Questions about 'data_info_maker.m' and the epochs about vsfa HOT 5 CLOSED

@stillbetter There exists several (at least two) versions of this dataset.
By the way, according to my records of the downloaded data, KonIQ-10k dataset also has several versions. If you feel confused, you may write an email to the dataset owner.

I downloaded this dataset a long time ago.
During the time, the dataset owner had renamed the video clips, i.e., simplifying the name.
You can re-make the KoNViD-1kinfo.mat with the right CSV file.
What you need to ensure is every video clip has its truly corresponding MOS.

Now I cannot access the konvid-1k-database. You may ask the dataset owner for confirmation.

from vsfa.

Got it.Thanks!

from vsfa.

Sorry for another question, I check the log file, I found the SROCC/PLCC/KROCC can reach a peak performance at around 100 epochs, then the figures get bad . Is that a little bit overfitting? Thanks!

from vsfa.

@stillbetter We use a standard procedure, i.e., training the model on the train data, saving the model based on validation data (That is, we monitor the validation performance, then save the best model.), and then testing the saved model on the test data.

If you see the training loss becomes smaller and smaller and the validation performance first increases and then decreases, you can say the model training is facing the overfitting issue.
This means there is no need to continue the training process unless you have a solution (e.g., add more data) to the overfitting issue. And if you have an elegant training strategy, the model performance may be further improved.

I did not tune the settings separately on each dataset for better performance.
I just set the default epochs as 2000 for all three datasets. The settings of other hyperparameters also matter, e.g., args.decay_interval = int(args.epochs/10).
For this dataset, smaller epochs may be enough if no elegant solution to the overfitting issue.
However, you should also take a look at the curves on the other two datasets.
For the other two datasets, epochs>2000 may obtain a better performance, which means the model did not reach the perfect convergence state in the default settings.

from vsfa.

Great！Thanks for the enthusiastic and patient answner.I'll bother if any question again.

from vsfa.