no more improve loss&accuracy. with hdf5 data

[labolas89]

first, i was try examples/hdf5_classification/nonlinear model. with after some modifications.

But the result of this model was not work when using the prediction.

this result likes, very high cost f(x) equal zero.

i know, I have added the option to be divided the accuracy for each of the labels in the source CAFFE.

ex) f(x) equal zero

I0229 15:12:19.077769 28263 solver.cpp:253]     Train net output #0: accuracy = 0.538889 [ 0 : 100% ] [ 1 : 0% ] [ 2 : 0% ] 

I0229 15:12:19.077790 28263 solver.cpp:259]     Train net output #1: loss = 0.917172 (* 1 = 0.917172 loss)

I0229 15:12:19.077800 28263 sgd_solver.cpp:106] Iteration 13500, lr = 0.001

I0229 15:12:21.739393 28263 solver.cpp:348] Iteration 14000, Testing net (#0)

I0229 15:12:21.958649 28263 solver.cpp:432]     Test net output #0: accuracy = 0.7165 [ 0 : 100% ] [ 1 : 0% ] [ 2 : 0% ] 

I0229 15:12:21.958739 28263 solver.cpp:438]     Test net output #1: loss = 0.77523 (* 1 = 0.77523 loss)

but, this code support only 3-label

Code has already been uploaded to the my github link.

i feed own my hdf5 data.

i makes dataset 1dim, 4 integer variable and 3 label. It's very small data

 

so. i use cifar10 based net.

here is github link. my net model, solver, hdf5 data folder

2d_train_test.prototxt

2d_solver.prototxt

new2d

this is my last train log

caffe.ubuntu.jace.log.2d.INFO

I0229 15:30:47.328443 28552 solver.cpp:253]     Train net output #0: accuracy = 0.85 [ 0 : 94.2688% ] [ 1 : 0% ] [ 2 : 91.4286% ] 

I0229 15:30:47.328470 28552 solver.cpp:259]     Train net output #1: loss = 0.38771 (* 1 = 0.38771 loss)

I0229 15:30:47.328481 28552 sgd_solver.cpp:106] Iteration 199500, lr = 1e-06

I0229 15:30:50.095059 28552 solver.cpp:499] Snapshotting to HDF5 file git/trained_data/2d_train_iter_200000.caffemodel.h5

I0229 15:30:50.096268 28552 sgd_solver.cpp:283] Snapshotting solver state to HDF5 file git/trained_data/2d_train_iter_200000.solverstate.h5

I0229 15:30:50.099409 28552 solver.cpp:328] Iteration 200000, loss = 0.308396

I0229 15:30:50.099468 28552 solver.cpp:348] Iteration 200000, Testing net (#0)

I0229 15:30:50.319504 28552 solver.cpp:432]     Test net output #0: accuracy = 0.869489 [ 0 : 94.6179% ] [ 1 : 0% ] [ 2 : 89.0563% ] 

I0229 15:30:50.319572 28552 solver.cpp:438]     Test net output #1: loss = 0.326559 (* 1 = 0.326559 loss)

I0229 15:30:50.319583 28552 solver.cpp:333] Optimization Done.

I0229 15:30:50.319591 28552 caffe.cpp:215] Optimization Done.

There have been performance improvements, i need more accuracy.

and then a simple form of such models now.

Because the feedforward-output system will be implemented as a circuit within the FPGA

what sure i do?

Could anyone help me or give me some advice? 

thanks.