4 ms·

Implementing a Convolutional Neural Network Using Only NumPy

126 points by vzhou842 7y ago

vzhou842 7y ago
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 https://repl.it/@vzhou842/A-CNN-from-scratch-Part-2
- stuqqq 7y ago
  You keep posting this a few times this week: https://news.ycombinator.com/from?site=victorzhou.com https://news.ycombinator.com/from?site=victorzhou.com
  - vzhou842 7y ago
    One of those was the prequel to this post: https://news.ycombinator.com/item?id=19981736 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.
    - modsiw 7y ago
      Did you change anything to help get traction? Seems to me that’s just a verbose way of saying you spammed it.
      - Cybiote 7y ago
        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.
        bjornsing 7y ago
        +1
  - Anon84 7y ago
    Doesn't HN get rid of duplicates?
  - minimaxir 7y ago
    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.
khasreto 7y ago
Wow nice job in the next part you can implement regression :) , This is done better than my lectures do at university haha. Greetings
neurohacker 7y ago
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-downsampling-of-labeled-images-using-python-and-numpy-e70ad3275589 https://towardsdatascience.com/countless-high-performance-2x...
panpanna 7y ago
Andrew Ng's coursed learn you to build CNN (and lots more) from scratch using only numpy. Hope does this compare to that?
- rahimnathwani 7y ago
  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.

126 points by vzhou842 7y ago

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 https://repl.it/@vzhou842/A-CNN-from-scratch-Part-2

stuqqq 7y ago
You keep posting this a few times this week: https://news.ycombinator.com/from?site=victorzhou.com https://news.ycombinator.com/from?site=victorzhou.com
- vzhou842 7y ago
  One of those was the prequel to this post: https://news.ycombinator.com/item?id=19981736 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.
  - modsiw 7y ago
    Did you change anything to help get traction? Seems to me that’s just a verbose way of saying you spammed it.
    - Cybiote 7y ago
      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.
      - bjornsing 7y ago
        +1
- Anon84 7y ago
  Doesn't HN get rid of duplicates?
- minimaxir 7y ago
  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.

khasreto 7y ago

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

neurohacker 7y ago

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-downsampling-of-labeled-images-using-python-and-numpy-e70ad3275589 https://towardsdatascience.com/countless-high-performance-2x...

panpanna 7y ago

Andrew Ng's coursed learn you to build CNN (and lots more) from scratch using only numpy. Hope does this compare to that?

rahimnathwani 7y ago
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.