In what order should I reproduce the paper?

[rixejzvdl649]

step1
pretrain_projector_image_encoder.sh
step2
pretrain_projector_video_encoder.sh
step3
finetune_dual_encoder.sh
step4
eval/vcgbench/inference/run_ddp_inference.sh
step5
eval/vcgbench/gpt_evaluation/vcgbench_evaluate.sh

#!/bin/sh


export DATASET_DIR=/mnt2/ninghuayang/data/videogpt_plus_dataset

BASE_LLM_PATH=microsoft/Phi-3-mini-4k-instruct
VISION_TOWER=OpenGVLab/InternVideo2-Stage2_1B-224p-f4
IMAGE_VISION_TOWER=openai/clip-vit-large-patch14-336
PROJECTOR_TYPE=mlp2x_gelu
#PRETRAIN_VIDEO_MLP_PATH=MBZUAI/VideoGPT-plus_Phi3-mini-4k_Pretrain/mlp2x_gelu_internvideo2/mm_projector.bin
#PRETRAIN_IMAGE_MLP_PATH=MBZUAI/VideoGPT-plus_Phi3-mini-4k_Pretrain/mlp2x_gelu_clip_l14_336px/mm_projector.bin
PRETRAIN_VIDEO_MLP_PATH=results/mlp2x_gelu_internvideo2/mm_projector.bin
PRETRAIN_IMAGE_MLP_PATH=results/mlp2x_gelu_clip_l14_336px/mm_projector.bin
OUTPUT_DIR_PATH=results/videogpt_plus_finetune

deepspeed videogpt_plus/train/train.py \
--lora_enable True --lora_r 128 --lora_alpha 256 --mm_projector_lr 2e-5 \
--deepspeed scripts/zero3.json \
--model_name_or_path "$BASE_LLM_PATH" \
--version phi3_instruct \
--dataset_use FINETUNING \
--vision_tower "$VISION_TOWER" \
--image_vision_tower "$IMAGE_VISION_TOWER" \
--mm_projector_type "$PROJECTOR_TYPE" \
--image_mm_projector_type "$PROJECTOR_TYPE" \
--pretrain_mm_mlp_adapter "$PRETRAIN_VIDEO_MLP_PATH" \
--pretrain_image_mm_mlp_adapter "$PRETRAIN_IMAGE_MLP_PATH" \
--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 $OUTPUT_DIR_PATH \
--num_train_epochs 1 \
--per_device_train_batch_size 24 \
--per_device_eval_batch_size 16 \
--gradient_accumulation_steps 2 \
--evaluation_strategy "no" \
--save_strategy "steps" \
--save_steps 50000 \
--save_total_limit 1 \
--learning_rate 2e-4 \
--weight_decay 0. \
--warmup_ratio 0.03 \
--lr_scheduler_type "cosine" \
--logging_steps 1 \
--tf32 True \
--model_max_length 4096 \
--gradient_checkpointing True \
--dataloader_num_workers 16 \
--lazy_preprocess True \
--report_to none

[rixejzvdl649]

In the official example, the two benchmarks each have their own weights

VideoGPT-plus/MBZUAI/VideoGPT-plus_Phi3-mini-4k/mvbench
VideoGPT-plus/MBZUAI/VideoGPT-plus_Phi3-mini-4k/vcgbench

[rixejzvdl649]

step1
pretrain_projector_image_encoder.sh
step2
pretrain_projector_video_encoder.sh
step3
finetune_dual_encoder.sh
step4
eval/vcgbench/inference/run_ddp_inference.sh
step5
eval/vcgbench/gpt_evaluation/vcgbench_evaluate.sh

So besides the above setp123. and step45, is there any other information or steps I missed?

[rixejzvdl649]

from .dataset_config import *

DataConfig = {
    "PRETRAINING": [CC3M_595K, COCO_CAP, COCO_REG, COCO_REC],

    "FINETUNING": [CONV_VideoChatGPT, VCG_HUMAN, VCG_PLUS_112K, CAPTION_VIDEOCHAT, CLASSIFICATION_K710, CLASSIFICATION_SSV2, CONV_VideoChat1, REASONING_NExTQA, REASONING_CLEVRER_QA, REASONING_CLEVRER_MC, VQA_WEBVID_QA],

    "VCGBench_FINETUNING": [CONV_VideoChatGPT, VCG_HUMAN, VCG_PLUS_112K, CAPTION_VIDEOCHAT, CONV_VideoChat1, VQA_WEBVID_QA],
    "MVBench_FINETUNING": [CLASSIFICATION_K710, CLASSIFICATION_SSV2, CONV_VideoChatGPT, REASONING_NExTQA, REASONING_CLEVRER_QA, REASONING_CLEVRER_MC, VQA_WEBVID_QA],

}

[rixejzvdl649]

I didn't use VCGBench_FINETUNING and MVBench_FINETUNING. Will there be any problems?

[mmaaz60]

Hi @rixejzvdl649,

Thank you for your interest in our work and providing the detailed information about your question. The steps you mentioned to reproduce our results (pretraining + finetuning + evaluation) are correct.

However, please note that we finetune two models/variants of VideoGPT+. The first variant finetuned using VCGBench_FINETUNING is used to evaluate on VCGBench and VCGBench-Diverse, and the second variant finetuned on MVBench_FINETUNING is used to evaluate on MVBench.

I hope it will help. Please let me know if you have any questions.

[qianwangn]

@mmaaz60 hello, does stage1 and stage2 can parallel?