context-art-classification icon indicating copy to clipboard operation
context-art-classification copied to clipboard

Reproducing the MTL model Results

Open mjantoun opened this issue 2 years ago • 0 comments

Hello,

I retrained the MTL model to reproduce the results. I kept the parameters the same as in the code (learning rate, batch size, patience, momentum, Loss weight for each attribute, ResNet50)

I got different accuracy for the Author, School and Timeframe attributes: Author Accuracy (after retraining the MTL model): 0.530 (reported: 0.603) School Accuracy (after retraining the MTL model): 0.671 (reported: 0.691) Timeframe Accuracy (after retraining the MTL model): 0.606 (reported: 0.632)

The only accuracy close to the one reported in the paper is for the Type attribute: Type: 0.790 (reported: 0.791)

Can you please let me know if you changed anything in the parameters?

mjantoun avatar May 26 '22 07:05 mjantoun