laurenceday / dreambooth-ion-cannon Goto Github PK

if i want to add new training images and retrain, what do I need to delete?

for example:

  rm -rf  logs/*
  rm trained_models/*.ckpt

# and outpust
  rm -rf outputs/*

anything else I should delete that won't get overwritten by running the notebook again?

in the "Downloading/Reconstituting Stable Diffusion v1.4" section of dreambooth_ion_cannon.ipynb the filepaths include the "workspace" folder.

running the notebook as is throws and indexError - we cloned this repo into /, not into workspace

The images in the readme and the lack of an instruction to cd into workspace make me think that it's a defunct folder.

hey thanks for this! I've fine-tuned a model on the man class and things worked fine. but I am trying a woman and the raw images and samples come out identical no matter what I do.

are there any other steps to trouble shoot this? I can confirm the files are named as:

[email protected]:/workspace/Dreambooth-Ion-Cannon$ tree training_samples/

training_samples/
└── woman
    ├── emma woman_001.png
    ├── emma woman_002.png
    ├── emma woman_003.png
    ├── emma woman_004.png
    ├── emma woman_005.png
    ├── emma woman_006.png
    ├── emma woman_007.png

and the notebook vars are initialized as:

target_name = "emma"
target_class = "woman"

not sure how else I could mess it up :)
any other hints where to look? I created a clean new vast instance and have done this twice now with identical results.

the raw gallery image comes out as the attached, both times (even though I'm using more training images now)

any other suggestions appreciated!

this might be useful to people, I made a justfile for syncing local/remote files using rsync. if it's useful I could make a PR ...

justfile

# https://github.com/casey/just

# assumes you have symlinked the right file to .env
set dotenv-load

default:
    @just --list

checkenv:
  @echo "vastPort: $vastPort vastLogin: $vastLogin"

sync-down: checkenv
  rsync -avu -e "ssh -p $vastPort -i ~/.ssh/rikai.pub -o StrictHostKeyChecking=no -l root" $vastHost:/workspace/Dreambooth-Ion-Cannon/outputs synced

sync-up: checkenv
  rsync -avu -e "ssh -p $vastPort -i ~/.ssh/rikai.pub -o StrictHostKeyChecking=no -l root" synced/inputs/emma/ $vastHost:/workspace/Dreambooth-Ion-Cannon/training_samples/woman

and a .env file in the same directory with your vast config

# .env 
vastPort=39443
vastHost=ssh5.vast.ai
vastLogin=root@$vastHost

you can get the port etc from the instances dashboard here

I would like to follow this rule while training my model:
Learning Rate: 5e-5:100, 5e-6:1500, 5e-7:10000, 5e-8:20000
Is there a way to do that?
Thanks!

I'm wondering why the regularisation images are so rough? these are images that are generated with another SD pass
is the goal to have lots of variety and avoid plain / trained photos of men/women etc?

!git clone https://github.com/laurenceday/Ion-Cannon-Regularisation-{target_class}.git

https://github.com/laurenceday/Ion-Cannon-Regularisation-man/blob/main/man/man_001.png

there's some quite odd stuff in there

eg