Help with overtrain

[flint jackal]

If I understand correctly, isn't it overtrain from here?

[tiny thorn]

i think so yes

[trail dawn]

Is your smoothing set to .987 on Tensorboard?

Hard to tell if its collapsing tbh. The spikes sometimes mean its not doing good, but I'm not sure here.

If that 34k~ step doesn't sound right, maybe try adjusting your batch size and retraining

[flint jackal]

This is what it says here https://docs.aihub.wtf/rvc/resources/epochs--tensorboard/#in-the-left-panel

[exotic terrace]

Ayo? @flint jackal level 2 !!!

[flint jackal]

batch size was 7. I can do maximum 8

my GPU can't handle above 8 (RTX 4060 Ti 8 GB)

[chilly sparrow]

get the 4060 Ti 16 GB variant, or 3060 12 GB is cheaper but still capable

[flint jackal]

do you know Turkey's economic situation? It was my childhood dream to buy this kind of GPU

I can't suddenly buy a different GPU yk

[trail dawn]

You can do 16 batch if you want

You should stick to certain numbers though

[flint jackal]

so I haven't even tried to go above 8 batch

[trail dawn]

Try 4 batch. This will make your model take longer to train probably but more betterer

[flint jackal]

I mean I tried but even when with 8 batch the vram is reaching its limits

[flint jackal]

i will try today

[trail dawn]

Or if you're short on time, do 12 batch

[flint jackal]

my last experiments were not very successful

i'm still learning

[trail dawn]

Every time you train its unique btw. RVC is hard to control

Kinda

[flint jackal]

I have time no problem with that

[trail dawn]

See how 4 batch does. Hopefully the tensorboard will be a little more smooth

Could also be a bad dataset. It doesn't suck right

You can DM a short snippet of your dataset if you want

[flint jackal]

the dataset is good

[flint jackal]

400 mb takes a while to upload

[chilly sparrow]

it will take more than 8 GB vram, and will be bottlenecked if insufficient

[flint jackal]

thats why

I checked my VRAM usage while training models and decided batch 7 was the best for me

[trail dawn]

Just a snippet

Like 30 seconds