def Split(self, db: ir_database.Database) -> List[np.array]:
"""Split the database."""
poj104 = super(TrainValTestSplitter, self).Split(db)
# Get the IDs of non-POJ-104 IRs.
with db.Session() as session:
total_count = (
session.query(sql.func.count(ir_database.IntermediateRepresentation.id))
.filter(
ir_database.IntermediateRepresentation.compilation_succeeded == True,
~ir_database.IntermediateRepresentation.source.like("poj-104:%"),
)
.scalar()
)
# Scale the train/val/test ratio to the total IR count.
train_val_test_counts = np.floor(self.ratios * total_count).astype(
np.int32
)
# Round up if there were missing values.
while train_val_test_counts.sum() < total_count:
train_val_test_counts[random.randint(0, 2)] += 1
assert total_count == train_val_test_counts.sum()
app.Log(
1,
"Splitting %s IRs into splits: %s train, %s val, %s test",
humanize.Commas(total_count + sum(len(s) for s in poj104)),
humanize.Commas(train_val_test_counts[0] + len(poj104[0])),
humanize.Commas(train_val_test_counts[1] + len(poj104[1])),
humanize.Commas(train_val_test_counts[2] + len(poj104[2])),
)
ir_ids = [
row.id
for row in session.query(ir_database.IntermediateRepresentation.id)
.filter(
ir_database.IntermediateRepresentation.compilation_succeeded == True,
~ir_database.IntermediateRepresentation.source.like("poj-104:%"),
)
.order_by(db.Random())
]
if not ir_ids:
raise ValueError("No results")
return [
np.concatenate((poj104[0], ir_ids[: train_val_test_counts[0]])),
np.concatenate(
(
poj104[1],
ir_ids[train_val_test_counts[0] : sum(train_val_test_counts[:2])],
)
),
np.concatenate((poj104[2], ir_ids[sum(train_val_test_counts[:2]) :])),
]
def proto_db(
request, ir_db: ir_database.Database
) -> unlabelled_graph_database.Database:
"""A test fixture which yields a graph database with random graph tuples."""
with ir_db.Session() as session:
ir_ids = [
row.id for row in session.query(ir_database.IntermediateRepresentation.id)
]
with testing_databases.DatabaseContext(
unlabelled_graph_database.Database, request.param
) as db:
with db.Session(commit=True) as session:
session.add_all(
[
unlabelled_graph_database.ProgramGraph.Create(
proto=random_programl_generator.CreateRandomProto(
graph_y_dimensionality=2
),
ir_id=ir_id,
)
for ir_id in ir_ids
]
)
yield db
def AnnotateGraphMetas(
ir_db: ir_database.Database,
proto_db: unlabelled_graph_database.Database,
df: pd.DataFrame,
ctx: progress.ProgressContext = progress.NullContext,
) -> Iterable[graph_tuple_database.GraphTuple]:
"""Add features and labels to graph metas in database."""
with ir_db.Session() as ir_session, proto_db.Session() as proto_session:
for _, row in df.iterrows():
relpath = row["relpath"]
with ctx.Profile(
2,
f"Processed graph {row['relpath']}:{row['data:dataset_name']}"
):
# Select the corresponding IR.
ir_id = (ir_session.query(
ir_database.IntermediateRepresentation.id).filter(
ir_database.IntermediateRepresentation.source ==
"pact17_opencl_devmap",
ir_database.IntermediateRepresentation.relpath ==
relpath,
).scalar())
# Check that we have an exact 1:1 mapping from the opencl devmap dataset
# to IR.
if ir_id is None:
raise ValueError(f"Expected one IR with relpath {relpath}")
# Load the program graph.
proto_row = (proto_session.query(
unlabelled_graph_database.ProgramGraph).filter(
unlabelled_graph_database.ProgramGraph.ir_id ==
ir_id).options(
sql.orm.joinedload(unlabelled_graph_database.
ProgramGraph.data)).scalar())
if proto_row is None:
raise ValueError(
f"Expected one proto for relpath {relpath} with ID {ir_id}"
)
proto: programl_pb2.ProgramGraph = proto_row.proto
# Add the null "selector vector" value.
for node in proto.node:
node.x.append(0)
# Add the graph-level features.
proto.x[:] = [row["wgsize"], row["transfer"]]
# Add 'y' graph feature as target.
proto.y[:] = row["y"].tolist()
# Create the graph tuple. Note the jumping through hoops with converting
# proto -> nx -> graph_tuple, because there is currently no direct
# proto -> graph_tuple conversion.
graph_tuple = graph_tuple_database.GraphTuple.CreateFromGraphTuple(
graph_tuple=graph_tuples.GraphTuple.CreateFromNetworkX(
programl.ProgramGraphToNetworkX(proto)),
ir_id=ir_id,
)
yield graph_tuple
def db_with_empty_ir(request,
db: ir_database.Database) -> ir_database.Database:
empty_ir_count = request.param
with db.Session(commit=True) as session:
session.add_all([
ir_database.IntermediateRepresentation.CreateEmpty(
source="foo",
relpath=str(i),
source_language=ir_database.SourceLanguage.C,
type=ir_database.IrType.LLVM_6_0,
cflags="",
) for i in range(empty_ir_count)
])
return db
def Split(self, db: ir_database.Database) -> List[np.array]:
"""Get the bytecode IDs for the POJ-104 app classification experiment."""
def GetBytecodeIds(session, filter_cb) -> np.array:
"""Return the IDs for the given filtered query."""
ids = np.array(
[
row.id
for row in (
session.query(ir_database.IntermediateRepresentation.id).filter(
ir_database.IntermediateRepresentation.compilation_succeeded
== True,
filter_cb(),
)
)
],
dtype=np.int32,
)
if not ids.size:
raise ValueError("No results")
return ids
with db.Session() as session:
return [
GetBytecodeIds(
session,
lambda: (
ir_database.IntermediateRepresentation.source == "poj-104:train"
),
),
GetBytecodeIds(
session,
lambda: (
ir_database.IntermediateRepresentation.source == "poj-104:val"
),
),
GetBytecodeIds(
session,
lambda: (
ir_database.IntermediateRepresentation.source == "poj-104:test"
),
),
]
def PopulateBytecodeTable(self,
db: ir_database.Database,
commit_every: int = 1000):
programs_df = self.programs_df.reset_index()
bar = progressbar.ProgressBar()
bar.max_value = len(programs_df)
# Process each row of the table in parallel.
pool = multiprocessing.Pool()
with db.Session(commit=True) as s:
for i, proto in enumerate(
pool.imap_unordered(ProcessOpenClProgramDfBytecode,
[d
for _, d in programs_df.iterrows()])):
bar.update(i)
s.add(
bytecode_database.LlvmBytecode(
**bytecode_database.LlvmBytecode.FromProto(proto)))
if not (i % commit_every):
s.commit()
def populated_ir_db(ir_db: ir_database.Database,
opencl_relpaths: Set[str]) -> ir_database.Database:
"""A test fixture which yields an IR database with 256 OpenCL entries."""
rows = []
# Create random rows using OpenCL relpaths.
for i, relpath in enumerate(opencl_relpaths):
ir = ir_database.IntermediateRepresentation.CreateFromText(
source="pact17_opencl_devmap",
relpath=relpath,
source_language=ir_database.SourceLanguage.OPENCL,
type=ir_database.IrType.LLVM_6_0,
cflags="",
text=CreateRandomString(),
)
ir.id = i + 1
rows.append(ir)
with ir_db.Session(commit=True) as session:
session.add_all(rows)
return ir_db
def Split(self, db: ir_database.Database) -> List[np.array]:
"""Split the database."""
with db.Session() as session:
all_ids = np.array(
[
row.id
for row in session.query(
ir_database.IntermediateRepresentation.id
).filter(
ir_database.IntermediateRepresentation.compilation_succeeded
== True,
ir_database.IntermediateRepresentation.source
== "pact17_opencl_devmap",
)
],
dtype=np.int32,
)
if not all_ids.size:
raise ValueError("No results")
kfold = sklearn.model_selection.KFold(self.k).split(all_ids)
return [all_ids[test] for (train, test) in kfold]