LoRA Trainer | Worker

This is a implementation https://github.com/kohya-ss/sd-scripts Derived from e1pupper's RunPod worker implementation https://github.com/e1pupper/loratrainer

1. Clone this repo
2. Cache the base model

Inputs

Name	Type	Description	Default	Required
zip_url	str	URL to the zip file containing the data	None	✔️
instance_name	str	Name of the model instance	None	✔️
class_name	str	Name of the model class	None	✔️
unet_lr	float	Learning rate of the U-Net model	0.0001
network_dim	int	Dimension of the neural network	256
lr_scheduler_num_cycles	int	Number of cycles of the learning rate scheduler	1
learning_rate	float	Global learning rate	0.0001
lr_scheduler	str	Type of the learning rate scheduler	'cosine'
lr_warmup_steps	int	Number of steps for the learning rate warmup	280
train_batch_size	int	Batch size for training	1
max_train_steps	int	Maximum number of training steps	1250
mixed_precision	str	Precision type used for mixed precision training	'fp16'
save_precision	str	Precision type used when saving the model	'fp16'
optimizer_type	str	Type of the optimizer used in training	'AdamW8bit'
max_data_loader_num_workers	int	Maximum number of workers for the data loader	0
steps	int	Number of steps to be taken during the training process	125

Example

{
    "input":{
        "zip_url": "https://github.com/runpod-workers/sample-inputs/raw/main/images/froggy.zip",
        "instance_name": "daiton",
        "class_name": "frog",
        "unet_lr": 0.0001,
        "network_dim": 256,
        "lr_scheduler_num_cycles": 1,
        "learning_rate": 0.0001,
        "lr_scheduler": "cosine",
        "lr_warmup_steps": 280,
        "train_batch_size": 1,
        "max_train_steps": 1250,
        "mixed_precision": "fp16",
        "save_precision": "fp16",
        "optimizer_type": "AdamW8bit",
        "max_data_loader_num_workers": 0,
        "steps": 125
    }
}

accelerate launch --num_cpu_threads_per_process 1 train_network.py \
--enable_bucket \
--pretrained_model_name_or_path="cache/v1-5-pruned.safetensors" \
--train_data_dir="/root/sd-scripts/input_imgs" \
--resolution=512,512 \
--output_dir="./training/model" \
--output_name="froggy_lora" \
--save_model_as=safetensors \
--network_module=networks.lora \
--cache_latents \
--bucket_reso_steps=64 \
--bucket_no_upscale