khoda81 / dethcod

A neural network based lossless data compression algorithm
MIT License
0 stars 0 forks source link

Model loss becomes nan #2

Open khoda81 opened 1 month ago

khoda81 commented 1 month ago

Actor and critic loss are very high and become nan after a few training steps. @martiny76

khoda81 commented 1 month ago

This seems to be caused by critic_head not being initialized properly which is fixed in f78ecc5b624878dd51f6901aa75ae50668d79859. But the original issue is not clear.