Designing an extension for TFRecords

[maxbendick]

Generally, I want to use a parser written in Python to render a binary file in a JupyterLab extension.

I'd like to create an extension that can display data from Tensorflow TFRecord files. I can see one way I might do this in the pdf-extension, but I think creating a mime-renderer extension would require writing a parser for TFRecords in JavaScript/TypeScript/etc.

A parser does exist for TFRecords in Tensorflow, but it's written in C++/Python.

Here's some approaches to reuse the Tensorflow's parser, both very naive:

1. The user launches a python server that reads the TFRecord. The server is registered in the TypeScript extension. The extension queries the server for relevant data, then renders.

2. In the TypeScript extension, generate python code to read the record, and execute the code in a kernel. Render the output.

Any thoughts? Maybe there are more approaches I'm not seeing.

[ellisonbg]

A custom mime renderer would work great, but I would first convert the TF Record to JSON on the server side. Then you can write a frontend renderer for that JSON data.

…

On Sun, Dec 31, 2017 at 3:34 PM, Max Bendick ***@***.***> wrote: Generally, I want to use a parser written in Python to render a binary file in a JupyterLab extension. I'd like to create an extension that can display data from Tensorflow TFRecord files. I can see one way I might do this in the pdf-extension <https://github.com/jupyterlab/jupyterlab/tree/master/packages/pdf-extension>, but I think creating a mime-renderer extension would require writing a parser for TFRecords in JavaScript/TypeScript/etc. A parser does exist for TFRecords in Tensorflow <https://www.tensorflow.org/api_docs/python/tf/TFRecordReader>, but it's written in C++/Python. Here's some approaches to reuse the Tensorflow's parser, both very naive: 1. The user launches a python server that reads the TFRecord. The server is registered in the TypeScript extension. The extension queries the server for relevant data, then renders. 2. In the TypeScript extension, generate python code to read the record, and execute the code in a kernel. Render the output. Any thoughts? Maybe there are more approaches I'm not seeing. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#3505>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AABr0EArb95HKJ3u_2t6tIJDYbvmiWd-ks5tGBoIgaJpZM4RP9FO> .

-- Brian E. Granger Associate Professor of Physics and Data Science Cal Poly State University, San Luis Obispo @ellisonbg on Twitter and GitHub [email protected] and [email protected]

[jasongrout]

@danielballan wrote an hdf5 viewer that uses a server-side library to parse the file and provide a server extension for data requests and a front-end renderer that seems similar to what you want to do as option (1). @danielballan - do you have your code somewhere?

[danielballan]

Yes. It's not pushed anywhere and I'll have to go digging, but I'll try to find time this weekend.

[danielballan]

@maxbendick Here's my work so far:

• a modification to the notebook server, adding a new configuration whereby new data formats ("external formats") such as HDF5 can be mapped to some separate server -- outside of the notebook server's purview -- such as h5serv (this piece was developed with some input form @ellisonbg and @jasongrout)
• a PR into phosphor (check out the GIF screencap demo) with a lazy-loading data grid that calls out to an h5serv endpoint to load chunks of data as needed
• a new JupyterLab extension -- the important file is this one -- which is not quite finished

The phosphor PR is the only piece that fully works, but I'll take this opportunity to revisit the work and connect the rest of the pieces....

[dnelson86]

See also #448