More documented examples and code inside the PyTorch Tutorials

[shyamalschandra]

I would like to have the following inside the pytorch tutorials:

1. Multiple examples of possible boilerplate code using Jupyter.
2. Snippet of source code implement the function (if available) for perusal instead of a terse, ambiguous description of the function.
3. More Jupyter or REPL examples with the PyTorch zoo of models so users can generalize.

[karandwivedi42]

I can take this up, but I am unclear on what you mean.

Multiple examples of possible boilerplate code using Jupyter.

Boilerplate code for training convnets on mnist/cifar type datasets?

Snippet of source code implement the function (if available) for perusal instead of a terse, ambiguous description of the function.

For the pytorch docs or the tutorials?

More Jupyter or REPL examples with the PyTorch zoo of models so users can generalize.

Something like https://github.com/pytorch/examples/ but in jupyter?

[shyamalschandra]

@karandwivedi42: What is the timeline and schedule for the completion of the documentation if your confusion is resolved? My personal email address is shyamalc@gmail.com. Could you please clarify what is unclear? Do you need step-by-step requirements?

[soumith]

hey shyamal,

• what do you mean by "multiple examples of possible boilerplate code"?
• in what part (as you were doing the tutorial) did you feel like you wanted the "Snippet of source code implement the function". Can you give an example?
• we have plenty of examples here: https://github.com/pytorch/examples/ , what do you see missing?

[shyamalschandra]

Hi @soumith and @karandwivedi42 ,

Could you please start with commenting every line of the "working" examples? (e.g. https://github.com/pytorch/examples/blob/master/dcgan/main.py)

Furthermore, could you attach paper references to papers with clear examples of the exact functionality of that method so one can drill-down to the details?

Let's start there. If you or someone else can do this to make the examples clearer and more linear, that would be a great help!

[soumith]

Thanks for your comments.

Furthermore, could you attach paper references to papers with clear examples of the exact functionality of that method so one can drill-down to the details?

We have paper references in the README of every example except for SNLI example. We can fix SNLI example.

Could you please start with commenting every line of the "working" examples? (e.g. https://github.com/pytorch/examples/blob/master/dcgan/main.py)

This is a subjective choice, but we feel it's better for examples to be concise and self-explanatory. commenting every line makes it verbose to the point that they look like tutorials. The tutorials are meant to be at the level of verbosity that you seem to want.

[your deleted comment]

"Is pytorch a community project or a marketing tool? Did you understand that?"

Please dont bring your snarkiness here, we dont appreciate it. Folks like @karandwivedi42 work hard as community members to make PyTorch the best it can be and they dont need it. I remember your entitled behavoir and threats on the torch7 forums and torch7 issues, and it looks like your attitude hasn't changed much.