Preprocessing

[cclough]

For the preprocessing for ResNet models (switching RGB images to BGR):

• Just to confirm, the pre-trained resnet models are trained on BGR?
• Do the pretrained resnet models expect images to be zero mean & unit variance also, or not?

Another question: when fine tuning, how important is the 2 epoch freeze first? important/not important?

Thank you

[qubvel]

1. yes, resnet models trained with images in BGR
2. no, images should be in range 0-255 (models have batchnorm as the first layer)
3. it is not very important to freeze, you can start train all model layers from the first epoch.

[nklein23]

I obtain better results when I provide normalized input data, strange..