Hi, I love that your video shows all the coding parts. I think your format is great! For an improvement, I would look into highlighting even more the code and the software architecture. For example, you could hand draw a simple graph that shows the architecture. You would have a little box for the dataset, another box for the torch model and so on. And then, you can "fill the boxes" with the coding parts. Great video!
Really appreciate your feedback! Great point! I am actually planning to look into some diagram creation tools since my drawing skills are terrible:D One YT channel that really stands out when it comes to diagrams is: ua-cam.com/users/AlfredoCanzianivideos . I will definitely try to take some inspiration from it:) Thanks again!
Great videos. I like your style of writing the code and going through it like this. I also enjoy the documentation that you write. Regarding the reason, the model is suboptimal in the div5 and div10 cases, my guess is it has to do with class imbalance. When you added sequence length and max restrictions, many of the examples in the data are eliminated, and 5 and 10 sequences probably have many of their members fall outside of your max value.
Thank you for your comment! I really appreciate it. Regarding the div5 and div10, yeh great point, the imbalance is probably one of the main reasons why the model is struggling in terms of accuracy. I was hoping that the LSTM model would at least be able to beat the dummy benchmark of always predicting the "dominant class" (= not divisible). So in the case of div10 this would be 90% accuracy and for div5 this would be 80%. However, it did worse than that. But yeh, probably paying more attention to the imbalance (using a different metric than accuracy, using class weights, oversampling, etc.) would make sense!
That is a pretty interesting idea actually! I did some googling and stumbled upon this resource: ai.stackexchange.com/questions/3850/can-a-neural-network-be-used-to-predict-the-next-pseudo-random-number Really interesting topic, I hope to look into it a little bit more in the future:) Thanks!
Great point! I agree that whenever there is some plotting involved Jupyter notebooks are superior to the IPython console. At the same time, I don't want to write all of my code inside of Jupyter notebooks since I find it less readable. However, that is of course highly subjective:) Anyway, thank you for your feedback!
Errata: - 9 is not a prime :D - You should not abuse the API of OEIS, since they actually already provide a simpler download of all sequences oeis.org/wiki/Welcome#Compressed_Versions
Great Video, Thanks
Glad you liked it!
Hi, I love that your video shows all the coding parts. I think your format is great! For an improvement, I would look into highlighting even more the code and the software architecture. For example, you could hand draw a simple graph that shows the architecture. You would have a little box for the dataset, another box for the torch model and so on. And then, you can "fill the boxes" with the coding parts. Great video!
Really appreciate your feedback! Great point! I am actually planning to look into some diagram creation tools since my drawing skills are terrible:D One YT channel that really stands out when it comes to diagrams is: ua-cam.com/users/AlfredoCanzianivideos . I will definitely try to take some inspiration from it:)
Thanks again!
@@mildlyoverfitted Great! I have already watched some of his videos. I'm looking forward to your next vid!
Great videos. I like your style of writing the code and going through it like this. I also enjoy the documentation that you write.
Regarding the reason, the model is suboptimal in the div5 and div10 cases, my guess is it has to do with class imbalance. When you added sequence length and max restrictions, many of the examples in the data are eliminated, and 5 and 10 sequences probably have many of their members fall outside of your max value.
Thank you for your comment! I really appreciate it.
Regarding the div5 and div10, yeh great point, the imbalance is probably one of the main reasons why the model is struggling in terms of accuracy. I was hoping that the LSTM model would at least be able to beat the dummy benchmark of always predicting the "dominant class" (= not divisible). So in the case of div10 this would be 90% accuracy and for div5 this would be 80%. However, it did worse than that. But yeh, probably paying more attention to the imbalance (using a different metric than accuracy, using class weights, oversampling, etc.) would make sense!
Please do jupyter notebook with numpy tricks that you did at the beginning of this video (when analyzing data). It was so informative.
Awesome video!!!!!❤️
Glad you liked it!!
you could check if you can predict Linear Congruent PRNGs (overflow) and Shift operator based PRNG
That is a pretty interesting idea actually! I did some googling and stumbled upon this resource: ai.stackexchange.com/questions/3850/can-a-neural-network-be-used-to-predict-the-next-pseudo-random-number
Really interesting topic, I hope to look into it a little bit more in the future:) Thanks!
It could be more informative if you use juoyter for the video. That would let you plot various things like histogram etc.
Great point! I agree that whenever there is some plotting involved Jupyter notebooks are superior to the IPython console. At the same time, I don't want to write all of my code inside of Jupyter notebooks since I find it less readable. However, that is of course highly subjective:) Anyway, thank you for your feedback!
Errata:
- 9 is not a prime :D
- You should not abuse the API of OEIS, since they actually already provide a simpler download of all sequences oeis.org/wiki/Welcome#Compressed_Versions