Convolution network for 1D data (time-series)

[Grigory V.Sapunov]

Hello,

I'm trying to apply Caffe to time-series 1D signal like temperature, Wolf number, etc.

I've created a data in hdf5, created network definition (mostly similar to MNIST one), run solver, and got an error:

I0120 01:49:05.769857  3623 net.cpp:67] Creating Layer data

I0120 01:49:05.770401  3623 net.cpp:356] data -> data

I0120 01:49:05.770464  3623 net.cpp:356] data -> label

I0120 01:49:05.770516  3623 net.cpp:96] Setting up data

I0120 01:49:05.770572  3623 hdf5_data_layer.cpp:57] Loading filename from train.txt

I0120 01:49:05.770745  3623 hdf5_data_layer.cpp:69] Number of files: 2

I0120 01:49:05.771117  3623 hdf5_data_layer.cpp:29] Loading HDF5 filesample_data_train.h5

I0120 01:49:05.788478  3623 hdf5_data_layer.cpp:49] Successully loaded 5000 rows

I0120 01:49:05.789269  3623 hdf5_data_layer.cpp:81] output data size: 10,1,1,300

I0120 01:49:05.789773  3623 net.cpp:103] Top shape: 10 1 1 300 (3000)

I0120 01:49:05.790540  3623 net.cpp:103] Top shape: 10 1 1 1 (10)

I0120 01:49:05.791241  3623 net.cpp:67] Creating Layer conv1

I0120 01:49:05.792071  3623 net.cpp:394] conv1 <- data

I0120 01:49:05.792824  3623 net.cpp:356] conv1 -> conv1

I0120 01:49:05.792940  3623 net.cpp:96] Setting up conv1

F0120 01:49:05.795249  3623 blob.cpp:13] Check failed: height >= 0 (-3 vs. 0)

Did I miss something? Are there any special issues when dealing with 1d convolutions?

Best regards,

Grigory

[Ran Manor]

I use Caffe with 1D convolutions, it works the same as the 2D one.
Make sure you use the separate kernel sizes, kernel_x and kernel_y, instead of kernel_size and that one of them is equal to 1.

[Grigory V.Sapunov]

Thanks, Ran! Exactly what I need.

(actually kernel_h and kernel_w)

On Tue, Jan 20, 2015 at 10:53 AM, Ran Manor <ran....@gmail.com> wrote:

I use Caffe with 1D convolutions, it works the same as the 2D one.
Make sure you use the separate kernel sizes, kernel_x and kernel_y, instead of kernel_size and that one of them is equal to 1.

--
You received this message because you are subscribed to the Google Groups "Caffe Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to caffe-users...@googlegroups.com.
To post to this group, send email to caffe...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/caffe-users/c2be1e35-984a-418d-a1b2-18be3ffabba4%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Abhishek Sharma]

Hi, Can you share the Layer network that you used? Thanks in advance.

[Royi]

Hi,

Could you share the code you used for training and the net?

I'm new to this package and would like to use it on 1D Signals.

Thank You.