Discussions Tab? #428

Hi,

This doesn't really fit here, because I can't find a discussions tab. I have refrained from posting on here, as most of my queries are really looking for 'best practice' rather than 'i need help'

For example now, i have just filled my ssd with training models and wondering what the best way to free up space is? what are the downsides?

There are also issues like - anyone trained 'applejack' with 500 epochs with a 2080 - should it really be taking me >5 hours? (its my first training)

This is a really great resource, but I'm not sure where really to discuss it?

Could we have a discussion tab here? maybe a subreddit? I'd certainly help with that if i can.

What community resources to people use?

Thanks for reading, and thanks mrq for putting the time into this.

Hi, This doesn't really fit here, because I can't find a discussions tab. I have refrained from posting on here, as most of my queries are really looking for 'best practice' rather than 'i need help' For example now, i have just filled my ssd with training models and wondering what the best way to free up space is? what are the downsides? There are also issues like - anyone trained 'applejack' with 500 epochs with a 2080 - should it really be taking me >5 hours? (its my first training) This is a really great resource, but I'm not sure where really to discuss it? Could we have a discussion tab here? maybe a subreddit? I'd certainly help with that if i can. What community resources to people use? Thanks for reading, and thanks mrq for putting the time into this.

because I can't find a discussions tab.

Gitea still doesn't really have an analogue to Github's discussion tabs, so the best is just opening issues here and hoping someone else (or I) will answer it, or crossing your fingers and digging up any of my ancient documentation in past issues.

For example now, i have just filled my ssd with training models and wondering what the best way to free up space is? what are the downsides?

If you're training through the web UI, there should be a setting to have it keep the last K checkpoints during training, so it shouldn't be filled up with unneeded checkpoints.

Beside that, you're free to delete all but the last checkpoint, as those aren't necessary.

should it really be taking me >5 hours? (its my first training)

From what I remember off my 2060 and 6800XT, it does take quite a bit of time, and I usually ended up having a day to finetune a model to adequate results. However, I think that has more to do with playing it safe with having a very small LR sustained for a majority of training, so good results ended up always taking a while.

> because I can't find a discussions tab. Gitea still doesn't really have an analogue to Github's discussion tabs, so the best is just opening issues here and hoping someone else (or I) will answer it, or crossing your fingers and digging up any of my ancient documentation in past issues. > For example now, i have just filled my ssd with training models and wondering what the best way to free up space is? what are the downsides? If you're training through the web UI, there should be a setting to have it keep the last K checkpoints during training, so it shouldn't be filled up with unneeded checkpoints. Beside that, you're free to delete all but the last checkpoint, as those aren't necessary. > should it really be taking me >5 hours? (its my first training) From what I remember off my 2060 and 6800XT, it does take quite a bit of time, and I usually ended up having a day to finetune a model to adequate results. However, I think that has more to do with playing it safe with having a very small LR sustained for a majority of training, so good results ended up always taking a while.

Ah, thats fair enough, there is quite a bit of detail in the 'issues' and i have been finding them a great resource.

One more question if you don't mind - is there any way to keep previous training runs. I just ran 50, decided it needed a bit more, so ran 100 but was surprised to see it started at 1 rather than 51?

thanks again

Ah, thats fair enough, there is quite a bit of detail in the 'issues' and i have been finding them a great resource. One more question if you don't mind - is there any way to keep previous training runs. I just ran 50, decided it needed a bit more, so ran 100 but was surprised to see it started at 1 rather than 51? thanks again

You'll need to re-create your training YAML and specifying the Resume State Path. It's a bit of a bother with how DLAS requires it to be defined, rather it be automatically deduced.

You'll need to re-create your training YAML and specifying the `Resume State Path`. It's a bit of a bother with how DLAS requires it to be defined, rather it be automatically deduced.

Labels Milestones

Discussions Tab? #428