search

zhang-can / ECO-pytorch

PyTorch implementation for "ECO: Efficient Convolutional Network for Online Video Understanding", ECCV 2018
BSD 2-Clause "Simplified" License
188 stars 93 forks source link

Cant reach 90.4% precision for 4segments ECOFULL. #63

Closed Rayaction closed 4 years ago

Rayaction commented 4 years ago

i use the recommended cmd: [UCF101 - ECO - RGB] command:

python main.py ucf101 RGB \ --arch ECO --num_segments 4 --gd 5 --lr 0.001 --lr_steps 30 60 --epochs 80 \ -b 32 -i 1 -j 1 --dropout 0.8 --snapshot_pref ucf101_ECO --rgbprefix img \ --consensus_type identity --eval-freq 1

but the result at 20 epoch is not quite saitisfied.: only 83%. wating to see the prec tommorrow. hope somebody could tell me are there still some tricks? I see a little bit different change in the configuration between this git and that made by MZO.

Rayaction commented 4 years ago

used the pretrained model from MZO git which pretrained by K400, trained for a night , waked up and checked its precision, best still around 83%, could u give me some advice?

Rayaction commented 4 years ago

ok now i use MZO's code and script and can get 91.89% on 16F, 92.9% on 32F.
btw guys, the first time i use the code, it cant reach that prec since the dataloader cant get right images from videos. check this out, it always get the first image. u should calculate the num of frames yourself, and then it works. secondly, data should be images not videos. thats all. good work mate