This library has been tested on various flavours of Linux, Android, MacOS and iOS. A Makefile is provided that should work in most of the cases. Android build files are also provided for a test application. For Android / iOS see section below.
Requirements: OpenBLAS (not required if compiled for ARM, ARM64 and x86_64), CuDNN version 4 or 5 if compiled with the CUDNN option, OpenCL if compiled with the OPENCL=1 option, gemmlowp if compiled with the LOWP=1 option, google protobuf is compiled with the ONNX=1 option. Check the CUDA and CUDNN directories in the Makefile if using CUDNN. Make with "make". Install with "(sudo) make install". Make options are:
- ONNX 0 is off, 1 is enable support of ONNX networks
- DEBUG 0 is off, 1 is on
- MEMORYDEBUG 0 checks memory leaks, 1 generates full dump of allocations in memdump.txt
- CUDNN 4 uses CuDNNv4, 5 uses CUDNNv5
- OPENCL 0 is off, 1 uses OpenCL
- LOWP 0 is off, 1 uses gemmlowp
- SANDYBRIDGE 0 is default (PENRYN), 1 uses SANDYBRIDGE routines
- HASWELL 0 is default (PENRYN), 1 uses HASWELL routines
On Intel architectures, the default is to use Penryn routines (e.g. Core 2 E8xxx CPUs), which will work on most machines. By activating SANDYBRIDGE (2nd generation Core) or HASWELL (4th generation) it will take advantage of new instructions present on more modern CPUs.
ONNX support needs to install protobuf. See examples for an example of how to install protobuf.
gemmlowp (https://github.com/google/gemmlowp) has to be installed into the lowp directory, if used.
Raspberry Pi / Raspbian installation: edit the Makefile and set 'USEQSML = no'
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>
This test reports running time in seconds for each iteration on the 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 an image file (JPEG, PNG, TGA, BMP), or a .t7 file containing a FloatTensor of dimension 3
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.
Create a new network from the given network. The new network will use OpenCL.
Create a new network from the given network. The new network will use gemmlowp (8 bit gemm). The output of the convolutions is supposed to be between -range/2 and +range/2. This feature is experimental.
Enables the use of 16 bit floats on OpenCL.
int THProcessFloat(THNETWORK *network, float *data, int batchsize, int width, int height, float **result, int *outwidth, int *outheight)
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.
int THProcessImages(THNETWORK *network, unsigned char **images, int batchsize, int width, int height, int stride, float **result, int *outwidth, int *outheight, int bgr)
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.
int THProcessYUYV(THNETWORK *network, unsigned char *image, int width, int height, float **results, int *outwidth, int *outheight)
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. Please note that the application is older of this library and will require some modifications in order to be compiled.
I don't have the time to build a complete Android example, but just to give some hints, I created a sample Android.mk and Application.mk file that will create an executable file that will run on Android devices. If you don't have a rooted device, you can use SSHServer from Ice Cold Apps, ssh into the device and copy the executable here: /data/data/com.icecoldapps.sshserver (the only accessible place where you can set the executable flag on files). images.c and test.c are required only for the test executable, leave them out if you only need the library.
An example Android app is here.
Add source C file in XCode, including all the ones in modules (leave out OpenBLAS-stripped and other files). images.c and test.c are required only for the test executable, leave them out if you only need the library.
In the project -> Build Phases -> You have to add -DUSEBLAS -DACCELERATE in the compiler options for files: SpatialConvolution.c, SpatialConvolutionMM.c thbasic.c, thapi.c
Add Accelerate Framework to Project -> Build Phases -> Link Binary with Libraries
An example iOs app is here.
WARNING: due to continuous changes in pytorch, this mode does not work anymore. While files exported with the previous version can be still imported by thnets, the thexport.py utility does not work anymore with the latest pytorch. The ONNX exported cna be used instead.
In order to run networks created with PyTorch, they have to be first exported with thexport.py. This is an example python code that will create a pymodel.net file that thnets will be able to load:
import torch
from torchvision import models
import thexport
net = models.alexnet().eval()
out = net.forward(torch.autograd.Variable(torch.FloatTensor(1,3,227,227)))
thexport.save('pymodel.net', out)
pymodel.net can be then loaded by thnets in the same way as Torch created model.net, just giving its directory, or also direct path to the file in this case. This exporter has been created, because torch.save() saves python code that cannot be interpreted by thnets. Not all thnets supported layers have been implemented, yet, but all the networks present in torchvision at the time of this writing are supported.
ONNX support has been tested with the networks in torchvision (pytorch) and the models in the ONNX repository. While all torchvision networks work (besides squeezenet, which is not exportable to ONNX), only resnet50 from the ONNX repository works, but this is only due to unimplemented layers like LRN. Other frameworks are still under test. If the passed path is a path to a file with the .pb, .proto or .onnx extension, the ONNX parser will be triggered. thnets has to be built with the ONNX=1 option.
Forward times in seconds:
| Image Size | Direct | MM | cuDNN | 16-bit cuDNN |
|---|---|---|---|---|
| 1280x720 | 2.461129 | 1.549271 | 0.088064 | 0.088064 |
| 1920x1080 | 7.196426 | 3.666830 | 0.195412 | 0.129979 |