Question about image preprocess?

[ztf-ucas]

@fchollet @joeyhng
When I use the pretrained vgg16 model(vgg16_weights_tf_dim_ordering_tf_kernels), Which way should I according to preprocess input? Tf(range is -1,1) or caffe(convert to bgr and zero-center)?
Keras2.1.2 tf backend
The pretrained model are downloaded from weights_path_no_top in Keras code.
Thank you very much!

[ztf-ucas]

I want to know which way do you train the model? Tf or caffe? And If my processing is inconsistent with the pre-trained model, will it affect performance?