search

mit-han-lab / temporal-shift-module

[ICCV 2019] TSM: Temporal Shift Module for Efficient Video Understanding
https://arxiv.org/abs/1811.08383
MIT License
2.05k stars 418 forks source link

Acc1 about training HMDB51 pretrained on Imagenet #219

Closed VanChilly closed 1 year ago

VanChilly commented 2 years ago

I use the code(also the script for ucf) train the HMDB51 datasets and the resnet50 pretrained on imagenet. Finally I got 51% top1 Acc, is that reasonable? I just want to know whether i got a reasomable results. Thank you very much

VanChilly commented 2 years ago

this is my command

python -u main.py hmdb51 RGB --arch resnet50 --num_segments 8 --gd 20 --lr 0.001 --lr_steps 10 20 --epochs 25 --batch-size 8 -j 8 --dropout 0.8 --consensus_type=avg --eval-freq=1 --shift --shift_div=8 --shift_place=blockres > log.log
zhoubin-me commented 2 years ago

same here, maximum 51% acc for 50 epochs at [20, 40] decay.

zhoubin-me commented 2 years ago

@VanChilly I think the result are obtained by a pre-trained model from K400 then finetune on HMDB

VanChilly commented 1 year ago

same here, maximum 51% acc for 50 epochs at [20, 40] decay.

So, it's clear that the result is reasonable for HMDB pre-trained on ImageNet. 👍

VanChilly commented 1 year ago

@VanChilly I think the result are obtained by a pre-trained model from K400 then finetune on HMDB

Yeah, the author mentioned that in the paper. Thank you.