IDLE provides nice cross-compiler for * TensorFlow and * Caffe with simple format, called UDLI (Unified Deep Learning Interface).
Overall architecture of IDLE is introduced in * Taewoo Kim, Eunju Yang, Soyoon Bae and Chan-Hyun Youn, “IDLE: Integrated Deep Learning Engine with Adaptive Task Scheduling on Heterogeneous GPUs”, to appear in IEEE TENCON, 2018
- Ubuntu 16.04 (>=)
- python 2.7
- TensorFlow-gpu >= 1.90
- Caffe 1.0
- cuDNN-6.0
- CUDA 9.0
Python 2.7
$ sudo apt-get install python-pip python-dev Liberaries.
$ sudo apt-get update
$ sudo apt-get install libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev protobuf-compiler
$ sudo apt-get install --no-install-recommends libboost-all-devInstall CUDA : * CUDA 9.0
BLAS.
$ sudo apt-get install libatlas-base-devOther dependencies
$ sudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-devInstall Caffe
// Download source code
$ cd {workspace}
$ git clone https://github.com/BVLC/caffe.git
$ cd caffe
$ cp Makefile.config.example Makefile.config// Change Make configuration --> check python include, lib path
$ vi Makefile.config
# NOTE: this is required only if you will compile the python interface. # We need to be able to find Python.h and numpy/arrayobject.h. PYTHON_INCLUDE := /usr/include/python2.7 \
/usr/local/lib/python2.7/dist-packages/numpy/core/include/
# We need to be able to find libpythonX.X.so or .dylib. PYTHON_LIB := /usr/lib// Compile
$ make clean
$ make all
$ make test
$ make runtestInstall * PyCaffe
- Dependencies
$ cd {workspace}/caffe/python
$ for req in $(cat requirements.txt); do sudo pip install $req; done- Setting environment variables
$ echo export PYTHONPATH={caffe_path}/python:$PYTHONPATH >> ~/.bashrc
$ echo export PYTHONPATH={python_path}:$PYTHONPATH >> ~/.bashrc
$ echo export CAFFE_ROOT={caffe_path} >> ~/.bashrc
$ echo export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH >> ~/.bashrc $ source ~/.bashrc- Recompiile
$ make clean
$ make pycaffe- Test PyCaffe
$ python
>> import caffe$ sudo apt-get install python-pip python-dev
$ pip install tensorflow-gpuParmiko
$ pip install parmikoPillow
$ pip install pillowTo support distributed training, all workers should share ssh-key each other.
- SSH key generation (local machine)
$ ssh-keygen -t -rsa
$ ls -al ~/.ssh- send each ssh-key to all workers
$ scp ~/.ssh/id_rsa.pub (remote_machine_id@ip_address):($destination)- register each worker key to authorized_keys
$ cat $destination/id_rsa.pub >> ~/.ssh/authorized_keysInstall IDLE on your environment.
$ cd {working directory}
$ git clone https://github.com/KAIST-NCL/IDLE.git
$ echo export PYTHONPATH={working directory}:$PYTHONPATH >> ~/.bashrc
$ source !$Sample DL-MDL files are wtored under /model folder.
all files with dlmdl extension shows the examples.
Edit the sample files for your environment.
$ cd {working directory}
$ python src/UDLI.py --framework=tf --input=model/VGG -rarguments are described in below
| Argument | Explanation | Example |
|---|---|---|
--framework |
Choose framework | --framework = tf , --framework=caffe |
--input |
Path of input file (.dlmdl file) | --input=model/VGG.dlmdl |
-r |
Execution or not (If -r is set, execution is conducted right after compilation | -r |
--compile_out |
compile output path | --compile_out=output/VGG |
--log_out |
path to save TensorFlow log | --log_out=/tmp/IDLE_LOG/TF |
This project is released under the *BSD 2-Clause license, see LICENSE for more information.