[Qustion] On the training the CNN networks about vsfa HOT 2 CLOSED

Finally, I succeeded in bringing CNN and LSTM together, and I think I could train in correct way now.

First, CNN does not have a time dimension, so I add the time dimension with "for loop" in CNN(for number of frames) and then LSTM goes after the same as before.

The problem is that the longer the epoch, the lower the performance. In my opinion, CNN is already a well-trained neural network, but the lack of learning data makes it rather overfit, and I don't know the exact reason yet.

Oh, by the way, I'm thinking of replacing CNN with Koncept512. Once replaced with koncept512 instead of resnet, performance did not improve significantly. I thought koncept512 was a neural network that measures the mos of images, so I thought it was better to transfer learning than a neural network that classifies images, but I'm not sure why it hasn't improved much yet.

Now I'm going to use multiple CNNs as an ensemble. What I want is for the error of mae to be 0.2 in the validation set. Is it possible? My research senior is asking for 0.2 because he cannot accept a 0.3 mae error. He asked why AI has an error performance of 0.3 out of 5 even if AI is trained, and he said mae that higher 0.1 is not AI.

from vsfa.

@SungmanHong Good luck! This may be related to overfitting problem, and I would reply to you in another issue #38

from vsfa.