How to use pretrained ResNext152 model

[RorschachChen]

Thanks for your codes. I had to admit it's a wonderful strategy.
However, when I use this package on the action recognition dataset Stanford40, I encounter the loss explosion problem, so I am thinking about using pre-trained model.
I had decreased the classes from 40 to 10. and turned the mode to fully supervised learning with exclude_unlabled as 'True'. Hope you have time to give a reply even a little hint.
Here I print out the loss at each step until loss explosion. The Res Loss increase like crazy.
AssertionError: Loss explosion: 226970.828125
0 batch
class Variable containing:
2.3374
const Variable containing:
1.00000e-02 *
2.4998
res Variable containing:
1.00000e-02 *
1.0730
1 batch
class Variable containing:
12.6847
const Variable containing:
275.5649
res Variable containing:
1.00000e+05 *
2.2668