fix #104 Add check for dot symbol and warn user

[vandrw]

When using the results of the SELFIES library for an ML model, I encountered some issues while preparing the data for training. If one uses the get_alphabet_from_selfies to generate a list of tokens, and then encodes each input according to it using selfies_to_encoding, it is possible to get a ValueKey error for the dot symbol. More details in #104.

This can be a bit unintuitive, as the key error can result from using the vocabulary generated from the same dataset that is fed to the encoding function.

With this pull request, I suggest adding a more descriptive error for this case, which could help users that are not very knowledgeable in chemistry understand what can be done.

[MarioKrenn6240]

One question on the PR (which might change variable types thus introduce unexpected behaviours): Why do you recast char_list as a list? char_list = list(split_selfies(selfies))

[vandrw]

As it currently is, char_list is a generator, which will be consumed when we search for the "." character. If the type is important, we could perform the line below in a try-except statement. This should also improve speed for long strings, as we are not evaluating twice. I will draft another commit shortly with this new version.

selfies/selfies/utils/encoding_utils.py

Line 51 in 120b776

integer_encoded = [vocab_stoi[char] for char in char_list]

[alstonlo]

Thanks for this PR! Sorry, I would prefer to merge the first commit since it reads more clearly!

char_list = list(split_selfies(selfies))

# Check if SELFIES string contains unconnected molecules
if "." in list(char_list) and not "." in vocab_stoi:
    raise ValueError(
        "The SELFIES string contains two unconnected molecules "
        "(given by the '.' character), but vocab_stoi does not "
        "contain the '.' key. Please add it or separate the molecules."
    )

integer_encoded = [vocab_stoi[char] for char in char_list]

I agree that the first cast to char_list is necessary, but I think there is a redundant one in the if-statement. Alternatively, maybe a one-pass approach as follows could work:

integer_encoded = []

for char in split_selfies(selfies):
    if (char == ".") and ("." not in vocab_stoi):
        raise ValueError("...")
    integer_encoded.append(vocab_stoi[char])

[vandrw]

That sounds better! I've added the suggested one-pass approach in a new commit to make it easier to merge. It passes all the tests, so feel free to merge it or let me know if I can help with anything else!

[MarioKrenn6240]

Thank you!