File size: 1,789 Bytes
fca4fc0 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 |
# Environment Variables
RANK=0,1,2,3,4,5,6,7
MASTER_PORT=29571
# Training Arguments
LOCAL_BATCH_SIZE=4
GRADIENT_ACCUMULATION_STEPS=4
# Path Arguments
export TRANSFORMERS_OFFLINE=1
export WANDB_PROJECT=vtimellm
MODEL_VERSION=vicuna-v1-5
OUTPUT_DIR=./outputs/
STAGE4=./outputs/vtimellm-vicuna-v1-5-7b-youcook2-stage4
RUN_NAME=vtimellm-vicuna-v1-5-7b-youcook2-stage5
deepspeed --include localhost:$RANK --master_port $MASTER_PORT vtimellm/train/train_dpo_mem.py \
--deepspeed ./scripts/zero2.json \
--lora_enable True --lora_r 64 --lora_alpha 128 \
--training_stage 3 --finetuning True \
--model_name_or_path ./checkpoints/vtimellm/vicuna-7b-v1.5 \
--version v1 \
--data_path ./data/YouCook2/dpo-vtimellm/mdpo-train.json \
--data_folder ./data/YouCook2/ \
--feat_folder ./data/YouCook2/clipvitl14-vtimellm.pth \
--pretrain_mm_mlp_adapter ./checkpoints/vtimellm/vtimellm-vicuna-v1-5-7b-stage1/mm_projector.bin \
--stage2_path ./checkpoints/vtimellm/vtimellm-vicuna-v1-5-7b-stage2 \
--stage3_path ./checkpoints/vtimellm/vtimellm-vicuna-v1-5-7b-stage3 \
--stage4_path $STAGE4 \
--output_dir $OUTPUT_DIR/$RUN_NAME \
--bf16 True \
--num_train_epochs 1 \
--per_device_train_batch_size $LOCAL_BATCH_SIZE \
--gradient_accumulation_steps $GRADIENT_ACCUMULATION_STEPS \
--evaluation_strategy "no" \
--save_strategy "no" \
--save_steps 50000 \
--save_total_limit 10 \
--learning_rate 2.5e-6 \
--freeze_mm_mlp_adapter True \
--weight_decay 0. --warmup_ratio 0.1 --lr_scheduler_type "cosine" \
--logging_steps 1 \
--tf32 True \
--model_max_length 2048 \
--gradient_checkpointing True \
--dataloader_num_workers 4 \
--lazy_preprocess True \
--report_to wandb \
--run_name $RUN_NAME \
--gamma 0.0 --beta 0.5 --dpo_alpha 1.0 --train4dpo |