refactor: Update image size and conversation logic for generating text from images

Author: Luodian

scripts/train/README.md

@@ -2,5 +2,9 @@
 
 We first release the basic training scripts for LLaVA-NeXT. It's based on previous LLaVA's training scripts and researchers familiar with LLaVA will find it easy to use.
 
-We will later release the more detailed training scripts for our LLaVA OneVision models including the mid stage, single-image final stage and one-vision final stage.
-> They are basically the same as the basic training scripts, but with some modifications, such as the data yaml.
+We will gradually release the more detailed training scripts for our LLaVA OneVision models including the mid stage, single-image final stage and one-vision final stage.
+> They are basically the same as the basic training scripts, but with some modifications, such as the data yaml.
+
+- `finetune_clip.sh`: This could be seen as the first image version LLaVA-NeXT (2024-01) training script, with `anyres` strategy and maximum 2x2 image grids.
+- `finetune_siglip.sh`: Same but with `siglip` encoder, each grid becomes 729 tokens.
+- `finetune_onevision.sh`: This is our latest training script, with `anyres_max_9` strategy and image grids weaving from 1x1 to 6x6, at most to 2304x2304 resolution. Inside the script, we also incorporate the multi-image and video data into training loop. the detail token strategy could be found in our paper.

scripts/train/finetune_onevision.sh

@@ -0,0 +1,70 @@
+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="Qwen/Qwen2-7B-Instruct" 
+# 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=plain
+
+BASE_RUN_NAME="llavanext-${VISION_MODEL_VERSION_CLEAN}-${LLM_VERSION_CLEAN}-mlp2x_gelu-pretrain_blip558k_plain"
+echo "BASE_RUN_NAME: ${BASE_RUN_NAME}"
+
+CKPT_PATH=$LLM_VERSION # this could also be the previous stage checkpoint
+
+ACCELERATE_CPU_AFFINITY=1 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 ${PROMPT_VERSION} \
+    --data_path ./onevision_data.yaml \
+    --image_folder ./onevision_data/images \
+    --video_folder ./onevision_data/videos \
+    --pretrain_mm_mlp_adapter="/checkpoints/projectors/${BASE_RUN_NAME}/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

scripts/train/finetune_siglip.sh scripts/train/finetune_siglip_a4.shscripts/train/finetune_siglip.sh renamed to scripts/train/finetune_siglip_a4.sh

@@ -16,10 +16,12 @@ PROMPT_VERSION=plain
 BASE_RUN_NAME="llavanext-${VISION_MODEL_VERSION_CLEAN}-${LLM_VERSION_CLEAN}-mlp2x_gelu-pretrain_blip558k_plain"
 echo "BASE_RUN_NAME: ${BASE_RUN_NAME}"
 
+CKPT_PATH=$LLM_VERSION # this could also be the previous stage checkpoint
+
 ACCELERATE_CPU_AFFINITY=1 torchrun --nproc_per_node="${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 ${LLM_VERSION} \
+    --model_name_or_path ${CKPT_PATH} \
     --version ${PROMPT_VERSION} \
     --data_path=llava_1_6.json \
     --image_folder your_image_folder \