I've trained st-gcn on your provided Kinetics dataset and find it converged much slower than on NTU RGB-D. The loss is about 3.0 even in epoch 60.
```
[ Mon Jun 11 03:58:19 2018 ] Training epoch: 58…
I was trying to fine-tuning I3D network using the new TensorFlow dataset/iterator API and get rid of the _feed_dict_ approach. It seems the network cannot build since it can't understand input tensor …