Replies: 5 comments 1 reply
-
|
@Timzoid Thank you for your feedback and thoughts on the updated versions. I really appreciate it :)
Unfortunatelly, I can't position instruments selection horizontally because Google Colab does not have that option in the interface. FYI, if you want arrangement/orchestration (inpaint desired instruments or change instruments), I have my Orchestrator implementation:
RE: Number of memory tokens: Number of memory tokens controls long-term structure of the composition. Basically, this number represents the amount of tokens/notes the model remembers each time it generates new notes/tokens. The more the better but the more memory tokens you use, the slower the generation is going to be. And in some cases, it may not have any benefit at all. It really depends on the composition and how you direct/control the model output. MuseNet had 4k number of memory tokens, LAMC has the same capability. Euterpe and Allegro have 2k max memory, which is still sufficient I think. In fact, you often can get away with 1k-1.5k memory toekens to make decent music. Now, re your Youtube videos/tutorials for LAMC. I most sincerely apologize for changing the whole thing and I am very sorry you have to delete that work now. However, I hope you agree that new versions are now easier to work with and they are more streamlined and polished. So if you want, we can discuss the changes/agree on what needs to be changed before you make another video/tutorial. I really appreciate your help and your promotion of LAMC on YouTube so I am willing to work with you so that you do not have to do vain work.
Hope this answers your questions and I again thank you for your feedback and thoughts. Please let me know what else you want to see and also please let me know if you want to make a new videos/tutorial on YouTube so that we can coordinate it to avoid any problems with changes and such.... Sincerely, Alex |
Beta Was this translation helpful? Give feedback.
-
|
@Timzoid I had a minute so I added notes to seed MIDI loader and Inpainting Generator. And I set inpating sampling default to 1 as you have asked. Check it out and let me know what you think. Alex |
Beta Was this translation helpful? Give feedback.
-
|
@Timzoid Ok. I think its done! :) I've incorporated all suggestions you had in all of the implementations. I added notes everywhere and added inpainting to Allegro Original Version. Let me know if you like it and if you have any suggestions. Alex |
Beta Was this translation helpful? Give feedback.
-
|
Great! The note will suffice on the custom seed loading. It's very simple and works well. Plus, the more I thought about it, having to load the cell first shouldn't be completely counterintuitive for the new user, since all the program cells have to be loaded before things work. I thought that might be the case about the vertical placement of instruments, but my chatbot told me it wasn't. Well, it's wrong A LOT. That's good to know about the memory tokens! It is so hard to find descriptions of these things online as they pertain to music generation, because all I get in searches is what that means for LLM that is generating words, not music. I thought about that being a possibility at one point, of what it does, but then forgot about it, finding that anything over 1000 seemed to work fine. In one respect, if you have a piece that is 2000 tokens long that is being inpainted, you maybe don't want it to be sourcing too much structure from the first 1000 tokens anyway. I don't mind about making a new tutorial, but I put in some notes in my pinned comment and the details section, and I may wait and see if I get a single person confused over seed loading before I bother making a new video about it. I just deleted my separate video on seed loading, and discovered it had very few views anyway, so no loss there. I made that tutorial video after only a few days of using LAMC, so it is probably time anyway to make a new one so I can explain the parameters a little better to people. And no need to apologize, I'm happy you're are engaged in taking my suggestions, but especially improving your own programs and writing new ones. I'll definitely try "Orchestrator." I think I explained before, I doubt my tutorial on LAMC would have about 1700 views now, if it weren't for the fluke that a few months after the launch of ChatGPT, channel traffic on my MuseNet tutorial went crazy, and still is. even though I changed the title to MuseNet is dead. But it's my link there that got people to watch the LAMC tutorial. Now going into fall and winter, and being as old and retired as I am, I am so glad to have your programs to use. Ultimately though, using GitHub is for the more serious hobbyist and some professionals, that have good ability with MIDI software. Whereas MuseNet, was loved by some people who didn't use MIDI at all, or very little anyway. And AI and MuseNet are what prompted me to really learn MIDI software, even though I'd use it for decades at a basic level. Lastly, a problem with the "inpaint" this morning on LAMC, October 11. Cell doesn't run and I get an error message: No problem for me, and so looking forward to trying inpainting on Allegro! Going there right now. |
Beta Was this translation helpful? Give feedback.
-
|
@Timzoid Thank you for reporting an error. I think I corrected it., Try it now and let me know if there are any other issues with it. Yes, a spike in interest for Music AI models (LAMC and others) was indeed because MuseNet went down. So I am very happy that we both were able to seize the moment and to fill in this gap. I am very grateful for your support with YouTube tutorials. It helps to make LAMC useful for people who are interested in Music AI. I will tell you for sure that when MuseNet first came out, I was 100% sure that it won't last long so LAMC and my other models/projects were created specifically to serve as an open-source and free replacement for MuseNet. So the spike in interest was very useful and good for that purpose. Anyway, thank you again for your feedback and thoughts and for reporting an issue with LAMC. Alex |
Beta Was this translation helpful? Give feedback.
-
|
@Timzoid PS. I am currently working on a better model/implementation that will feature full MIDI note events range and 4k memory. It should also be very fast like MuseNet was. So there is definitely a lot of work still to be done and there is a lot of room for improvement. |
Beta Was this translation helpful? Give feedback.
-
I've used "inpaint" almost daily for the last few months, and I notice this morning you made the change I suggested about uploading a custom seed file through a button, rather than having to go through the CoLab directory process including Copy Path. So that's good. This is definitely easier. The bad part is, it is still hidden, requiring that the cell be run to see the button. That took me 10 seconds to figure out. The problem is, that's going to be confusing for new users. Still, not as convoluted as the old method, so this is a big improvement. Maybe it would be enough to put a note: (Cell must be run to upload custom MIDI file.)
So, you've added MI to inpaint and that's nice, except no matter what instrument combination I choose, it just ignores it. Also, is there a way you could put the instrument list laterally instead of vertically? Each time I need to scroll up, like to load a new seed, I have to scroll by all those instruments. According to the AI chatbot I use, this could be an issue of "overfitting," which is the training model tends to pick whatever it is trained on the most, which is piano music.
Could you set the default of "number of samples per inpainted note" to 1 instead of 4. Having it at 4 quadruples the processing time of the inpaint sequence and maybe it's just me, but I get the best results at 1, and it is so much faster. If I set that on 6, for example, it hammers away on one note for the whole piece. So, I think for newbies, best to just have the default be 1.
I had a discussion with an AI chatbot about samples per note. Is more ever better? The chatbot said, not always.
In addition to my YouTube tutorial on How to Use Los Angeles Music Composer, a few months ago I made a second one on Seed Loading and using the "Inpaint" feature of the "Original Version", and now I'll just delete that, because both things are changed. Of course, I described seed loading in my main video about LAMC, but I'll just put some notes in the comments how seed loading has changed, and in the details section, and hope people will see that.
I see a value in the "Number of memory tokens slider" but don't completely understand the effect of that. It seems to work the same whether it's 4000 or 1000. I probably need to experiment with it more. I tried really low settings and they mess up the sequence. Anyway, I felt I really mastered the two simple settings on the old version. I assume the default on that was a sample rate of 1. But I can see where this new format inpaint, same one as was on Euterpe-X, is fine.
I am using Allegro Music Transformer, "Composer Version," more now. Are you going to do an "inpaint" feature for the original version of that? In my process, I am starting to use the composer version more, but even things I do on that from a seed generated by Huggingface, for example, I tend to augment it by doing multiple inpaints on the thing before I start slice and dicing it into a composition in MIDI software.
Beta Was this translation helpful? Give feedback.
All reactions