Problem with vl_nnconv

[TalhaMetu]

Hi all,

I am using mathconvnet 1.0 beta17. I can compile the toolbox correctly. Everytime i try to execure the vl_nnconv, i get two errors. 1) First error is due to "opts" is given as cell array at vl_simplenn, i get "Cell contents reference from a non-cell array object." When i changed the code in vl_simplenn in vl_nnconv it doesnt give this error anymore. 2) Second error is "The option name is not a string (argument number 8)" At opts argument i didnt changed anything. I get this errors even if i use basic examples at mathconvnet websites and examples at mathconvnet folder i get same errors.

This is very imporant for me. Thank you

[lenck]

Hi, I think this happens when you use vl_arparse from vlfeat, not from MatConvNet. The vlfeat version does not support nested structures... At least this is what happens to me, usually... You can check this with which vl_argparse and then eventually fix it by first calling the vlfeat vl_setup and then the MatConvNet vl_setupnn (so that MatConvNet paths have a priority).

[yhqlower]

@lenck exactly! I've had the same problem several days ago. Since I did not use the files in /vlfeat/toolbox/misc, I just removed this directory out of the matlab path.

[vedaldi]

Hi, yes sorry for that, we shock update vlfeat.

However, there is a super easy fix: cal vl_setup before vl_setupnn (this will use MCN version).

On 5 Jan 2016, at 11:14, Colin Yan <notifications@github.com mailto:notifications@github.com> wrote:

@lenck https://github.com/lenck exactly! I've had the same problem several days ago. Since I did not use the files in /vlfeat/toolbox/misc, I just removed this directory out of the matlab path.

— Reply to this email directly or view it on GitHub https://github.com/vlfeat/matconvnet/issues/376#issuecomment-168963447.

[TalhaMetu]

Hi all, Thank you for your time and answers. But the point is i dont have vlfeat and vl_setup. So i am a little confused :/

[lenck]

When it crashes and you run which vl_argparse what does it return?

[TalhaMetu]

\Pattern Recognitiom Term Project\Without Cuda Codes\matconvnet-1.0-beta17\matlab\vl_argparse.m

[TalhaMetu]

When i delete opts i get: An input is not a SINGLE array nor it is empty. error. Thank you

[lenck]

Can you try it with the latest master from a fresh directory? You can get e.g. here: https://github.com/vlfeat/matconvnet/archive/master.zip There has been some bugs, so it might have been fixed...

[TalhaMetu]

Hi, i've just tried but again get same errors. Besides if i do:nt write opts.disableDropout=false; inside vl_simplenn i get sme errors. Also vl_convnn gives: Cell contents reference from a non-cell array object. if i give opts as a cell array not like (:). Thank you

[lenck]

Can you post the whole error? Opts should not be a cell array, that does not sound like the original quick start example...

[TalhaMetu]

training: epoch 01: batch 1/ 5: Cell contents reference from a non-cell array object.

Error in vl_simplenn (line 263) res(i+1).x = vl_nnconv(res(i).x, l.weights{1}, l.weights{2}, ...

Error in cnn_weakly_label_train>process_epoch (line 283) res = vl_simplenn(net, im, dzdy, res, ...

Error in cnn_weakly_label_train (line 138) [net,stats.train] = process_epoch(opts, getBatch, epoch, train, learningRate, imdb, net) ;

Error in cnn_weakly_label (line 41) [net, info] = cnn_weakly_label_train(net, imdb, @getBatch, ..

Also my l looks like this :

[TalhaMetu]

Also same error with mnist example

[lenck]

Ah, I see... That may be a bad initialization - in the beta-17 the layer opts has been added to the network structure. You can fix it if you add:

net = vl_simplenn_tidy(net) ;

before calling the [net, info] = cnn_weakly_label_train(net, imdb, @getBatch, .. in cnn_weakly_label...

[lenck]

And the bug in cnn_mnist has been fixed in the master branch... It was exactly the same issue :)

[TalhaMetu]

But i use master branch and i get still same error at mnist :( When i used that code in my program it says :+1: An input is not a SINGLE array nor it is empty. he res(i).x and other stride etc is not empty

I am so sorry about occupying you. Thank you for your time and answers

[TalhaMetu]

Can this be about the compiler that use on MATLAB? It is Microsoft Windows SDK 7.1 (C++). Although i dont get any errors when i am compiling it

[vedaldi]

Hi, I should add that _tidy needs to be called only once after loading the network.

Basically, it will upgrade the network structure to the latest MatConvNet version. It also fills in default values for unspecified parameters, so it is a bit easier to create new simplenn models.

On 5 Jan 2016, at 16:25, Karel Lenc notifications@github.com wrote:

Ah, I see... That may be a bad initialization - in the beta-17 the layer opts has been added to the network structure. You can fix it if you add:

net = vl_simplenn_tidy(net) ;

before calling the [net, info] = cnn_weakly_label_train(net, imdb, @getBatch, .. in cnn_weakly_label...

---

Reply to this email directly or view it on GitHub: https://github.com/vlfeat/matconvnet/issues/376#issuecomment-169075235

[TalhaMetu]

For the ones who suffers from same problem with me: I am using PASCAL VOC2012 dataset. Even if error is "An input is not a SINGLE array nor it is empty." You should give your input res(i).x as single(res(i).x) . I think this is because of MATLAB and GPU data type differencies but i am not that sure.

Thank you all guys for your valuable time and your study :)

[liquidmetal]

Was this ever resolved? I'm running into this exact issue.

[zgplvyou]

the same issue to you！！How to solve it？

[k8280627]

TalhaMetu adding the single() works for me!!

[OluwoleOyetoke]

Hi @k8280627. I am having a similar challenge here. Whenever I try to convert my DAG structure tos simplenn -----net_dag= vl_simplenn_tidy(net_dag) ;----, I get the error 'Cell contents reference from a non-cell array object'....Do you have a solution to this problem?

[k8280627]

@OluwoleOyetoke I have never had this problem, but from the error message, is it possible that your net_dag is not a cell-array? could you convert it to a cell-array?

[ghazalee70]

I am also getting the same error although all the inputs and parameters are gpu arrays: input{1}: 256 256 3 10 params{1}: 7 7 3 64 params{2}: 64 1 Error using vl_nnconv An input is neither SINGLE or DOUBLE nor it is empty. Can anyone please help me regarding this issue?

[hhamin]

@OluwoleOyetoke Ihave the same problem when I've tried using cnn_vgg_faces.m. the problem is due to that net.layers have some empty structures. I'll try remove these empty structures.

[OluwoleOyetoke]

Okay, @hhamin . I hope the problem is solved now