search

solomesolo / Pallace

2 stars 1 forks source link

Testing performance #4

Closed solomesolo closed 4 years ago

solomesolo commented 4 years ago

Comparing performance. Choosing the best model on that stage.

bogdan-fesenko commented 4 years ago

We have 3-5% improvement in Kappa score (0.679->0.700) and with 360360px image size training compared to 320320px. It would be better to train a network with a bigger size, increasing the input size step by step. 512*512 can require too much GPU memory to train. Also, we can use another network architecture, that will fuse coarse-grained analysis of the whole image size with fine-grained features of the potential region with abnormality with original scale crop image.

bogdan-fesenko commented 4 years ago

Final best result: Ensemble of 4 DenseNet-169 models. Metrics: Kappa score : 0.720 AUC ROC score : 0.870 Accuracy : 0.863