github30 / dlinput-tf

Optimized data input pipeline for deep learning frameworks

nvda-zmq-op

This is a tensorflow operator that allows you to pull tensors from an upstream pipeline using ZMQ

Requirements

Please check or install your requirements in following order.

• gnu gcc & g++ >= 5
• c++11
• python3
• cmake

Install instructions

Check permission of requirements.sh, makesure it can be executed.

python3 -m pip install .
or
pip3 install .

Examples

There are examples in ./example/ folder. It's .ipynb file and can be run by jupyter notebook.

python3 -m pip install --upgrade pip
python3 -m pip install jupyter
python3 -m pip install ipykernel
python3 -m ipykernel install --user
jupyter notebook --ip=127.0.0.1

Init submodule tensorcom git submodule update --init --recursive

Install pytorch for tensorcom server-mnist.

Then you can run mnist data server in tensorcom/ dir with ./serve-mnist zpub://127.0.0.1:7880

Tests & Dev

1. Build source

mkdir build
cd build
cmake ..
make

2. After running the build steps you can invoke the tests at DLinput-TF/ with python3 test/zmq_ops_test.py

3. Examples

   • To init submodel: git submodule update --init --recursive

   • perftest and cmp need add following codes to tensorcom/zcom.py if using imagenet

     #insert under line 122
     self.socket.setsockopt(zmq.SNDBUF, 10240)
     self.socket.setsockopt(zmq.SNDHWM, 200)
     self.socket.setsockopt(zmq.RCVBUF, 10240)
     self.socket.setsockopt(zmq.RCVHWM, 200)
     

   • Simple_training need to fix this bug as PR#24522 in installed tensorflow python source code currently.

     TODO: If it is merged into master and be released in the future, delete this instruction.

     locate training_utils.py
     

     modify the one in /home/<your username>/.local/lib/python3.5 and the other in /usr/local/lib/python3.5

     e.g. on my machine one is /home/xxx/.local/lib/python3.5/site-packages/tensorflow/python/keras/engine/training_utils.py

     search standardize_single_array

     # change 
     if (x.shape is not None
         and len(x.shape) == 1
         and (expected_shape is None or len(expected_shape) != 1)):
     
     # to
     if tensor_util.is_tensor(x):
         x_shape_ndims = array_ops.rank(x)
     else:
         x_shape_ndims = len(x.shape)
     if (x_shape_ndims == 1
         and (expected_shape is None or len(expected_shape) != 1)):
     

Documentation

Python documentation is automatically generated at build, use help() in python on entities to access the documentation

Structure

DLinput-TF
|--zmq_ops
    |--kernel       # c++ kernel codes
        |-- zmq_ops.cc  # Op registration file
        |-- zmq_base.h  # zmq connection resource base header & source file
    |--zmq_ops.py   # wrap kernel
    |--__init__.py  # python package init file
|--test # unittests
|--example  # some examples

Coding style: Google 4 indent

format all c++ header and source files in current directories

./format.sh