About Dataset
General
Since imageNet weights are not ideal as a starting point for this task. I have retrained the model on the chest X dataset. You can use the models here as a starting point for your training. that will significantly boost your CV and LB. Unfortunately, I will not share the trained models for two reasons.
1- Keep the LB stable.
2- A lot of Kagglers are just copying the notebook and using the weights. "which was really disappointing"
Notes
To load the model you need to add a classifier (Linear Layer) because I didn't drop the last layers when I saved the weights.
New starting points will be added soon.
Results:
single model, single fold ResNet200D CV:96.7, LB:96.7
single model, single fold Inceptionv3 CV:95.7, LB:95.9
single model, single fold DenseNet121 CV:94.9, LB:95.7