search

kaist-ina / NAS_public

This is an official repository of the paper, "Neural Adaptive Content-aware Internet Video Delivery"
Other
80 stars 23 forks source link

Some training problems. #6

Closed lhwwhl closed 4 years ago

lhwwhl commented 4 years ago

Dear Hyunho Yeo:

         I have read your article carefully recently and think this is a great project! In the process of learning, I have encountered some problems.

  1. Start training from a content-agnostic model (as described in this article). Does this content-agnostic model refer to using the provided DNN model with DIV2K training set?
  2. When can you open source dash.js module? I have a question, how does dash.js call pytorch trained models in the browser?

Looking forward to your reply!

jaykim305 commented 4 years ago

Hi, I am one of the collaborators and am happy to answer your questions on behalf of Hyunho.

  1. Yes. The content-agnostic model from our paper is trained with DIV2K dataset (https://data.vision.ee.ethz.ch/cvl/DIV2K/). Especially we used bicubic downsampled trainset of each scale.

  2. Sorry to notice you that it would be unavailable to open up dash.js source due to technology transfer issues. And, to answer your question, we didn't integrate pytorch to dash.js browser. We implemented seperate flask server with pytorch that handles request and runs DNN processing.

lhwwhl commented 4 years ago

Does the separate flask server run on localhost? Then using dash.js in the browser connects the flask server to handles request and runs DNN processing.

jaykim305 commented 4 years ago

Yes, the server runs on localhost. As you said, dash.js browser sends video chunks to the DNN server by sending the http request and DNN server handles the request.