BigDL-Tutorials
BigDL-Tutorials copied to clipboard
intel-analytics
new structure of the future BigDL tutorial
http://gluon.mxnet.io is a document-like website. I think it's different from the purpose of this project. This project provides a list of notebooks on mnist to demonstrate how to use bigdl and user can try by themselves. I strongly suggest you create another repo. Just like our doc website to hold your tutorial.
@yiheng Thanks for the review. Basically, we want to build a document-like website similar to mxnet tutorial page, which has a navigation bar with progressive instructions, enough details and various applications to fully present the flavor of deep learning with BigDL. That's the final goal so we can combine all the notebooks, emphasize their relativeness and provide users more friendly experience. But the current tutorial might be prone to be more single-tasking(MNIST digit recognition) and less elegant. Thus inadvertently, the tutorial needs a better organization with a document-like wiki support if the BigDL tutorial repository is expanded with more other topics in the future. Based on this wish, @jason-dai suggests me to put a proposal of some viable navigation structure just like what we can see in the navigation of http://gluon.mxnet.io. The proposal is temporarily put in the readme for everyone to review so don't worry about the readme change since it's just a proposal for open suggestion and ideas.
Although I wasn't told that the project was focused on MNIST. I can start a new repo. It's just that before finalizing the tree structure of the directory, we can probably have some reviews and ideas from peers about how to make the tutorial more scalable and think of what meaningful aspect of deep learning our repo is missing. Such aspects like overfitting, regularization, batch normalization which are basic but we can see in the MXNet tutorial they cherry picked almost every fundamental topic to teach novices from bottom to top hand by hand. It forms the quality of its tutorial which tends to be very user-friendly. We can make a better tutorial of the same quality.
I think one option is to create another folder for the new tutorials, which can be implemented by porting the MXNet tutorials (https://github.com/zackchase/mxnet-the-straight-dope) to BigDL. In the first stage, we can just implement these using the Keras APIs in BigDL.
