I'm working on a Bacteria detection project.
My question is : is it releavant to use pretrained weights and only train a model on few last layers to detect unusual objects ? Or should i train the model from scratch ? (i have arround 2000 images of bacterias and I did'nt use any data augmentation yet) I tried to train from scratch yoloV3 on the tiny-yolo model but i have a min loss of 4... Any suggestions ? Any prefered model ?
I advice not to train the model from scratch as it will take a lot of your time and compute power. The darknet feature extractor of YOLO (darknet-19 in YOLOv2 and darknet-53 in the case of YOLOv3) are first trained on Imagenet dataset for getting a feature extractor which is a lengthy process.
You should use the feature extractor as it is and train the layers after it. YOLOv2 should work fine at some reasonable speed if you don't want very precise bounding boxes, but if you cannot work with some localization errors and want tight bounding boxes, go with YOLOv3.
For more tranining guidelines follow this link. AlexeyAB has curated all the steps to be followed to train YOLO on custom dataset.
