SilvioGiancola / SoccerNet-code

SoccerNet: A Scalable Dataset for Action Spotting in Soccer Videos
106 stars 25 forks source link

i3d. py and rgb_scratch - rgb_imagenet checkpoints #8

Closed draxler1 closed 4 years ago

draxler1 commented 4 years ago

Dear Sir, I am trying to use feature_extractor using i3d script you provide. Line number 20 in extract_i3d_spatial_features.py that is "import i3d" not working, can ypu please provide me i3d.py class file in your repository.

Also,

_CHECKPOINT_PATHS = {
    'rgb': 'kinetics-i3d/data/checkpoints/rgb_scratch/model.ckpt',
    'rgb_imagenet': 'kinetics-i3d/data/checkpoints/rgb_imagenet/model.ckpt'
}

are missing in main deepmind kinetic-i3d repository. Those folder contained: checkpoint model.ckpt.data-00000-of-00001
model.ckpt.index model.ckpt.meta

can you please share 'kinetics-i3d/data/checkpoints/rgb_scratch/model.ckpt' and 'kinetics-i3d/data/checkpoints/rgb_imagenet/model.ckpt'

Thanks.

SilvioGiancola commented 4 years ago

Dear @draxler1,

Make sure you have cloned the SoccerNet-code repo with the kinetics-i3d submodule, or re-initialize the submodule, that's the reason of your issue importing i3d, your kinetics-i3d folder looks empty to me be empty.

The code for feature extraction originate from its original github repo, I have added it here as a submodule. The models your are looking for are available here and here.

I recommend you to use the feature provided with the the dev kit as I3D require 8 GPUs for the flow estimation and is extremely slow. If you insist on extracting your own features, you can follow the instruction on the kinetics-i3d repo as they may have released new models trained on kinetics-600.

Best

SilvioGiancola commented 4 years ago

Actually, you should clone the I3D repo following those instructions.

draxler1 commented 4 years ago

Thanks for quick response @SilvioGiancola I have followed exactly what you specified in reply. Now, following error popping up.

File "extract_i3d_spatial_features.py", line 232, in <module>
    main(args.video_list_dir, args.feature_dir, overwrite=args.overwrite)  File "extract_i3d_spatial_features.py", line 104, in main
    _NUM_CLASSES, spatial_squeeze=True, final_endpoint='PreLogits')
  File "/floyd/home/src/feature_extraction/i3d-feat-extract/i3d.py", line 137, in __init__
    raise ValueError('Unknown final endpoint %s' % final_endpoint)
ValueError: Unknown final endpoint PreLogits

Seeking help :-(
SilvioGiancola commented 4 years ago

I'm sorry, but I can't provide support to extract I3D features as it is not the scope of SoccerNet and it's not my code but rather deepmind's. Please check with them to run it on your own videos. I am already providing the I3D features for all the videos of SoccerNet [pre-zipped here].