# -*- coding:utf-8 -*-
#
# Copyright (C) 2019-2021, Saarland University
# Copyright (C) 2019-2021, Maximilian Köhl <koehl@cs.uni-saarland.de>
from __future__ import annotations
import dataclasses as d
import typing as t
import asyncio
import json
import fractions
import pathlib
import struct
import subprocess
import tempfile
from .. import jani, model
from ..tools import modest
from . import generic
from .abstract import Oracle
from .dump_nn import dump_nn
if t.TYPE_CHECKING:
import torch # type: ignore
HEADER = struct.Struct("!II") # (num_features, num_actions)
DECISION = struct.Struct("!I") # (action,)
@d.dataclass(frozen=True)
class ModesOptions:
max_run_length_as_end: bool = True
additional_options: t.Sequence[str] = ()
def apply(self, command: t.List[str]) -> None:
if self.max_run_length_as_end:
command.append("--max-run-length-as-end")
command.extend(self.additional_options)
_DEFAULT_OPTIONS = ModesOptions()
def _apply_actions_observations(
command: t.List[str], actions: generic.Actions, observations: generic.Observations
) -> None:
if actions is generic.Actions.EDGE_BY_LABEL:
command.append("--select-by-label")
else:
assert actions is generic.Actions.EDGE_BY_INDEX
if observations is generic.Observations.LOCAL_AND_GLOBAL:
command.append("--observations-local-global")
elif observations is generic.Observations.OMNISCIENT:
command.append("--observations-omniscient")
else:
assert observations is generic.Observations.GLOBAL_ONLY
def _create_vector_decoder(num_features: int) -> struct.Struct:
return struct.Struct(f"!{num_features}f")
def _create_available_decoder(num_actions: int) -> struct.Struct:
return struct.Struct(f"!{num_actions}?")
@d.dataclass()
class OracleServer:
oracle: Oracle
host: str = "0.0.0.0"
port: t.Optional[int] = None
server: t.Optional[asyncio.AbstractServer] = None
async def _handle_connection(
self, reader: asyncio.StreamReader, writer: asyncio.StreamWriter
) -> None:
address = writer.get_extra_info("peername")
print(f"New connection from {address}.")
num_features, num_actions = HEADER.unpack(await reader.readexactly(HEADER.size))
print(f"Features: {num_features}, Actions: {num_actions}")
state_decoder = _create_vector_decoder(num_features)
available_decoder = _create_available_decoder(num_actions)
try:
while True:
state = state_decoder.unpack(
await reader.readexactly(state_decoder.size)
)
available = available_decoder.unpack(
await reader.readexactly(available_decoder.size)
)
decision = self.oracle(state, available)
writer.write(DECISION.pack(decision))
await writer.drain()
except asyncio.IncompleteReadError:
pass
async def start(self) -> None:
self.server = await asyncio.start_server(
self._handle_connection, self.host, self.port
)
assert self.server.sockets is not None
self.port = self.server.sockets[0].getsockname()[1]
await self.server.start_serving()
async def stop(self) -> None:
assert self.server is not None
self.server.close()
await self.server.wait_closed()
_NO_PARAMETERS: t.Dict[str, t.Any] = {}
def _load_result(output_file: pathlib.Path) -> t.Mapping[str, fractions.Fraction]:
result = json.loads(output_file.read_text(encoding="utf-8-sig"))
return {
dataset["property"]: fractions.Fraction(dataset["value"])
for dataset in result["data"]
if "data" in dataset
for dataset in dataset["data"]
if "property" in dataset
}
async def _check_oracle(
model_path: pathlib.Path,
automaton_name: str,
oracle: Oracle,
output_file: pathlib.Path,
*,
parameters: t.Mapping[str, t.Any],
toolset: modest.Toolset,
options: ModesOptions,
actions: generic.Actions,
observations: generic.Observations,
) -> None:
server = OracleServer(oracle)
await server.start()
address = f"localhost:{server.port}"
print(f"Address: {address}")
arguments: t.List[str] = [
"modes",
str(model_path),
"-O",
str(output_file),
"json",
"-R",
"oracle",
"--socket",
address,
"-A",
automaton_name,
"--threads",
"1",
]
options.apply(arguments)
for param_name, param_value in parameters.items():
arguments.extend(["-E", f"{param_name}={param_value}"])
_apply_actions_observations(arguments, actions, observations)
process = await asyncio.subprocess.create_subprocess_exec(
toolset.executable, *arguments
)
await process.wait()
async def check_oracle_async(
network: model.Network,
instance: model.Instance,
oracle: Oracle,
*,
parameters: t.Mapping[str, t.Any] = _NO_PARAMETERS,
toolset: modest.Toolset = modest.toolset,
options: ModesOptions = _DEFAULT_OPTIONS,
actions: generic.Actions = generic.Actions.EDGE_BY_INDEX,
observations: generic.Observations = generic.Observations.GLOBAL_ONLY,
) -> t.Mapping[str, fractions.Fraction]:
"""Checks an arbitrary Python function implementing a decsion agent."""
with tempfile.TemporaryDirectory() as temp_dir:
temp_path = pathlib.Path(temp_dir)
model_path = temp_path / "model.jani"
model_path.write_text(jani.dump_model(network), encoding="utf-8")
output_file = temp_path / "output.json"
assert instance.automaton.name is not None
await _check_oracle(
model_path,
instance.automaton.name,
oracle,
output_file,
parameters=parameters,
toolset=toolset,
options=options,
actions=actions,
observations=observations,
)
return _load_result(output_file)
[docs]def check_oracle(
network: model.Network,
instance: model.Instance,
oracle: Oracle,
*,
parameters: t.Mapping[str, t.Any] = _NO_PARAMETERS,
toolset: modest.Toolset = modest.toolset,
options: ModesOptions = _DEFAULT_OPTIONS,
actions: generic.Actions = generic.Actions.EDGE_BY_INDEX,
observations: generic.Observations = generic.Observations.GLOBAL_ONLY,
) -> t.Mapping[str, fractions.Fraction]:
"""
Checks an arbitrary Python function.
Arguments:
network: A JANI automaton network.
instance: An instance of an automaton in the provided network. The decision-making
agent trained on the resulting environment is assumed to act by resolving the
non-determinism in this automaton.
orcace: The Python function to check.
parameters: Allows defining values for parameters of the JANI model.
toolset: Specifies the location of the Modest Toolset.
actions: Specifies the action space for the environment.
observations: Specifies the observation space for the environment.
"""
loop = asyncio.get_event_loop()
return loop.run_until_complete(
check_oracle_async(
network,
instance,
oracle,
parameters=parameters,
toolset=toolset,
options=options,
actions=actions,
observations=observations,
)
)
[docs]def check_nn(
network: model.Network,
instance: model.Instance,
nn: torch.nn.Module,
*,
parameters: t.Mapping[str, t.Any] = _NO_PARAMETERS,
toolset: modest.Toolset = modest.toolset,
options: ModesOptions = _DEFAULT_OPTIONS,
actions: generic.Actions = generic.Actions.EDGE_BY_INDEX,
observations: generic.Observations = generic.Observations.GLOBAL_ONLY,
) -> t.Mapping[str, fractions.Fraction]:
"""
Checks a PyTorch neural network.
Arguments:
network: A JANI automaton network.
instance: An instance of an automaton in the provided network. The decision-making
agent trained on the resulting environment is assumed to act by resolving the
non-determinism in this automaton.
nn: The PyTorch neural network.
parameters: Allows defining values for parameters of the JANI model.
toolset: Specifies the location of the Modest Toolset.
actions: Specifies the action space for the environment.
observations: Specifies the observation space for the environment.
"""
with tempfile.TemporaryDirectory() as temp_dir:
temp_path = pathlib.Path(temp_dir)
model_path = temp_path / "model.jani"
model_path.write_text(jani.dump_model(network), encoding="utf-8")
nn_path = temp_path / "nn.json"
nn_path.write_text(dump_nn(nn))
output_file = temp_path / "output.json"
command: t.List[str] = [
str(toolset.executable),
"modes",
str(model_path),
"-O",
str(output_file),
"json",
"-R",
"NN",
"-NN",
str(nn_path),
"-A",
str(instance.automaton.name),
"--threads",
"1",
]
for param_name, param_value in parameters.items():
command.extend(["-E", f"{param_name}={param_value}"])
options.apply(command)
_apply_actions_observations(command, actions, observations)
subprocess.check_call(command)
return _load_result(output_file)
