As a data scientist, I can surely tell you that the image size and batch size arguments are most of the time to play around with. There are even libraries such as Optuna which find the best combination in terms of a chosen metric using cross-validation and Bayesian optimization, for any framework, and for any hyperparameter combination, but it would take respectively 10x time. Usually batch sizes are in the 2^n range. Also to avoid false positives, you can set minimum confidence value to set label to the minimum conf value of TP label, which here is around 0.8. As I said, all these steps are very dependent on the dataset and are almost always different.
What was the intended use case? Is it possible to use the trained model in Python code and you wanted it to run live while playing? Wouldn't this just help you "find" one on video, therefore after playing when you can't catch it anymore? Or was this all just a way to learn something new and you just wanted to see if you could find something in an image?
I was writing a script to automate traveling to space time distortions in pokemon legends arceus. (it waited for the message, flew there, visited the three fixed spawns, and if it was a "known" not-shiny porygon it'd continue on and loop -- stopping if it either missed, wasn't a porygon (likely a shiny!), or got knocked off by an alpha). I ended up getting the shiny porygon 2 without the full automation (just the waiting and navigation to the distortion) and I got the porygon z using by "bad" detection approach (thresholding on hsv) without the trained model. in theory I would have additionally cross referenced my hsv-based approach with the ml model but I never got around to that because I didn't need to
I've gone over this thresholding approach in a few of my other videos (try the ones further on in the playlist ideally): th-cam.com/video/chvgQUX7QaI/w-d-xo.html&pp=gAQBiAQB
fwiw they were fairly small (720p) and I didn't need near-instant response (the model timing was ~10s of ms) so within the reponse time budget I had in mind (order of ~500-1000ms)
As a data scientist, I can surely tell you that the image size and batch size arguments are most of the time to play around with. There are even libraries such as Optuna which find the best combination in terms of a chosen metric using cross-validation and Bayesian optimization, for any framework, and for any hyperparameter combination, but it would take respectively 10x time. Usually batch sizes are in the 2^n range. Also to avoid false positives, you can set minimum confidence value to set label to the minimum conf value of TP label, which here is around 0.8. As I said, all these steps are very dependent on the dataset and are almost always different.
Somhow I knew it will be to detect pokemon :D
Detecting pixel color will take you just so far. Great work and a nice video :-)
don't worry, he went back to pixel colors XD
@@sadhlife universe retains its balance
As a data scientist who follows this channel to become a better dev, this was wonderful. "Because someone told me to" ❤️
*blinks twice*
This is now a data science channel
What was the intended use case?
Is it possible to use the trained model in Python code and you wanted it to run live while playing?
Wouldn't this just help you "find" one on video, therefore after playing when you can't catch it anymore?
Or was this all just a way to learn something new and you just wanted to see if you could find something in an image?
I was writing a script to automate traveling to space time distortions in pokemon legends arceus. (it waited for the message, flew there, visited the three fixed spawns, and if it was a "known" not-shiny porygon it'd continue on and loop -- stopping if it either missed, wasn't a porygon (likely a shiny!), or got knocked off by an alpha). I ended up getting the shiny porygon 2 without the full automation (just the waiting and navigation to the distortion) and I got the porygon z using by "bad" detection approach (thresholding on hsv) without the trained model. in theory I would have additionally cross referenced my hsv-based approach with the ml model but I never got around to that because I didn't need to
This was a great tutorial I wanted to know how to train a yolo model. Could you do the tutorial for finding pink on the screen?
I've gone over this thresholding approach in a few of my other videos (try the ones further on in the playlist ideally): th-cam.com/video/chvgQUX7QaI/w-d-xo.html&pp=gAQBiAQB
Those file image sizes could be an issue
fwiw they were fairly small (720p) and I didn't need near-instant response (the model timing was ~10s of ms) so within the reponse time budget I had in mind (order of ~500-1000ms)
now we know you play pokémon :D
can i get code used in this video
it usually gets transcribed by viewers through pull request and ends up at github.com/anthonywritescode/explains
The very first