Kohya SDXL style run

style.sh

#!/bin/bash

# Config Start

# Configurations
ckpt="checkpoint path here" # base checkpoint to finetune
image_dir="training path here, should have images as subfolder" # folder containing folders with repeats_conceptname
reg_dir="" #optional, just point this to an empty folder if you don't care
output="output path here" # safetensors output folder

learning_rate=1
lr_warmup_ratio=0
train_batch_size=5
num_epochs=15
save_every_n_epochs=1
scheduler="linear"
network_dim=32
text_encoder_lr=0
unet_lr=1

# Config End

echo "Image directory: $image_dir"
ls "$image_dir"

while IFS= read -r -d $'\0' dir; do
    dirname=$(basename "$dir")
    IFS='_' read -r -a parts <<< "$dirname"
    
    if [ "${#parts[@]}" -ne 2 ]; then
        echo "Directory name $dirname does not follow expected format."
        continue
    fi
    
    repeats=${parts[0]}
    concept=${parts[1]}
    echo "Processing: $dirname, Repeats: $repeats, Concept: $concept"
    
    imgs=$(find "$dir" -type f \( -iname "*.png" -o -iname "*.bmp" -o -iname "*.gif" -o -iname "*.jpg" -o -iname "*.jpeg" -o -iname "*.webp" \) | wc -l)
    img_repeats=$((repeats * imgs))
    
    echo -e "\t$concept: $repeats repeats * $imgs images = $img_repeats"
    total=$((total + img_repeats))
done < <(find "$image_dir" -mindepth 1 -maxdepth 1 -type d -print0)

# Calculations based on total images
total=$imgs
mts=$((total / train_batch_size * num_epochs))
lr_warmup_steps=$(echo "$mts * $lr_warmup_ratio" | bc)

echo "Total images with repeats: $total"
echo "Max training steps $total / $train_batch_size * $num_epochs = $mts"
echo "LR Warmup Steps: $lr_warmup_steps"

# Activate the virtual environment
source ./venv/bin/activate

# Launch training script with parameters
accelerate launch --num_cpu_threads_per_process 16 sdxl_train_network.py \
    --cache_latents \
    --enable_bucket \
    --network_train_unet_only \
    --min_bucket_reso=512 \
    --max_bucket_reso=2048 \
    --bucket_reso_steps=256 \
    --max_data_loader_n_workers=1 \
    --persistent_data_loader_workers \
    --pretrained_model_name_or_path="$ckpt" \
    --train_data_dir="$image_dir" \
    --reg_data_dir="$reg_dir" \
    --resolution=1024,1024 \
    --optimizer_type="Prodigy" \
    --optimizer_args "weight_decay=0.01" "betas=0.9,0.999" "d_coef=0.8" "use_bias_correction=True" "safeguard_warmup=False" \
    --output_dir="$output" \
    --train_batch_size=$train_batch_size \
    --lr_scheduler="$scheduler" \
    --lr_warmup_steps=$lr_warmup_steps \
    --max_train_steps=$mts \
    --multires_noise_discount=0.3 \
    --prior_loss_weight=1 \
    --gradient_checkpointing \
    --xformers \
    --mixed_precision=bf16 \
    --save_every_n_epochs=$save_every_n_epochs \
    --seed=1234 \
    --save_precision=bf16 \
    --logging_dir="" \
    --caption_extension=.txt \
    --save_model_as=safetensors \
    --network_module=networks.lora \
    --text_encoder_lr=$text_encoder_lr \
    --unet_lr=$unet_lr \
    --network_dim=$network_dim \
    --network_alpha=$network_dim \
    --output_name="yd_sdxl_linearv6"