oo-method-test-split.py

#!/usr/bin/python

import datasets

import itertools

import os

import pyarrow as pa
import pyarrow.parquet as pq

BASE_DATASET = "ejschwartz/oo-method-test"

def setexe(r):
    r['Dirname'], r['Exename'] = os.path.split(r['Binary'])
    return r

class OOMethodTestDataset(datasets.ArrowBasedBuilder):

    BUILDER_CONFIGS = [
        datasets.BuilderConfig(
            name="combined",
            version=datasets.Version("1.0.0"),
            description="All data files combined",
        ),
        datasets.BuilderConfig(
            name="byrow",
            version=datasets.Version("1.0.0"),
            description="Split by example (dumb)",
        ),
        datasets.BuilderConfig(
            name="byfuncname",
            version=datasets.Version("1.0.0"),
            description="Split by function name",
        ),
        datasets.BuilderConfig(
            name="bylibrary",
            version=datasets.Version("1.0.0"),
            description="Split so that library functions (those appearing in >1 exe) are used for training, and non-library functions are used for testing",
        ),
        datasets.BuilderConfig(
            name="bylibrarydedup",
            version=datasets.Version("1.0.0"),
            description="Split so that library functions (those appearing in >1 exe) are used for training, and non-library functions are used for testing. Only one example per function name is retained.",
        )

    ]

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

    def _info(self):
        return datasets.DatasetInfo(
            features = datasets.Features({'Binary': datasets.Value(dtype='string', id=None),
                 'Addr': datasets.Value(dtype='string'),
                 'Name': datasets.Value(dtype='string'),
                 'Type': datasets.ClassLabel(num_classes=2, names=['func', 'method']),
                 'Disassembly': datasets.Value(dtype='string'),
                 'Dirname': datasets.Value(dtype='string'),
                 'Exename': datasets.Value(dtype='string')}))

    def _split_generators(self, dl_manager):
        ds = datasets.load_dataset(BASE_DATASET)['combined']

        ds = ds.map(setexe, batched=False)

        if self.config.name == "combined":

            return [
                datasets.SplitGenerator(
                    name="combined",
                    gen_kwargs={
                        "ds": ds,
                    },
                ),
            ]
        
        elif self.config.name == "byrow":

            ds = ds.train_test_split(test_size=0.1, seed=42)
            #print(ds)

            return [
                datasets.SplitGenerator(
                    name="train",
                    gen_kwargs={
                        "ds": ds['train'],
                    },
                ),
                datasets.SplitGenerator(
                    name="test",
                    gen_kwargs={
                        "ds": ds['test'],
                    },
                ),

            ]
        
        elif self.config.name == "byfuncname":

            unique_names = ds.unique('Name')
            nameds = datasets.Dataset.from_dict({'Name': unique_names})

            name_split = nameds.train_test_split(test_size=0.1, seed=42)
            #print(name_split)

            train_name = name_split['train']['Name']
            test_name = name_split['test']['Name']

            return [
                datasets.SplitGenerator(
                    name="train",
                    gen_kwargs={
                        "ds": ds.filter(lambda r: r['Name'] in train_name),
                    },
                ),
                datasets.SplitGenerator(
                    name="test",
                    gen_kwargs={
                        "ds": ds.filter(lambda r: r['Name'] in test_name),
                    },
                ),

            ]
        
        elif self.config.name in ["bylibrary", "bylibrarydedup"]:
            # A function (name) is a library function if it appears in more than one Exename

            # this is (('func', 'oo.exe'): 123)
            testcount = set(zip(ds['Name'], ds['Exename']))

            # sorted pairs by function name
            testcount = sorted(testcount, key=lambda x: x[0])

            # group by function name
            grouped = itertools.groupby(testcount, lambda t: t[0])

            grouped = {k: [b for _,b in g] for k, g in grouped}

            library_func_names = {f for f, exes in grouped.items() if len(exes) > 1}
            library_func_names_dedup = {(f, exes[0]) for f, exes in grouped.items() if len(exes) > 1}
            nonlibrary_func_names = {f for f, exes in grouped.items() if len(exes) == 1}

            train_filter_fun = None
            if self.config.name == "bylibrary":
                train_filter_fun = lambda r: r['Name'] in library_func_names
            elif self.config.name == "bylibrarydedup":
                train_filter_fun = lambda r: (r['Name'], r['Exename']) in library_func_names_dedup
            else:
                assert False, "Invalid configuration"

            return [
                datasets.SplitGenerator(
                    name="train",
                    gen_kwargs={
                        "ds": ds.filter(train_filter_fun),
                    },
                ),
                datasets.SplitGenerator(
                    name="test",
                    gen_kwargs={
                        "ds": ds.filter(lambda r: r['Name'] in nonlibrary_func_names),
                    },
                ),

            ]

        else:
            assert False
    
    def _generate_tables(self, ds):

        # Converting to pandas is silly, but the old version of datasets doesn't
        # seem to have a way to convert to Arrow?
        for i, batch in enumerate(ds.to_pandas(batched=True)):
            yield i, pa.Table.from_pandas(batch)