search

ultralytics / yolov3

YOLOv3 in PyTorch > ONNX > CoreML > TFLite
https://docs.ultralytics.com
GNU Affero General Public License v3.0
10.21k stars 3.45k forks source link

Ultralytics Pretrained Weights #236

Closed AmIr-KinG closed 5 years ago

AmIr-KinG commented 5 years ago

Hi Can you please upload the weights that result to ultralytics yolov3 mAP reported in mAP section? It seems like fine-tuning using those weights would be simpler than darknet pre-trained weights.

glenn-jocher commented 5 years ago

@AmIr-KinG mAP results in https://github.com/ultralytics/yolov3#map are provided with the official darknet weightfiles, yolov3-spp.weights, as you can see from the namespace arguments.

For training from scratch you may want to visit https://github.com/ultralytics/yolov3/issues/205.

Note that it is impossible to 'fine-tune' an already trained model like yolov3-spp.weights, otherwise the original authors would have obviously done this themselves. The model has settled into a very specific local minima at an extremely low final LR, and resuming training again from that point (especially with settings defined for epoch 0) will cause your loss to jump initially (and mAP to drop).

AmIr-KinG commented 5 years ago

I understand that I can't fine-tune darknet weights. but since I can't train from scratch due to some limitations, I intend to fine-tune your weights which were obtained by training from scratch. But it doesn't seem like I can find the weights anywhere.

glenn-jocher commented 5 years ago

@AmIr-KinG all of the pytorch weights are available for download from our Google Drive folder. https://github.com/ultralytics/yolov3#pretrained-weights