make fine-tuning script public?
#5
by
wdli
- opened
Hi thanks for the great work! Do you have plan to make the fine-tuning script public?
Hello thank for the reply. I was using the diffusers script that the link points to. But I got this error:ValueError: Module down_blocks.0.attentions.0.transformer_blocks.0.attn1.to_k is not a LoRACompatibleConv or LoRACompatibleLinear module.
when trying to load the lora on stable diffusion. It's pretty weird. Have you ever encounterd this?
I'm using stable diffusion 2-1-base. and I stored the lora checkpoint in my local folder. The comamnd I'm using is:
export MODEL_NAME="stabilityai/stable-diffusion-2-1-base"
export OUTPUT_DIR="<output_path>"
export TRAIN_DATA_DIR="<dataset_path>"
accelerate launch --multi_gpu --num_processes=2 train_text_to_image_lora.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--train_data_dir=$TRAIN_DATA_DIR \
--dataloader_num_workers=8 \
--resolution=512 \
--train_batch_size=16 \
--gradient_accumulation_steps=1 \
--max_train_steps=15000 \
--learning_rate=1e-04 \
--max_grad_norm=1 \
--lr_scheduler="cosine" \
--lr_warmup_steps=0 \
--mixed_precision="fp16" \
--output_dir=${OUTPUT_DIR} \
--report_to=wandb \
--checkpointing_steps=200 \
--seed=42