Skip to main content

Which model for Experimentations?

Created on February 21|Last edited on February 21

Introduction

This report results from experimenting on which model I should use going forward when running experiments.
The models used are, in descending order of their size:
  • convnext_small
  • convnext_tiny
  • resnet18 (which I was using for the previous experiments)

The training was done on 10 classes for 10 epochs using Adam, and unfreezing after 1 epoch (Therefore 11 epochs in total).

Results


Run set
19



Run set
19


Discussion and Conclusion

The bigger the model, the better the result I got as you can see in the charts above. However, more time is taken waiting for it to train for the bigger models.
Convnext_tiny is the best model to experiment with because it gives relatively good results but also trains in considerable time.