how to tain the FCN on other dataset

[Jianzhong He]

hello everyone,
i'm a newer to the caffe, recently i want to run the fcn on the kitti dataset, but there are too problems.
when i set the num_output equals to 21, then follow problems occurs
Check failed: outer_num_ * inner_num_ == bottom[1]->count() (453620 vs. 465750) Number of labels must match number of predictions; e.g., if softmax axis == 1 and prediction shape is (N, C, H, W), label count (number of labels) must be N*H*W, with integer values in {0, 1, ..., C-1}.
and if i set it to 3 or other,another problems occurs

error==cudaSuccess (77 vs. 0) an illegal memory access was encountered

is there someone can help me?
thanks a lot.

[Grayson Adkins]

I'm having a similar issue. I started with the pascal context FCN-32s network (60-output) and ran into issues when changing the number of outputs to 2, 21 and 25; however when the number of outputs is 7, 11 or 60 it works...Seems strange.

Also, make sure that your image labels match the number of outputs of the network. For example, make sure that for a 21-output network (labels 0 to 20), you don't have pixel-wise labels of 21, 22, 23, etc.

This is not my issue, so I still don't have a solution.

[Vikraman K]

Hi, 

Where you able to train on a new dataset? If yes, could you please help me out. Its urgent.

Thanks,

K.Vikraman

[Mina Rezaei]

Dear Jianzhong

I would like to know more about data preparation and your dataset format,
Thanks in advance!