Resnet Model In Keras And Tf.keras Give Different Output For The Same Image

For a given image, I'm extracting ResNet features i.e. after all conv layers and global max pooling, which gives a 2048 length vector per image. Earlier, I was using keras==2.3.1

Solution 1:

The models themselves keep updating and the weights with better results are used. Here is the releases folder where you can see releases of different weight files. The folder linked is for keras, which is most probably also used by tf.keras. The problem you are facing is probably because of the version of library rather than method of calling the model.

For example, here you can find the keras_applications code and check the weight path which points to release v0.2(linked is keras but similarly for tensorflow.keras).

The older library versions have older links, updating the libraries also updates the paths.