nadavbar / node-cntk Goto Github PK

Allow the user to specify input and output variables once instead of each time when calling eval.

The tutorial should show how to take a trained Keras model, save it as a CNTK model using the underlying CNTK backend and then load it using this module and host it.

Add a sample that shows how to serve the model in a webserver.
One option is to have a website that lets you draw digits on a canvas and then sends it to the server for evaluation on the sample MNIST model.

@nadavbar

I am an Computer Engineering working Buenos Aires working for a small startup and we are doing some pretty exciting things down here. Our startup also does a lot with work microsoft primarily in IoT. However one of the projects we are working on will involve the use of Microsoft Cognitive Toolkit. I personally am not a big fan of python , but am a very big fan of C++ and node.js. Therefore running into this library really got me fired up. The way you designed this port is among the lines to how I would have gotten started, but it would be very interesting to create support for training. If possible could you send me an email at [email protected]. I would love to possibly sync up via skype and discuss your desired milestones for this project.

Thank you ,

Michael Roberts

The API currently only supports setting a default device. In general we should support setting device for each operation separately in case people want to distribute work across several GPUs.

Hi, is there any way to feed sequence data to node-cntk? I have a pre-trained model that I can feed one-hot data, but the sequence axis seems to throw it off, giving errors such as:

Error: Error occured during call to evalModel: NDArrayView::CopyFrom: The source view shape '[1023 x 1 x 4]' is not same as the shape '[1023 x 4]' of this NDArrayView.

(on a 4-word input with output vocabulary of 1,023) which seems to refer to a missing sequence axis. Feeding in two copies of the same sequence (as in, a batch of two elements) gives no response whatsoever, error or result. Any thoughts? Thanks

In case the model have just one input and outptut, support options to just pass input parameters without variable names, and without the need to specify output params.