OpenGVLab / VideoMamba

[ECCV2024] VideoMamba: State Space Model for Efficient Video Understanding
https://arxiv.org/abs/2403.06977
Apache License 2.0
833 stars 59 forks source link

Evaluation of video_sm using trained weights #87

Open ManyThanks1204 opened 1 month ago

ManyThanks1204 commented 1 month ago

!/bin/bash

export CUDA_VISIBLE_DEVICES=1 export MASTER_PORT=$((12000 + RANDOM % 20000)) export OMP_NUM_THREADS=1

VIDEO_PATH='name.mp4' OUTPUT_DIR='k10' WEIGHT_PATH='checkpoint-latest.pth'

python run_class_finetuning.py \ --model videomamba_middle \ --finetune ${WEIGHT_PATH} \ --eval_data_path ${VIDEO_PATH} \ --output_dir ${OUTPUT_DIR} \ --batch_size 1 \ --num_sample 1 \ --input_size 224 \ --short_side_size 224 \ --test_num_segment 4 \ --test_num_crop 1 \ --eval I used this infer.sh and tried to evaluate a video, but it raised an error: "FileNotFoundError: [Errno 2] No such file or directory: 'you_data_path/test.csv'". Could you please give a detailed guide on how to evaluate a video? Thank you for your great work!

Andy1621 commented 1 month ago

If you want to use this script, please follow DATA.md to prepare the csv. After testing, it will generate some test files with softmax score.

If you only want to infer one video, please prepare the data transform as in the dataset, and them feedforward the sampled frames, thus obtain the softmax score.