search

LZU-SIAT / AS-CAL

Augmented Skeleton Based Contrastive Action Learning with Momentum LSTM for Unsupervised Action Recognition
MIT License
30 stars 9 forks source link

Main method confusing among 'moco', 'ins', 'end' #1

Open atifshahcs opened 3 years ago

atifshahcs commented 3 years ago

Hello, I am using your code, but have some confusing with mom_flag. Would you please answer the following query.

  1. You have used three different flags in the code: moco, ins, end, so which one is the main part, and should reflect the reported results in your paper.
  2. The train_moco function uses two split, x1 for model training and x2 for model_ema, while the 'ins' used whole data for model training. Even in linear evaluation the ins method is not implemented.

Would be appreciated if you could answer these queries. Looking forward. Thank you.

Mikexu007 commented 3 years ago

@atifshahcs

  1. “moco” is the main part. "ins" is from “Unsupervised feature learning via non-parametric instance discrimination”. "end" is traditional end to end training.
  2. "moco" and "ins" are just different pre-training methods. After pre-training, the pre-trained model can be used for linear evaluation or other tasks.