Then you can run mnist data server in tensorcom/ dir with ./serve-mnist zpub://127.0.0.1:7880
Tests & Dev
Build source
mkdir build
cd build
cmake ..
make
After running the build steps you can invoke the tests at DLinput-TF/ with python3 test/zmq_ops_test.py
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