How to do fine-tuning? #11
Comments
|
I wrote wrong command so far. Below one is correct. Then, I have converted weight model with below message. And I also have got error in models.py, So I have modified like below. from: model.load_state_dict(torch.load(weights, map_location='cpu')['model']) |
Dear @dsgh2 , I am so sorry to bother you, but could you please do me a favor? I want to sparse train resnet152_trident.cfg on my custom dataset. However, an error occur when I input the following command:
I find that all the Resnet-structure can not been sparsity train on this project. I give the content of resnet152_trident.cfg as following, looking forward to hearing from you. Thank you very much. [net] Testing#batch=1 Trainingbatch=64 learning_rate=0.001 policy=sgdr [convolutional] [maxpool] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] Conv 4[convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] TridentNet - large objects - Start[convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] Conv 5[convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [yolo] TridentNet - large objects - EndTridentNet - medium objects - Start[route] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] Conv 5[convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [upsample] [route] [convolutional] [convolutional] [yolo] TridentNet - medium objects - EndTridentNet - small objects - Start[route] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] Conv 5[convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [convolutional] [convolutional] [shortcut] [convolutional] [upsample] [route] [convolutional] [convolutional] [yolo] TridentNet - small objects - End |
|
Well, I have never tried without yolov3. My pruned yolov3 model fails to detect anything even though fine-tuning is done. |
Thank you very much for your reply. Looking forward to hearing from you, good luck! |
I am using latest python 3 and pytorch version under latest anaconda.
As we know, there are four steps to train custom model
by step 3, I have got pruned model. And the result is a .pt file.
In original paper, the author describe that the last fine tuning step is done by original darknet, not ultralytics's yolov3. So I am trying to run using darknet.
Then how to do fine-tuning with the .pt file?
I have tried to convert using convert_pt_weights.py, it fails to convert because of decoding error like below.
command:
python convert_pt_weights.py weights/last.pt ./pruned.weights
result:
UnicodeDecodeError: 'utf-8' codec can't decode byte 0x80 in position 0: invalid start byte
Please let me know a way to do last step.
Thanks in advance.
The text was updated successfully, but these errors were encountered: