How to fine-tune LLaVA-OV our my own datasets?

[FeiElysia]

Dear authors,

Thanks for your promising work, I am trying to fine-tune LLaVA-OV on my own datasets, I modified the finetune_onevision.sh as follows:

export OMP_NUM_THREADS=8
export NCCL_IB_DISABLE=0
export NCCL_IB_GID_INDEX=3
export NCCL_SOCKET_IFNAME=eth0
export NCCL_DEBUG=INFO

LLM_VERSION="pretrained_ckpts/llava-onevision-qwen2-0.5b-ov" 
# for 7b model we recommend bs=1, accum=2, 16 nodes, 128 gpus, lr=1e-5, warmup=0.03
# for 72b model we recommend bs=1, accum=1, 32 nodes, 256 gpus, lr=1e-5, warmup=0.03
LLM_VERSION_CLEAN="${LLM_VERSION//\//_}"
VISION_MODEL_VERSION="google/siglip-so400m-patch14-384"
VISION_MODEL_VERSION_CLEAN="${VISION_MODEL_VERSION//\//_}"

############### Pretrain ################

PROMPT_VERSION="qwen_1_5"

BASE_RUN_NAME="llavanext-${VISION_MODEL_VERSION_CLEAN}-${LLM_VERSION_CLEAN}-mlp2x_gelu-pretrain_blip558k_plain"
echo "BASE_RUN_NAME: ${BASE_RUN_NAME}"

RANK=${RANK:-0}
ADDR=${ADDR:-"127.0.0.1"}
PORT=${PORT:-"29501"}
NNODES=${NNODES:-1}
NUM_GPUS=${NUM_GPUS:-1}
MID_RUN_NAME="finetune-qwen05b-test"

CKPT_PATH=$LLM_VERSION # this could also be the previous stage checkpoint

torchrun --nproc_per_node="${NUM_GPUS}" --nnodes="${NNODES}" --node_rank="${RANK}" --master_addr="${ADDR}" --master_port="${PORT}" \
    llava/train/train_mem.py \
    --deepspeed scripts/zero3.json \
    --model_name_or_path ${CKPT_PATH} \
    --version ${PROMPT_VERSION} \
    --data_path ./onevision.yaml \
    --image_folder /datasets/source_images \
    --pretrain_mm_mlp_adapter pretrained_ckpts/llava-onevision-projectors/0.5b/mm_projector.bin \
    --mm_tunable_parts="mm_vision_tower,mm_mlp_adapter,mm_language_model" \
    --mm_vision_tower_lr=2e-6 \
    --vision_tower="/${VISION_MODEL_VERSION}" \
    --mm_projector_type mlp2x_gelu \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --group_by_modality_length True \
    --image_aspect_ratio anyres_max_9 \
    --image_grid_pinpoints  "(1x1),...,(6x6)" \
    --mm_patch_merge_type spatial_unpad \
    --bf16 True \
    --run_name $MID_RUN_NAME \
    --output_dir "/checkpoints/${MID_RUN_NAME}" \
    --num_train_epochs 1 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 2 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 1000 \
    --save_total_limit 1 \
    --learning_rate 1e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 32768 \
    --gradient_checkpointing True \
    --dataloader_num_workers 4 \
    --lazy_preprocess True \
    --report_to wandb \
    --torch_compile True \
    --torch_compile_backend "inductor" \
    --dataloader_drop_last True \
    --frames_upbound 32

# You can delete the sdpa attn_implementation if you want to use flash attn

But I got the following error:

size mismatch for 0.weight: copying a param with shape torch.Size([896, 1152]) from checkpoint, the shape in current model is torch.Size([0]).
size mismatch for 0.bias: copying a param with shape torch.Size([896]) from checkpoint, the shape in current model is torch.Size([0]).
size mismatch for 2.weight: copying a param with shape torch.Size([896, 896]) from checkpoint, the shape in current model is torch.Size([0]).
size mismatch for 2.bias: copying a param with shape torch.Size([896]) from checkpoint, the shape in current model is torch.Size([0]).

It seems that the projector is not initialized correctly. The code can run when I comment out --pretrain_mm_mlp_adapter. If there any special setting when loading projector?

[hyYY-1024]

哥们你解决这个问题了吗

[NicoZenith]

don't bother with the adapter, check the new training script, you can just start from the already fine-tuned llava-onevision model in --model_name_or_path

[FeiElysia]

Thank you for your comment. Yes, you can run the script if you don't use --pretrain_mm_mlp_adapter. If you still want to use your own adapter, you can set deepspeed to scripts/zero2.json. This can address the mismatching. There are some bugs when using lazy load with zero3.json