Requirements for library: OpenBLAS (already part of the library for ARM and x86_64), CuDNN version 4 if compiled with the CUDNN=1 option.
Requirements for test: libpng and libjpeg
Check the CUDA and CUDNN directories in the Makefile if using CUDNN. Make with "make". Install with "(sudo) make install". Make options are:

• DEBUG 0 is off, 1 is on
• MEMORYDEBUG 0 checks memory leaks, 1 generates full dump of allocations in memdump.txt
• CUDNN 0 is off, 1 uses CuDNN

export LD_LIBRARY_PATH=/usr/local/lib:/opt/OpenBLAS/lib (add CUDA and CUDNN lib directories, if using CUDNN)
./test -m <model_dir> -i <input_file>

The model directory must contain 2 files:

• model.net the network file saved in .t7 format
• stat.t7 contains a table with a 'std' and 'mean' FloatTensor of dimension 3

Input file can be a .jpg or .png file, or a .t7 file containing a FloatTensor of dimension 3

A demo model can be downloaded from teradeep/demo-apps

Initializes the library.

Loads the network contained in the path directory and returns a THNETWORK object or 0, if the network cannot be loaded. The reason can be obtained with THLastError().

Makes the loaded network suitable for images bigger of the eye size.

Changes every occurrence of SpatialConvolution in the network to SpatialConvolutionMM (nn_type>0) or viceversa (nn_type=0). SpatialConvolutionMM modules with padW or padH different of 0 will not be changed to SpatialConvolution, as the SpatialConvolution module does not support them. If nn_type=2 and supported (currently only on ARM), it will use virtual MM, which takes much less memory.

Create a new network from the given network. The new network will use CuDNN.

Enables the use of 16 bit floats on CUDA.

Runs the network on the float data. Float data is organized as a coniguous array of size batchsize x 3 x height x width, where 3 is the number of color planes. The order of the color components is expected to be red, green, blue.

Returns the number of categories in the output and the size of the output in outwidth and outheight.
Result will point to the array with the data and must not be freed.
The data is a contiguous array of size batchsize x number of categories x outheight x outwidth.

Runs the network on the series of images. Images is an array with batchsize pointers and each element points to the start of the image. Images are arrays of size height x stride x 3, where only the first width of each line long stride contains data. If bgr is zero, the order of the color components is expected to be red, green, blue, otherwise it's expected to be blue, green, red.

Returns the number of categories in the output and the size of the output in outwidth and outheight.
Result will point to the array with the data and must not be freed.
The data is a contiguous array of size batchsize x number of categories x outheight x outwidth.

Runs the network on an image in the YUYV format. This function is useful when the image comes from a camera, where the YUYV format is common.

Returns the number of categories in the output and the size of the output in outwidth and outheight.
Result will point to the array with the data and must not be freed.
The data is a contiguous array of size batchsize x number of categories x outheight x outwidth.

Frees the network and all associated data and outputs.

Returns an error code describing the reason of the last error. It is now used only for THLoadNetwork and can give these results:

• 0 Ok
• -1 The file cannot be opened
• -2 The file cannot be read till the end
• -3 The file contains some elements, which were not implemented in this library
• -4 The file is corrupted
• -5 The file contains torch objects not expected in that file

See: https://github.com/teradeep/demo-apps.git The neural network model can be downloaded from teradeep/demo-apps

Forward times in seconds: