Invalid compilation caused by categorical Aggregate with default=None

[langosco]

Raising a separate issue to track a problem mentioned in #11.

Programs containing multiple categorical aggregates can sometimes compile invalidly (i.e. the output of the compiled model doesn't match the output of the rasp program).

This probably happens because the value None is just propagated forward unchanged in the rasp program. In the compiled model, Nones are converted to 0s, on which further operations can be performed (if I'm not mistaken).

Example:

from tracr.rasp import rasp
from tracr.compiler import compiling, validating

sel = rasp.Select(rasp.indices, rasp.tokens, rasp.Comparison.EQ)
sop = rasp.Aggregate(sel, rasp.indices)
program = rasp.Aggregate(sel, sop)


model = compiling.compile_rasp_to_model(program, vocab={1,2,3,4}, max_seq_len=5, compiler_bos="BOS")
compiled_output = model.apply(["BOS", 1, 2, 3, 4]).decoded
rasp_output = program([1, 2, 3, 4])


# The output of the compiled model does not match the output of the RASP program:
print(rasp_output)  # [2.0, 3.0, None, None]
print(compiled_output) # ['BOS', 2, 3, 0, 1]

# The validator doesn't catch the error:
print(validating.validate(program, [1, 2, 3, 4])) # []

[langosco]

Here's a shorter example (a single Select-Aggregate) of a categorical aggregate compiling incorrectly:

example_program = rasp.categorical(rasp.Aggregate(
    rasp.Select(rasp.tokens, rasp.tokens, predicate=rasp.Comparison.FALSE),
    rasp.tokens,
))

compiled = compiling.compile_rasp_to_model(example_program, vocab=set(range(5)), max_seq_len=5)

print(example_program([1,2,3]))  # [None, None, None]
print(compiled.apply(["compiler_bos", 1,2,3]).decoded)  # ["compiler_bos", 1, 1, 1]

For general inputs, compiled will return the minimum of the input sequence (1 in the example above). IMO the desired behavior here would be to consistently return 0.