Details about the selection of structures in the dp test command

[vwcruzeiro]

Hello everyone,

I would like to use the dp test command to find out the errors in the energies, forces, and virials for my full training and test sets. Let's assume that my training set is located at data/set.000 and that my test set is at data/set.001. When I execute dp test -m graph.pb -s data -n 100 it is unclear to me from which set.00? folder the 100 structures will be picked from. For this reason, I have moved my training set to training_data/set.000 and my test set to test_data/set.000; then, by ensuring that type.raw is placed at both training_data and test_data folders, when I run dp test pointing the -s flag to these different folders I should select structures only from either the training or test sets, correct?

Assuming what I explained above is correct, my question is more specifically about the -n flag. If I want, for example, to evaluate the full test set, I can set -d to test_data and -n equal to the number of structures in the test set. But, will this ensure that all structures are picked? Let me explain what I mean: in case the selection is completely random, a given structure could be picked more than once while another is never picked. Thus, if there something in the code to prevent a given structure to be selected more than once?

Thank you!

[tuoping]

You can take a look at deepmd/utils/data.py. As you can see, the test_set is not taken randomly, but taken orderly, with tag "shuffle_test" to determine wether to shuffle the order. That is to say, no structures are selected more than once.

[vwcruzeiro]

Thank you very much for your reply!