Can we use the weights of a model trained on RGB i

问题:

actually I am trying to use transfer learning from pre trained inception model on Image Net to my dataset which is in gray scale , so i just want to know can i use the same weights or some changes should be made first before using the weights.

Thanks in advance.

回答1:

There are a few options available:

1. Convert grayscale images to colour images using standard OpenCV functions.
2. Add a convolutional layer in front that has depth 3 as output. Proceed as per normal.
3. Select one set of weights from the first layer of your existing network.
4. Average the weights from the first layer of your existing network.

Each option has different trade-offs respectively:

1. This is probably what would work well but you pay the price of extra computation, which may not be a lot in this case.
2. You might essentially be learning filters that try to map to RGB, so this might be similar to 1.
3. This is a hack, and may not work like how you would expect.
4. Same as 3.