Hey, author here. Happy to hear any comments or answer any questions.

Runnable code from the article: https://repl.it/@vzhou842/A-CNN-from-scratch-Part-2

You keep posting this a few times this week: https://news.ycombinator.com/from?site=victorzhou.com

HN allows a small number of reposts for a given post if the original didn't get much traction.

In this case however, this particular is a Part 2 to a Part 1 that was successful on HN, which may not be sufficiently different.

One of those was the prequel to this post: https://news.ycombinator.com/item?id=19981736 (titles are similar because content is similar)

The other two were reposts because they didn’t gain traction.

Did you change anything to help get traction? Seems to me that’s just a verbose way of saying you spammed it.

The page contains superb, high quality work. I find the negativity in this branch incomprehensible. I and I suspect, many others, did not see any of the other times the author posted it. I am probably not alone in being glad they persevered at sharing it.

+1

Doesn't HN get rid of duplicates?

If you want to downsample segmentation produced from convnet output using only numpy, I have this technique to contribute :)

https://towardsdatascience.com/countless-high-performance-2x...

Wow nice job in the next part you can implement regression :) , This is done better than my lectures do at university haha. Greetings

Andrew Ng's coursed learn you to build CNN (and lots more) from scratch using only numpy.

Hope does this compare to that?

Victor's CNN posts cover roughly the same ground as section 1 (of 4) of Andrew's CNN course.

Andrew's explanations in the videos are really well crafted, and cover the 'why' of everything clearly.

Sections 2-4 of Andrew's course cover more advanced CNN architectures, and also more advanced use cases that image classification.