There is possibly a different answer depending on your usecase, so please answer the following questions: 

The reason for asking these questions is: 

If there is little overlap between vgg-F data and yours, then it is unlikely that the vgg-F network will help with your problem out of the box. In particular, if there is no relationship between the labels of the two domains, it's hard to see that the final layers of the network will help with your problem as these should be tuned specifically for representing vgg-F's targets rather than yours. If this is the case, then it might be reasonable to use the output from layers closer to the middle than to the end as these may yield a more general representation of the data that may help your problem. 

This is quite speculative right now :-) If the tasks of yours and vgg-F match, then there is no reason not to use the full network.

Either way, it is probably worth fine tuning it to your data if possible. 

Thank you Niall for your answer,

vgg-F network was trained on ImageNet dataset, and my data are MRI images which is completely different. But I got very good classification accuracies using the features extracted from the 7th layer without tuning. (which is a known method in the literature)

yes I will fine tune the network. But before tuning, my question at this stage is normally is there a convention or reasonable justification about wether to use features extracted after activating the 7th layer with ReLU function or before?

This is probably something you have to answer empirically with experimentation. Try cutting it at different places and see what happens.

There is no straightforward rule to decide whether you will get a better representation with or without ReLu from 7th layer features. Applying ReLu may increase the performance but that you need to investigate experimentally..