5) there you are, your network will start running Hope this helps Thanks and Regards Sagar Hukkire

…

On Apr 9, 2017 4:14 AM, "wxde" ***@***.***> wrote: @sagarax009 <https://github.com/sagarax009> my 3D images is medical dataset , how to sent the .mhd farmat to vnet model? from xu — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#24 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ATFxy58IhdM7cw03HggMYFG4NOxl2ikmks5ruD8UgaJpZM4MvSyF> .

Answer 10 · 2017-04-09T08:02:17.000Z

@sagarax009
I am glad to heard from you ,your reply is so detail , thank you very much
from wxde

Answer 11 · 2017-04-09T08:21:12.000Z

@Fausto I tried to calculate by hand correct if I am wrong 1) stage 1 , simply kernel size = filter size i.e. 5x5x5 2) Stage 2, stacking of two convolution layer with one down sampling Receptive field For example X direction (5+5-1) = 9 ( m+m-rank of filter)then there down sampling by 2 with stride 2 , how result is 22 here then. Thanks and Regards Sagar Hukkire you have to use a kernel size 5x5x5. then when there is the downsampling the behavior is similar to a pooling 2x2x2. in the up sampling path, the deconvolutions also contribute to increase the receptive field! I calculated it once by hand and another time with matlab, through the package matconvnet.

On Apr 8, 2017, at 8:00 PM, Sagar Hukkire ***@***.***>

wrote:

@faustomilletari <https://github.com/faustomilletari> link for receptive field calculation ? I followed many articles like

dilated convolution and all but no way I can get same numbers as your papers?

So it will be great at one layer how you are able to get those dimension

for receptive field

Thanks Sagar — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <https://github.com/

faustomilletari/VNet#24#issuecomment-292753933>, or mute the thread < https://github.com/notifications/unsubscribe-auth/ AMtsvomaNmVYZey8Cpk5p5ErAl9Y5CS5ks5ruB-bgaJpZM4MvSyF>.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#24 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ATFxy0KY3CbiTzHQRonDI9Mb3l6JuDOcks5ruCIqgaJpZM4MvSyF> .

Answer 12 · 2017-04-09T16:07:18.000Z

I think it’s done “per block”, not per convolutional stage. (per resolution block)

…

On 09 Apr 2017, at 04:21, Sagar Hukkire ***@***.***> wrote: @Fausto I tried to calculate by hand correct if I am wrong 1) stage 1 , simply kernel size = filter size i.e. 5x5x5 2) Stage 2, stacking of two convolution layer with one down sampling Receptive field For example X direction (5+5-1) = 9 ( m+m-rank of filter)then there down sampling by 2 with stride 2 , how result is 22 here then. Thanks and Regards Sagar Hukkire you have to use a kernel size 5x5x5. then when there is the downsampling the behavior is similar to a pooling 2x2x2. in the up sampling path, the deconvolutions also contribute to increase the receptive field! I calculated it once by hand and another time with matlab, through the package matconvnet. > On Apr 8, 2017, at 8:00 PM, Sagar Hukkire ***@***.***> wrote: > > @faustomilletari <https://github.com/faustomilletari> > link for receptive field calculation ? I followed many articles like dilated convolution and all but no way I can get same numbers as your papers? > > So it will be great at one layer how you are able to get those dimension for receptive field > > Thanks > Sagar > > — > You are receiving this because you were mentioned. > Reply to this email directly, view it on GitHub <https://github.com/ faustomilletari/VNet#24#issuecomment-292753933>, or mute the thread < https://github.com/notifications/unsubscribe-auth/ AMtsvomaNmVYZey8Cpk5p5ErAl9Y5CS5ks5ruB-bgaJpZM4MvSyF>. > — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#24 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ATFxy0KY3CbiTzHQRonDI9Mb3l6JuDOcks5ruCIqgaJpZM4MvSyF> . — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#24 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AMtsvjtXDgVR3RCCpxTLTnAA1Nb803qKks5ruJT5gaJpZM4MvSyF>.

Answer 13 · 2017-04-12T22:44:20.000Z

@faustomilletari

I got it in another paper where Author has cited Vnet ..haha its nice to understand. Yes its block wise

Answer 14 · 2017-09-13T08:18:29.000Z

Hi @sagarhukkire , were you able to obtain the same receptive field in the upsampling (deconvolutions) path? So far I have only been able to obtain the receptive field in the downsampling path..., I can't obtain the 476 in the R-stage 4

Answer 15 · 2017-09-13T08:32:36.000Z

Refer Appendix of paper "Medical image segmentation using CNN" There is formula which is needed for receiptive field calculation Thanks and Regards Sagar Hukkire

…

On Sep 13, 2017 10:18 AM, "Roger Trullo" ***@***.***> wrote: Hi @sagarhukkire <https://github.com/sagarhukkire> , were you able to obtain the same receptive field in the upsampling (deconvolutions) path? So far I have only been able to obtain the receptive field in the downsampling path... — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#24 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ATFxy6OU1ecklMFa3q9-_cPCxysdgnTMks5sh4_WgaJpZM4MvSyF> .

Answer 16 · 2017-09-13T08:49:06.000Z

Thanks @sagarhukkire !
EDIT:
I was able to reproduce the receptive field in the paper, I made a small script , I will share it later so other people struggling can use it.