Multilabels classification:
Created on October 26|Last edited on October 26
Comment
Best results
Showing first 10 runs
Showing first 10 runs
Showing first 10 runs
Showing first 10 runs
Run set
1282
Is Hamming similarity the best measure?
It is not best. If dataset contains 10 classes, and there are 2 occurring on average, the simple model can guess all labels to be zero. This gives as pretty good 80% accuracy, which is meaningless despite being huge. It is better when we have average number of labels close to max. no. of classes.
Similarly for 8 classes, and 2 labels occurring on average, anything below 75% is below the baseline.
Minibatch is better:
Run set
1282
- Minibatch is better.
Adam is better:
Run set
1282
- Now here Adam is much better. The max values reached are significantly higher. Most of top-20 use Adam.
Activations?
Run set
1282
- On average tanh is better.
Add a comment