Training on FLIC

[ekkravchenko]

Hi!

I tried to replicate your results on FLIC and got an error when slicing the label. I used the 'genJSON.m' as it is, 'genLMDB.py' with changes of 'FLIC' dataset location, and modified 'genProto.py' as mentioned in the tutorial. In 'genProto.py' I changed the number of stages to be 4, slice parameter to be 10 (# joints + background) and 'outCH' (# of output channel of the convolutional layer before loss) to be 10. When I try to train the network with obtained .prototxt the label of size 16x30x46x46 is getting sliced into 16x10x46x46 and 16x20x46x46 correspondingly, resulting in the following error:

F1010 13:48:06.188012 50300 euclidean_loss_layer.cpp:18] Check failed: bottom[0]->count(1) == bottom[1]->count(1) (21160 vs. 42320) Inputs must have the same dimension.
*** Check failure stack trace: ***
    @     0x2b7b002d6e6d  (unknown)
    @     0x2b7b002d8ced  (unknown)
    @     0x2b7b002d6a5c  (unknown)
    @     0x2b7b002d963e  (unknown)
    @     0x2b7af8d287da  caffe::EuclideanLossLayer<>::Reshape()
    @     0x2b7af8dd83ec  caffe::Net<>::Init()
    @     0x2b7af8dd9265  caffe::Net<>::Net()
    @     0x2b7af8de6fba  caffe::Solver<>::InitTrainNet()
    @     0x2b7af8de80bc  caffe::Solver<>::Init()
    @     0x2b7af8de83ea  caffe::Solver<>::Solver()
    @     0x2b7af8e00053  caffe::Creator_SGDSolver<>()
    @           0x40f52e  caffe::SolverRegistry<>::CreateSolver()
    @           0x408592  train()
    @           0x40601c  main
    @     0x2b7b0a61cb15  __libc_start_main
    @           0x40688d  (unknown)

Could you let me know what adjustments should I perform to make it work?

[hongge831]

I have the same problem， help

[hongge831]

@ekkravchenko hvae you solve the problem?

[dingruoxiu]

有人解决这个问题吗 ？

[1Rookiechang]

have you download the pretrained model？can you share it with me.
because the download site cannot open.thank you very much!
[email protected]

[centosrhel]

@shihenw I have encountered a similar problem. could you please provide some help about modifying related code? Thanks

[nzzfsw]

you only need to add "number_parts: 9" in transform_param