Fine tune a pre-trained computer vision model to test the effectiveness of transfer learning.
Domain: medical field - colonoscopy. 27k training images, 4k validation
Model: YOLOv8x
In order to obtain a higher accuracy, I went a step further and performed hyperparameter variation experiments.
Batch size: 8, 16
Learning rate pairs (initial, final): (0.001, 0.000001), (0.0001, 0.01)
Highest accuracy obtained (mAP): 95.8%
Best performing variation: batch size of 16, learning rate pair of (0.0001, 0.01)