search

jsikyoon / pathnet

Tensorflow Implementation of PathNet: Evolution Channels Gradient Descent in Super Neural Networks
https://arxiv.org/pdf/1701.08734.pdf
MIT License
102 stars 24 forks source link

the training costs of the two selected individuals are different. #12

Open czhhbp opened 5 years ago

czhhbp commented 5 years ago

During the implementation of the tournament selection method, the training costs of the two selected individuals are different. The second path individual trains a different net weights baseed on what the first individual has trained. So the second individual always have more training epochs than the first one. It may be wrong when both individuals are compared.

jsikyoon commented 5 years ago

Hi,

Thank you for your interesting about my repo.

This algorithm as I remember, is to use trained NN in previous task.

I can't fully understand second have more training epochs, but when both tasks are similar or commonsense, trained NN in first task helps training for second task.