Standalone Tensorflow Projector

[lhausermann]

Sorry to double open this issue but did not get any answer from StackExchange forum :
http://stackoverflow.com/questions/41643365/standalone-tensorflow-projector

Tensorflow Projector is a great way to visualize a mutli dimensional dataset and I was wondering if you are planning to open source the version running on the web (ie outside tensorboard), so we would be able to visualize any dataset from a simple TSV file (and not only word embedding or other learnt representation from TF)

Thanks

[Baschdl]

Hi @lhausermann, I published a small script to use the embedding projector only for visualization of your data. https://gist.github.com/Baschdl/e965ba3499bb6e905f92befa7f77bbf2
You must only load the data and define the input variables accordingly. Run the script once and you can open the projector with tensorboard --log-path ../path/to/log.

[zihaolucky]

@Baschdl I guess @lhausermann wants the one like http://projector.tensorflow.org/ but allows you to upload your own data, which means it doesn't depend on TF. But the issue is how to generate the TF's checkpoint file.

[aselle]

@dandelionmane, could you comment on the stackoverflow issue please http://stackoverflow.com/questions/41643365/standalone-tensorflow-projector
however, I am going to close the issue here since it is more of a stackoverflow question I think. Thanks everybody.

[teamdandelion]

I commented on StackOverflow that it has only been released as a TensorBoard plugin.
@dsmilkov would be the person to talk to about releasing a standalone version.

[teamdandelion]

Going to re-open and let @dsmilkov decide if this should be closed or not.

[dsmilkov]

We didn't open source the standalone version, however it's pretty easy to use the existing standalone version with your own data. Just fork https://github.com/tensorflow/embedding-projector-standalone/ and edit the oss_demo_projector_config.json to point to your datasets.

[harveyslash]

what is the format of the tensors.bytes that the standalone projector seems to use?

[justindujardin]

@harveyslash The .bytes files are binary files filled with a bunch of 32 bit floating point numbers. You can use a numpy array with dtype=numpy.float32 and tofile to generate them:

vectors = numpy.zeros(vector_shape, dtype=numpy.float32)
vectors.tofile('my_tensors.bytes')

Here's a complete example of exporting spaCy Vocab vectors for use in the standalone projector.
https://github.com/explosion/spaCy/blob/master/examples/vectors_tensorboard_standalone.py