Clarity on range normalization of inputs

[hari-aindra]

Normalization range for input to the network is not clearly defined.

The tutorial says it should be normalized to the range 0-1 here features['image'] = tf.cast(features['image'], tf.float32) / 255.0

The training code on the other hand normalizes it to -1 to +1 here im = (im - 127.5) / 127.5

Among these, which is the correct one?

[akolesnikoff]

The correct scaling depends on where you download the model from. Originally, all BiT models were trained with the [-1, 1] scaling and expect such scaling as input.

However, there is one exception: if you download BiT model from https://tfhub.dev, then the expected scaling is [0, 1], as tensorflow hub enforces such convention. If I am not mistaken, in this repository, TF2 colab is the only example, that uses hub models and, thus, expects [0, 1] scaling.

[hari-aindra]

That solves the issue. Thanks

[hari-aindra]

Do the tensor-hub model with 0-1 scaling have same performance(Accuracy) as of the original BIT models ?

[lucasb-eyer]

Yes, they are exactly the same models and just do the conversion internally.