Overfitting traning data.
Closed this issue · 3 comments
I was training the model with cityscapes dataset with 28 classes.
Training images - 2975
Testing - 500
Epoch - 300 (broke at 62 as not getting empirical data to go ahead)
mIOU for various epochs are as follows.
1st Epoch - 32.65
7th Epoch - 34.06
15th Epoch - 13.81
25th Epoch - 12.68
35th Epoch - 12.91
45th Epoch - 17.25
55th Epoch - 20.57
60th Epoch - 17.36
62th Epoch - 20.57
When I tried to run the inference on the model was clearly able to see that it was trying to overfit the training image. I was continuously able to see the structure of the Mercedes logo.
Did you solve the problem? if not can you give more details? setup and what you tired so far?
I have the same problem, could you help me to solve this?
can you share you validation/train loss graphs throughout training?