hellbell / adnet Goto Github PK

Hi, @hellbell ,

When will the training code be released?

Thanks~

When i train the net , i meet this error "FILTERS are larger than the DATA (including padding)." which makes me very confused. Can you tell me how to solve it? Thank you vert much!

Is there any tensorflow version for training RL or SL based AD net.

Hi, I meet this error
'Undefined function or variable 'test_demo''
when i use the command 'adnet_demo' to test the demo

How can i solve it ?

Excuse me,I meet a problem:
When I run the 'adnet_compile' append this mistake:
This code is run on Win10 64bit, MATLAB 2016a, and Cuda-9.1 with NVIDIA GTX 970M

Misuse of mex
MEX cannot find the library 'nppi' specified with the -l option.
  MEX finds a file with one of the following names:
  Nppi.lib
  Libnppi.lib
  Use the -L option to specify the path to this library.

Error build_cropRectanglesMex_on_windows (line 39)
Mex(mopts{:}) ;

Error run (line 96)
Evalin('caller', [script ';']);

Error adnet_compile (line 23)
     Run build_cropRectanglesMex_on_windows.m

Could you help me? Thanks!

Does anyone test the results with the updated model? My result on the OTB-100 is 0.574, but the result in the paper is 0.646. Does the test environment affect the result so much?

I executed adnet_train.v code and get net_basic.mat. But adnet_demo.m (including adnet_test function) does not work. It seems the own trained model's layer compositions are different with your given model 'net_rl.mat'. How to fix this error?

The models folder does not contain files like net_init.mat, imagenet-vgg-m-conv1-3.mat, etc. Please upload the files to help us reproduce the training results.
Thank You.

Could you upload the script of test_demo ? thanks.

I'm new to deep learning. I would like to know how this code works and how to train this code in matlab. Thanks in advance.

I think uploaded model is not same with your latest model

model's layer is not same with that you explain in paper and also accuracy is lower than paper's.

would you please check your model and re-upload. I want your model for comparing with my one in same experiment condition.

Thank you for your team's such meaningful work! I have a question, why don't you split the adnet into the two nets: policy net and confience net to train repectively? After the sl step, you use the rl to train the fc1-fc6 except for the fc7 and in online adaptation step, you use the fc7's parameter directly. Don't you worry about that the parameters' change in rl step may lead the confidence branch of the net less accurate?

Thanks for your great work!
I'm confused about the training of RL, in the code it seems you use the "accumulate_gradients_dagnn.m" to update parameters for RL, which is the same with SL stage. However, in the paper you stated that use "stochastic gradient ascent" to maximize tracking score. Can you help to explain it? Appreciate for the help!