Skip to content

Latest commit

 

History

History
76 lines (61 loc) · 2.92 KB

docker-helper-commands.md

File metadata and controls

76 lines (61 loc) · 2.92 KB
Raw

Shortcut for rebuilding and relaunching

Before launching it, make sure you have dataset directory here populated with subject images.

The following Docker run command mounts the image's /output directory to the local one named output, so we can get the output of the training outside the container easily.

docker rm -f train-flux-kohya-sd-scripts && \
    docker build -t kopyl/train-flux-kohya-sd-scripts . && \
    docker run \
        --name train-flux-kohya-sd-scripts \
        -d \
        -v ./output:/output \
        -v ./dataset:/dataset \
        -v ./dataset-config.toml:/dataset-config.toml \
        -v ./sample_prompts.txt:/sample_prompts.txt \
        --gpus all \
        --shm-size 8G \
        --env OUTPUT_MODEL_NAME="finetuned-model" \
        --env SAVE_EVERY_N_EPOCHS=70 \
        --env MAX_TRAIN_EPOCHS=300 \
        --env SAMPLE_EVERY_N_EPOCHS=10 \
        --env LEARNING_RATE="1.1e-5" \
        kopyl/train-flux-kohya-sd-scripts && \
    docker logs train-flux-kohya-sd-scripts -f

docker exec -it train-flux-kohya-sd-scripts bash

The environment variables are just the defaults. Unless not provided, the same values are going to be used

The most common use case:

  1. Clone the repo with (git must be installed):
git clone https://github.com/kopyl/flux-training-docker-kohya.git
  1. cd into the repo with
cd flux-training-docker-kohya
  1. Create directory with name dataset and put your images there;
  2. Change class_tokens inside dataset-config.toml file to your subject token. E.g: if it's a man, change 'sks girl' to 'sks man';
  3. Change the class token in inside sample_prompts.txt file. E.g: if you change the class token, you need to change it in the prompts list as well; (prompt formatting syntax)
  4. Launch the docker container to start the training with:
docker run \
    --name train-flux-kohya-sd-scripts \
    -d \
    -v ./output:/output \
    -v ./dataset:/dataset \
    -v ./dataset-config.toml:/dataset-config.toml \
    -v ./sample_prompts.txt:/sample_prompts.txt \
    -v ./run-training.sh:/run-training.sh \
    --gpus all \
    --shm-size 8G \
    kopyl/train-flux-kohya-sd-scripts
  1. Sit back and watch the logs with
docker logs train-flux-kohya-sd-scripts -f

Each 10 epochs there are going to be generated the images from your prompts inside /output/sample directory (which you mount to ./output/sample on your host with the docker run command above). When the training is finished, you get the model name finetuned-model.safetensors inside /output directory. For more info on how to generate images with the train models, read But how am I supposed to use that training model in Diffusers? section.