关于结果复现的问题

[Cooperxjt]

您好，我正在尝试复现您的工作，我下载了数据集和预训练模型，但是在第一个ScienceQA任务的训练上，结果就比较低。准确率只有72%左右，和文中的80%有较大的差距。下面是我用的训练代码和测试代码，其他部分没有修改。

训练代码：
################## VICUNA ##################
PROMPT_VERSION=v1
MODEL_VERSION="vicuna-7b-v1.5"
################## VICUNA ##################

################## LLaMA-2 ##################
# PROMPT_VERSION="llava_llama_2"
# MODEL_VERSION="Llama-2-7b-chat-hf"
################## LLaMA-2 ##################

deepspeed --include localhost:0,1,2,3,4,5,6,7 --master_port 29600 ETrain/Train/LLaVA/train_mem.py \
    --deepspeed ./scripts/zero3_offload.json \
    --lora_enable True --lora_r 128 --lora_alpha 256 --mm_projector_lr 2e-5 \
    --expert_num 8 \
    --model_name_or_path ./checkpoints/LLaVA/Vicuna/vicuna-7b-v1.5 \
    --pretrain_mm_mlp_adapter /root/paddlejob/workspace/env_run/zhangshuo/pretrained_model/llava-v1.5-mlp2x-336px-pretrain-vicuna-7b-v1.5/mm_projector.bin \
    --version $PROMPT_VERSION \
    --data_path ./playground/Instructions_Original/ScienceQA/train.json \
    --image_folder ./cl_dataset \
    --vision_tower /root/paddlejob/workspace/env_run/zhangshuo/pretrained_model/clip-vit-large-patch14-336 \
    --mm_projector_type mlp2x_gelu \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --image_aspect_ratio pad \
    --group_by_modality_length True \
    --bf16 True \
    --output_dir ./checkpoints/LLaVA/CoIN/ScienceQA_llava_MOE_lora \
    --num_train_epochs 1 \
    --per_device_train_batch_size 14 \
    --per_device_eval_batch_size 16 \
    --gradient_accumulation_steps 8 \
    --evaluation_strategy "no" \
    --save_strategy "epoch" \
    --learning_rate 2e-4 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --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 none

测试代码
#!/bin/bash

gpu_list="${CUDA_VISIBLE_DEVICES:-0}"
IFS=',' read -ra GPULIST <<< "$gpu_list"

CHUNKS=${#GPULIST[@]}

if [ ! -n "$1" ] ;then
    STAGE='Finetune'
else
    STAGE=$1
fi

if [ ! -n "$2" ] ;then
    MODELPATH='./checkpoints/Instruction/Only_Pretrain_1.5/ScienceQA/llava-1.5-7b-lora'
else
    MODELPATH=$2
fi

RESULT_DIR="./results/CoIN/LLaVA/ScienceQA"

for IDX in $(seq 0 $((CHUNKS-1))); do
    CUDA_VISIBLE_DEVICES=${GPULIST[$IDX]} python -m ETrain.Eval.LLaVA.CoIN.model_vqa_science \
        --model-path $MODELPATH \
        --model-base ./checkpoints/LLaVA/Vicuna/vicuna-7b-v1.5 \
        --question-file ./playground/Instructions_Original/ScienceQA/test.json \
        --image-folder ./cl_dataset \
        --answers-file $RESULT_DIR/$STAGE/${CHUNKS}_${IDX}.jsonl \
        --num-chunks $CHUNKS \
        --chunk-idx $IDX \
        --temperature 0 \
        --conv-mode vicuna_v1 &
done

wait

output_file=$RESULT_DIR/$STAGE/merge.jsonl

# Clear out the output file if it exists.
> "$output_file"

# Loop through the indices and concatenate each file.
for IDX in $(seq 0 $((CHUNKS-1))); do
    cat $RESULT_DIR/$STAGE/${CHUNKS}_${IDX}.jsonl >> "$output_file"
done

python -m ETrain.Eval.LLaVA.CoIN.eval_science_qa \
    --base-dir ./cl_dataset/ScienceQA \
    --result-file $output_file \
    --output-file $RESULT_DIR/$STAGE/output.jsonl \
    --output-result $RESULT_DIR/$STAGE/output_result.jsonl \

我使用的预训练模型链接如下：
https://huggingface.co/liuhaotian/llava-v1.5-mlp2x-336px-pretrain-vicuna-7b-v1.5
https://huggingface.co/lmsys/vicuna-7b-v1.5