Spaces:
Running
on
T4
Status Log:
Current project status:
- For transitions between 30 second segments are imperfect.
- implemented a fade between the two segments that lasts "overlap" time, evaluating.
- Interrupt button is not working correctly.
- There has been a major improvement in consistency with Melody Conditioned music!
6/18/23 - Getting a lot of timeouts waiting for queue
Update
6/18/23 - major consistency upgrade for melody conditioned music!
Fades between segments changed also
6/19/23 1 am: restarted server due to max overlap second max duration render. Changed UI to only allow overlap of 15 as max instead of 29
6/19/23 1:20 am: restarted again, as hackers edited the HTML to submit a billion second song - update to keep items within reasonable limits
6/25/23 - Add the ability to choose prompt conditioning sample segments ( -1 Conditions each verse of new music based on the parallel verse of the uploaded music)
- Update title to Melody Conditioning file name upon load
- Separate title and settings inclusions to background
- Fixed a bug in my 6/19 code... stupid logical mistake (modified and re-implemented due to hackers)
6/25/23 - Minor update to lock Melody Model when doing Melody Conditioning
6/26/23 - fixed bug in unconditioned melodies
6/26/2023 - MAJOR UPDATE
Transitions are about 50% improved
UI simplified
Add "HARMONY ONLY" that allows conditioning to remove drums
Ability to download .WAV files
6/27/23 - UI overhaul
Add Mic input
6/28/23 - Fix bug in short melody (under 30 seconds) generating long duration music
Begin review colab issue
6/29/23 - Minor update to prevent some more hacking, also add indicator of duration as seconds (s)
6/30/23 - Fixed a logical error that was exploitable to create long queues.
Facebook research is either down or deleted due to hackers... we cannot see the models
Ok, it is back
I looked deeply into adding effects, like low pass, splice, etc. However the library needed requires a custom torch implementation. This would have fixed the overlap issue and I am pursuing an alternative.
There was some bug in gradio 3.34.0 and another package 2 days ago. I have resolved it for this project and advised other musicgen projects
Ok, the system was working last night, but another gradio update seems to have broken it. This is now resolved
7/24/23 - I have noticed that the interrupt button is being abused and may limit it to only YOUR submissions.
7/25/23 - Updated gradio launch parameter, but ran out of memory. reverted.
7/26/23 - limited the queue to 10, after reboots the queue is filling up within a few minutes. API is disabled. I believe these are valid users based upon some of the prompts used.
Concurrent users being set to 2 instead of one revealed some batching issues, but they are not addressed because the disk size and GPU memory limits do not allow concurrent users. This is set to 1 concurrent user again.
The interrupt queue is being exploited and may force me to add a token that will only allow you to interrupt your own submissions. I do not see how to remove a submission from the queue until it is processing. Investigating.
7/30/23 - Oops there was an error in the 6/27 update integrated by the other developers that was preventing conditioned melodies from generating past 30 seconds. I should have looked closer. This is now resolved.
Once again, Apologies. We have people submitting maximum (720s) length requests. These are sometimes interrupted by other users to keep the queue moving. If the queue is less than 5 please allow the other users to complete their long requests. You are always welcome to run this program in your own space or the desktop version is available.
I have noticed that @ 21 seconds the Melody Model is adding a weird tone. Looking at how to resolve this.
NVIDIA GPU driver update pushed Torch and TorchAudio package updates... but HuggingFace has not implemented the driver and this broke the system. I fixed the package versions and will change back when the new NVIDIA drivers are installed
I once again plead that you keep the extremely long song generations to a minimum. You can test segments of 10-25 seconds very quickly with settings before creating the long version of a song.
Upgrading to use 6 new Models
- add meta data to videos for settings
- upgrading audiocraft version where needed
- possibly making variable length segments (instead of default 30 seconds)