search

fmthoker / skeleton-contrast

41 stars 5 forks source link

Question about Semi-sup. Evaluation. #8

Closed dingli-dean closed 2 years ago

dingli-dean commented 2 years ago

Hello, thx for sharing this great work. In the repo, I noticed that you have mentioned the experiments about Semi-sup. Evaluation, and the related code for dataset are shown in feeder diretory. However, there still lacks of feeder for this mode, the specific detail is shown as follows: from feeder.feeder_downstream_semi_supervised import Feeder Line 48 in dataset.py. Could you share the feeder_downstream_semi_supervised.py file? Thx again for sharing this work.

fmthoker commented 2 years ago

Here is the feeder_downstream_semi_supervised file. You may have to adapt it to make it work. feeder_downstream_semi_supervised.txt

wb05025 commented 2 years ago

Hi, Thanks a lot for sharing the semi-supervised feeder.

I used the feeder and did modification on your action_classification.py according to the description on your paper (train encoder and classifier together). However, the best accuracy I can reach on the cross-sub 60 dataset using 20% semi-rate was 63.4%, which is a bit lower than your paper result (70.8%).

I'm curious if there is any chance you would consider sharing your original codes for semi-supervised classification or your finetuned hyperparameters for this part? Thank you so much!

fmthoker commented 2 years ago

The hyperparameters of semi-supervised settings are in the supplementary of the arxiv version (https://arxiv.org/pdf/2108.03656.pdf). Also, note the reported numbers are based on sequence-representation, so you have to fine-tune the corresponding trained encoder. I hope this helps.