Failed to reproduce linear probing top 1 acc

[potatowarriors]

Hi, I am reproducing the experiment you conducted because it is very interesting. Among them, a linear probing experiment was conducted, but it failed to obtain the top 1 acc of 38.9% presented in Table 3. The top 1 acc I got was 33.9% in 1x1 view condition and 34.7% in 2x3 view condition. If possible, can you share the code that you used to run the linear-prob experiment? It would be a great help to me if you share

thanks

[potatowarriors]

Hi thanks for sending the UCF101 training config but what I asked about in issue#57 was about linear probing. Can you also tell me the linear prob training config?

…

2022. 11. 21. 오후 5:52, xiaojieli0903 ***@***.***> 작성: Set the path to save checkpoints OUTPUT_DIR='/home/gao2/disk/work_dirs/video_mae/ucf_videomae_pretrain_base_patch16_224_frame_16x4_tube_mask_0.75_videos_e3200/eval_lr_5e-4_epoch_100_bs1-16' path to UCF101 annotation file (train.csv/val.csv/test.csv) DATA_PATH='/home/gao2/code/mmaction2/data/ucf101/' path to pretrain model MODEL_PATH='/home/gao2/disk/work_dirs/video_mae/ucf_videomae_pretrain_base_patch16_224_frame_16x4_tube_mask_0.75_videos_e3200/checkpoint.pth' batch_size can be adjusted according to number of GPUs this script is for 8 GPUs (1 nodes x 8 GPUs) OMP_NUM_THREADS=1 python3 -m torch.distributed.launch --nproc_per_node=1 --master_port 12322 run_class_finetuning.py --model vit_base_patch16_224 --data_path ${DATA_PATH} --finetune ${MODEL_PATH} --log_dir ${OUTPUT_DIR} --output_dir ${OUTPUT_DIR} --data_set UCF101 --nb_classes 101 --batch_size 16 --input_size 224 --short_side_size 224 --save_ckpt_freq 50 --num_frames 16 --sampling_rate 4 --num_sample 2 --opt adamw --lr 5e-4 --warmup_lr 1e-8 --min_lr 1e-5 --layer_decay 0.7 --opt_betas 0.9 0.999 --weight_decay 0.05 --epochs 100 --test_num_segment 5 --test_num_crop 3 --fc_drop_rate 0.5 --drop_path 0.2 --use_checkpoint --dist_eval --enable_deepspeed --eval — Reply to this email directly, view it on GitHub <#57 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AVUOXEGA5MNHKFF2RUSUDELWJMZ6HANCNFSM6AAAAAARE6JX7Y>. You are receiving this because you authored the thread.

[honglin-chen]

Hi @potatowarriors, I am trying to reproduce the top 1 acc number in Table 3 as well, but fail to do so. Would you mind sharing about how the 33.9% number was obtained? Thanks in advance!