Guidelines on setting up config for custom datasets?

[bhosalems]

Hi, I am using action former as a baseline in one of our works, I want to setup the config for our dataset. Is there any guide on how to set the parameters in the config, for now, I have populated almost all the fields from the THUMOS config file.

[tzzcl]

For most custom datasets, I think hyperparameters from THUMOS14 is OK, and you just need to change the stride, feature dimension, etc.

[bhosalems]

Okay, what's the number of frames?

[tzzcl]

num_frames means each clip uses num_frames of input frames to generate that figure, for example, I3D uses 16 frames to generate one feature vector.

[bhosalems]

I don't understand what parameter that would be from the I3D extraction script, I have little knowledge of the I3D feature extraction method recommended by you here https://github.com/Finspire13/pytorch-i3d-feature-extraction?

These are the parameters - --mode: rgb or flow --load_model: path of the I3D model --input_dir: folder of converted videos --output_dir: folder of extracted features --batch_size: batch size for snippets --sample_mode: oversample, center_crop or resize --frequency: how many frames between adjacent snippet --usezip/no-usezip: whether the frame images are zipped

[happyharrycn]

We have limited capacity and can only address questions directly related to this repo. For I3D, please refer to their paper and the repo for more details.