search

agentmorris / MegaDetector

MegaDetector is an AI model that helps conservation folks spend less time doing boring things with camera trap images.
MIT License
116 stars 24 forks source link

model layer structure for frozen training #78

Closed agentmorris closed 1 year ago

agentmorris commented 1 year ago

I am training for specific species using a frozen backbone as discussed here: https://docs.ultralytics.com/tutorials/transfer-learning-froze-layers/

I can’t find any documentation on the layer structure for megadetector v5. Is it the same as yolov5 with layers model.0.' - 'model.9 making up the backbone?

Thanks

Issue cloned from Microsoft/CameraTraps, original issue posted by ehallein on Oct 25, 2022.

agentmorris commented 1 year ago

We did no modification to the architecture; MegaDetector is a trained instance of YOLOv5x6. I can't find a great representation of YOLOv5x6 specifically, but you can visualize the layers with, e.g., netron.app.

Hope that helps!

(Comment originally posted by agentmorris)

agentmorris commented 1 year ago

Excellent, knowing the model used gives me the answer. So did you train at 1280 pixels?

(Comment originally posted by ehallein)

agentmorris commented 1 year ago

Yes. More specifically, we trained with --imgsz 1280, which means "set the long side of each image to 1280 and adjust the short side to maintain aspect ratio". This appears to be standard for YOLOv5x6.

(Comment originally posted by agentmorris)