Higher performance at early epochs, but no progress at the end. How can i treat the model

Hi, all.

now I modify CNN model for pose estimation on MSCOCO, and check the performance changes. Some modified models actually have enhanced performance at the early epoch such as 6 epoch among 32, but the enhancement is gone when epochs going on and on.

Is there any idea to hold enhancement of modified model to the end of learning? If this phenomenon is typical, then i expect some ways to handle it.

Thank you in advance.

I’m not sure to understand the question properly.
It seems your model is learning at the beginning but stops learning in further epochs.
If that’s the case, it’s a common observation.
You could try to change some hyperparameters, e.g. lower the learning rate when no learning takes place.
Have a look at lr_scheduler.ReduceLROnPlateau to automatically lower the learning rate once a plateau is reached.

1 Like