MobileNet-SSD: Assertion `min_arg < max_arg' failed.

[Dudule33]

Hello, i am running MobileNet-SSD train and test with lmdb data, to fine tune an SSD model. I get the following error message, when running with a debug compiled caffe version, CPU only: include/boost/random/uniform_real_distribution.hpp:159: boost::random::uniform_real_distribution<RealType>::uniform_real_distribution(RealType, RealType) [with RealType = float]: Assertion `min_arg < max_arg' failed.

This appends, i believe, in the BasePrefetchingDataLayer thread.

I try to compile/run with Python2 and Python3, without any change. I try with my own lmdb data, but also with downloaded lmdb data, without any change too.
I change also boost version from 1.65 to 1.67, without any effects.

The behaviour changes when modifying some solver.prototxt parameters, like test_initialization (true or false). It seems that, depending on solver parameters, training is running correctly for less or more iterations, but always ending with the same error. Anyone got this problem? Thanks for help

[jtdutta1]

Were you able to solve this issue? I'm still stick in this. I got this after i tried to resolve the Check failed: a <= b (0 vs. -1.19209e-07) by modifying the math_functions.cpp code as reported by various sites and issues in github. But now i run into this error. So i was curious to see if you have solved it?

[Dudule33]

Hello Jett,

In a very simple manner, i can say that i have modified the math_functions.cpp code like that:

/template <typename Dtype>//
//void caffe_rng_uniform(const int n,  Dtype a,  Dtype b, Dtype* r) {//
//  CHECK_GE(n, 0);//
//  CHECK(r);//
/*/  if (a<b)//
//  { //
//   a=b;//
//  }//
/*/***// CHECK_LE(a, b); *_
_/

But i'm not sure that it will solve your problem./_ _/

In my remembers, i got this issue during the learning phase of my model (fine tuning).

I have met a lot of problems before succeeding to train my model, especially because of software version.

I hope that my message help you.

Let me know about it, may be i could give you better idea.

Regards.

Alain.

Le 29/04/2019 à 07:16, Jeet Dutta a écrit :

Were you able to solve this issue? I'm still stick in this. I got this after i tried to resolve the |Check failed: a <= b (0 vs. -1.19209e-07)| by modifying the math_functions.cpp code as reported by various sites and issues in github. But now i run into this error. So i was curious to see if you have solved it?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/chuanqi305/MobileNet-SSD/issues/112#issuecomment-487453939, or mute the thread https://github.com/notifications/unsubscribe-auth/AJ7FHFV4OEXEKOAFFXAQ6A3PS2AA7ANCNFSM4FMUMHGQ.

[jtdutta1]

I shall try this method and notify you if I succeeded it not. Thank you.

[jtdutta1]

May I ask if you rebuilt the caffe distribution again after these modifications?

[jtdutta1]

Hi i failed to notice that the parameters a and b are constants which is why it was continuously giving read only errors. I removed the consts as you showed in the code now I don't get these errors. Thank you!!