Hi folks. I have built an object classification model. To me, the curve looks like underfitting. But I used an early stopping option, which means the model training would have stopped automatically after it stopped improving, so I know that training time or number of epochs is not responsible for the underfitting. Can anyone else interpret this curve for me? If it is underfitting, can the curve itself suggest what the problem may be? Any help really appreciated.
Looks to me like your early-stopping has stopped too soon. The training cost has not stabilized.
I think its better to look at the trend in the validation loss (orange curve) rather than the training loss. It seems that the validation loss is not decreasing any more beyond around 85. In fact it started increased (sign of overfitting) and has maintained this trend for a while. That indicates you probably should stop training.
BTW your x-axis says # of batches processed. Generally the number of epochs should be plotted on the x-axis. Otherwise the trends can be confusing. E.g., losses may just happen to increase for some batches by accident (e.g. if these batches just happen to have a bigger portion of harder to classify data points), although the training isn’t reallycomplete yet
If this is from mini-batch training, the cost may not decrease uniformly, since you may at any time encounter new examples that were not previously included. This can cause the cost to suddenly increase.
This is not necessarily a problem. It just makes the cost curves difficult to interpret.
Can you say a bit more about your training method?
1 Like