'preprocess_input' defaults to 'caffe' rather than 'tf'

[ialhashim]

I was very confused when my 'cat' pictures where predicted as 'lampshade'. Shouldn't Tensorflow be the default?

keras/keras/applications/imagenet_utils.py

Line 151 in b0f1bb9

def preprocess_input(x, data_format=None, mode='caffe'):

[fchouteau]

Hi,

The preprocess_input function defaults to caffe because it is imported without modifications in keras.applications.vgg16 and keras.applications.resnet50 where those models were trained using caffe2 preprocessing. I suppose this was put in imagenet_utils to have sane defaults for those classic networks. (example: https://github.com/keras-team/keras/blob/master/keras/applications/vgg16.py#L29)

For more recent models such as Xception and MobileNet (where the training was under tf preprocessing), the preprocess_input function is actually redefined see: https://github.com/keras-team/keras/blob/master/keras/applications/xception.py#L267 for tf and https://github.com/keras-team/keras/blob/master/keras/applications/densenet.py#L325 for torch

[ialhashim]

I see, I was testing Densenet so actually I should be using 'torch' or better yet import the corresponding preprocessing function and the model definition from 'densenet' (or other trained models).

[swghosh]

I had the same question. Thanks, @ialhashim @fchouteau for the clarification.

Does that mean the recommended way of using preprocessing_input from models like vgg16 and resnet50 be mode='caffe' (the default one) itself for getting correct predictions on a Tensorflow backend too?