Image Preprocessing to finetune/predict using pre trained VGG model

[sudhashbahu]

The example at https://github.com/tflearn/tflearn/blob/master/examples/images/vgg_network_finetuning.py shows how to fine tune the pre-trained VGG16 model.

However the trained model might have applied mean subtraction and Std normalization where the mean/STD values are calculated over entire training dataset.

if we need to fine-tune the model (or use it as-is for classification) shouldn't we have to apply the exact same preprocessing to new images?

The fine tuning example above applied Normalization to the input image but this will not match that over the entire pre-training dataset.

[aymericdamien]

You are right! this was missing. I updated the example, you can have a look.

[sudhashbahu]

Thank you for the quick action.

[sudhashbahu]

Are you sure the Normalize=True should be retained in the image preloader?

And the original VGG 16 authors seems to indicate BGR images instead of RGB and they did not apply normalization, only zero mean. It all depends on how the VGG16.tflearn was trained.

[aymericdamien]

Yes, it has been reversed to RGB, so it should be working as it.
Ops, 'Normalize' is a typo, thanks for pointing that out! I fixed it.