MXConsole

MXConsole is a standalone TensorFlow's Tensorboard port intended for MxNet, but now is plotform netural. Now this project is in very early stage.

Structure

MXConsole use extracted TensorFlow's filesystem and record_reader and record_writer code as a python extension. Thus we don't need to build the whole TensorFlow. meanwhile, without TensorFlow's Tensor and ops. Platform relavant summary related code is needed . Now we only provide api that generate summary from numpy.ndarray. part of these api is merged from dmlc/tensorboard.

Build tools

1. bazel
2. node.js and bower(actually we only need bower to do bower install, so just npm install -g bower should be fine`)
3. optional: gulp and other node.js module(for frontend build only)

Installation

• normal installation

  1. clone this repo git clone https://github.com/bravomikekilo/mxconsole --recursive
  2. cd mxconsole/tensorflow_fs
  3. ./configure to configure tensorflow_fs build, choose your python binary and default libs and choose only the jemalloc support.
  4. cd .. && ./update_native.sh to build and update tensorflow_fs
  5. python setup.py setup.py bdist_wheel to build python wheel
  6. pip install dist/MXConsole-0.0.1a1-py2.py3-none-any.whl install the python wheel
  7. You now can use the MXConsole like python -m mxconsole --logdir path/of/your/logs

• refresh the native library

  1. $ ./update_native.sh

some logs can be generated from carpedm20/DCGAN-tensorflow, just train that model on mnist is just fine. You also can try the example merged from dmlc/tensorboard, under the demo folder.