How to use transfer learning when changing number of input channels?

[pourmand1376]

Search before asking

• I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

Hi, I have trained a full YOLOv5 model with grayscale images (1 channel). Now I want to use that model and train another model which gets 5 images in a row (5 channels) and predicts the output. This is easy with yolov5 ch parameters.

However, I want to use the learned features from 1 channel version. How can I achieve that? (Actually this is related to a previous issue discussed here).

Additional

No response

[glenn-jocher]

@pourmand1376 transfer learning is easy, you can start from any pretrained model even if it doesn't match the architecture 100%. What you do is define the --cfg of the model you want to train and the pretrained --weights you want to start from. All layers with matching names and shapes will transfer, the rest will be initialized randomly. You'll see a report in the console telling you how many layers were trasnferred, i.e. 'successfully transferred 190/272 layers`:

python train.py --cfg yolov5s.yaml --weights pretrained_model.pt

[pourmand1376]

That's pretty amazing! I didn't know YOLOv5 does this automatically.

[wangfurong123]

@pourmand1376 Hello, how did you implement the training of grayscale images (one channel) based on the yolov5 model? Can you provide code? This question is bothering me, I hope you can answer it, thank you.

[glenn-jocher]

@wangfurong123 greyscale images are handled automatically, just train normally.