Last epoch performance or best epoch performance?
Closed this issue · 2 comments
Hi, I am reproducing your Table 3. I want to ask whether the results in Table 3 come from last epoch performance or best epoch performance?
On PathMnist with ResNet-18(28), my last epoch AUC is 0.985 which is close to the paper, but my best epoch AUC is 0.994. The following figure is the test AUC during training procedure.
Moreover, do you think the learning rate is too large? As you can see in the picture, the test performance fluctuates greatly and does not converge during the training. I wander how do you select lr=0.001, do you observe that such lr performs best on most datasets?
As in the paper and the code, we use the best performing model on the validation set, which is the standard practice in machine learning.
As in the paper and the code, we use the best performing model on the validation set, which is the standard practice in machine learning.
Gotcha! Thanks for the reply.