An Unofficial implementation of 'ACGNet: Action Complement Graph Network for Weakly-supervised Temporal Action Localization'
File Structure:
- model.py contains model building code for ACGNet.
- loss.py contains code implementation for 'Easy Positive Mining' loss in the paper.
Note that the memory footprint could be huge, so batch_size might need to be set a lot smaller than original framework, or cut off num_segments by a large margin.
from model import ACGNet
from loss import loss_EPM
for data, label in data_loader:
orig_F, new_F, A_prime = acg_net(_data)
_, score_act = net(new_F)
loss_epm = loss_EPM(A_prime, orig_F, new_F, score_act)