def train(self, output_dir):
with tempfile.TemporaryDirectory() as tmpdir:
loss_fn = nn.Sequential(
OrderedDict([
("loss",
Apply(
module=Loss(),
in_keys=[
"rule_probs",
"token_probs",
"reference_probs",
"ground_truth_actions",
],
out_key="action_sequence_loss",
)),
("pick",
mlprogram.nn.Function(Pick("action_sequence_loss")))
]))
collate = Collate(word_nl_query=CollateOptions(True, 0, -1),
nl_query_features=CollateOptions(True, 0, -1),
reference_features=CollateOptions(True, 0, -1),
actions=CollateOptions(True, 0, -1),
previous_actions=CollateOptions(True, 0, -1),
previous_action_rules=CollateOptions(
True, 0, -1),
history=CollateOptions(False, 1, 0),
hidden_state=CollateOptions(False, 0, 0),
state=CollateOptions(False, 0, 0),
ground_truth_actions=CollateOptions(True, 0,
-1)).collate
qencoder, aencoder = \
self.prepare_encoder(train_dataset, Parser())
transform = Map(self.transform_cls(qencoder, aencoder, Parser()))
model = self.prepare_model(qencoder, aencoder)
optimizer = self.prepare_optimizer(model)
train_supervised(tmpdir,
output_dir,
train_dataset,
model,
optimizer,
loss_fn,
EvaluateSynthesizer(test_dataset,
self.prepare_synthesizer(
model, qencoder,
aencoder),
{"accuracy": Accuracy()},
top_n=[5]),
"accuracy@5",
lambda x: collate(transform(x)),
1,
Epoch(100),
evaluation_interval=Epoch(100),
snapshot_interval=Epoch(100),
threshold=1.0)
return qencoder, aencoder
def prepare_synthesizer(self, model, qencoder, aencoder):
transform_input = Compose(
OrderedDict([("extract_reference",
Apply(module=mlprogram.nn.Function(tokenize),
in_keys=[["text_query", "str"]],
out_key="reference")),
("encode_query",
Apply(module=EncodeWordQuery(qencoder),
in_keys=["reference"],
out_key="word_nl_query"))]))
transform_action_sequence = Compose(
OrderedDict([("add_previous_action",
Apply(
module=AddPreviousActions(aencoder,
n_dependent=1),
in_keys=["action_sequence", "reference"],
constants={"train": False},
out_key="previous_actions",
)),
("add_action",
Apply(
module=AddActions(aencoder, n_dependent=1),
in_keys=["action_sequence", "reference"],
constants={"train": False},
out_key="actions",
)), ("add_state", AddState("state")),
("add_hidden_state", AddState("hidden_state")),
("add_history", AddState("history"))]))
collate = Collate(word_nl_query=CollateOptions(True, 0, -1),
nl_query_features=CollateOptions(True, 0, -1),
reference_features=CollateOptions(True, 0, -1),
actions=CollateOptions(True, 0, -1),
previous_actions=CollateOptions(True, 0, -1),
previous_action_rules=CollateOptions(True, 0, -1),
history=CollateOptions(False, 1, 0),
hidden_state=CollateOptions(False, 0, 0),
state=CollateOptions(False, 0, 0),
ground_truth_actions=CollateOptions(True, 0, -1))
return BeamSearch(
5, 20,
ActionSequenceSampler(aencoder, is_subtype, transform_input,
transform_action_sequence, collate, model))
def test_rescore(self):
def transform(state: str) -> Environment:
return Environment({"x": torch.tensor([int(state)])})
collate = Collate(x=CollateOptions(False, 0, 0))
sampler = SamplerWithValueNetwork(MockSampler(), transform, collate,
MockValueNetwork())
zero = SamplerState(0, sampler.initialize(0))
samples = list(sampler.batch_k_samples([zero], [3]))
assert [
DuplicatedSamplerState(SamplerState(0, "00"), 1),
DuplicatedSamplerState(SamplerState(1, "01"), 1),
DuplicatedSamplerState(SamplerState(2, "02"), 1)
] == samples
def test_REINFORCESynthesizer():
synthesizer = MockSynthesizer()
synthesizer.model.model.weight.data[:] = 10.0
optimizer = torch.optim.SGD(synthesizer.model.parameters(), 0.1)
synthesizer = REINFORCESynthesizer(
synthesizer=synthesizer,
model=synthesizer.model,
optimizer=optimizer,
loss_fn=Loss(),
reward=Reward(),
collate=Collate(x=CollateOptions(False, 0, 0)).collate,
n_rollout=1,
device=torch.device("cpu"),
baseline_momentum=0.9,
max_try_num=1,
)
input = Environment({"x": torch.tensor(1.0)})
for i, x in enumerate(synthesizer(input)):
assert i < 100
if x.output == 1:
break
def prepare_synthesizer(self, model, qencoder, cencoder, aencoder):
transform_input = Compose(
OrderedDict([("extract_reference",
Apply(module=mlprogram.nn.Function(tokenize),
in_keys=[["text_query", "str"]],
out_key="reference")),
("encode_word_query",
Apply(module=EncodeWordQuery(qencoder),
in_keys=["reference"],
out_key="word_nl_query")),
("encode_char",
Apply(module=EncodeCharacterQuery(cencoder, 10),
in_keys=["reference"],
out_key="char_nl_query"))]))
transform_action_sequence = Compose(
OrderedDict([("add_previous_action",
Apply(
module=AddPreviousActions(aencoder),
in_keys=["action_sequence", "reference"],
constants={"train": False},
out_key="previous_actions",
)),
("add_previous_action_rule",
Apply(
module=AddPreviousActionRules(
aencoder,
4,
),
in_keys=["action_sequence", "reference"],
constants={"train": False},
out_key="previous_action_rules",
)),
("add_tree",
Apply(
module=AddActionSequenceAsTree(aencoder),
in_keys=["action_sequence", "reference"],
constants={"train": False},
out_key=["adjacency_matrix", "depthes"],
)),
("add_query",
Apply(
module=AddQueryForTreeGenDecoder(aencoder, 4),
in_keys=["action_sequence", "reference"],
constants={"train": False},
out_key="action_queries",
))]))
collate = Collate(word_nl_query=CollateOptions(True, 0, -1),
char_nl_query=CollateOptions(True, 0, -1),
nl_query_features=CollateOptions(True, 0, -1),
reference_features=CollateOptions(True, 0, -1),
previous_actions=CollateOptions(True, 0, -1),
previous_action_rules=CollateOptions(True, 0, -1),
depthes=CollateOptions(False, 1, 0),
adjacency_matrix=CollateOptions(False, 0, 0),
action_queries=CollateOptions(True, 0, -1),
ground_truth_actions=CollateOptions(True, 0, -1))
return BeamSearch(
5, 20,
ActionSequenceSampler(aencoder, is_subtype, transform_input,
transform_action_sequence, collate, model))
def reinforce(self, train_dataset, encoder, output_dir):
with tempfile.TemporaryDirectory() as tmpdir:
interpreter = self.interpreter()
collate = Collate(
test_case_tensor=CollateOptions(False, 0, 0),
variables_tensor=CollateOptions(True, 0, 0),
previous_actions=CollateOptions(True, 0, -1),
hidden_state=CollateOptions(False, 0, 0),
state=CollateOptions(False, 0, 0),
ground_truth_actions=CollateOptions(True, 0, -1),
reward=CollateOptions(False, 0, 0)
)
collate_fn = Sequence(OrderedDict([
("to_episode", Map(self.to_episode(encoder,
interpreter))),
("flatten", Flatten()),
("transform", Map(self.transform(
encoder, interpreter, Parser()))),
("collate", collate.collate)
]))
model = self.prepare_model(encoder)
optimizer = self.prepare_optimizer(model)
train_REINFORCE(
output_dir, tmpdir, output_dir,
train_dataset,
self.prepare_synthesizer(model, encoder, interpreter),
model, optimizer,
torch.nn.Sequential(OrderedDict([
("policy",
torch.nn.Sequential(OrderedDict([
("loss",
Apply(
module=mlprogram.nn.action_sequence.Loss(
reduction="none"
),
in_keys=[
"rule_probs",
"token_probs",
"reference_probs",
"ground_truth_actions",
],
out_key="action_sequence_loss",
)),
("weight_by_reward",
Apply(
[("reward", "lhs"),
("action_sequence_loss", "rhs")],
"action_sequence_loss",
mlprogram.nn.Function(Mul())))
]))),
("value",
torch.nn.Sequential(OrderedDict([
("reshape_reward",
Apply(
[("reward", "x")],
"value_loss_target",
Reshape([-1, 1]))),
("BCE",
Apply(
[("value", "input"),
("value_loss_target", "target")],
"value_loss",
torch.nn.BCELoss(reduction='sum'))),
("reweight",
Apply(
[("value_loss", "lhs")],
"value_loss",
mlprogram.nn.Function(Mul()),
constants={"rhs": 1e-2})),
]))),
("aggregate",
Apply(
["action_sequence_loss", "value_loss"],
"loss",
AggregatedLoss())),
("normalize",
Apply(
[("loss", "lhs")],
"loss",
mlprogram.nn.Function(Div()),
constants={"rhs": 1})),
("pick",
mlprogram.nn.Function(
Pick("loss")))
])),
EvaluateSynthesizer(
train_dataset,
self.prepare_synthesizer(model, encoder, interpreter,
rollout=False),
{}, top_n=[]),
"generation_rate",
metrics.use_environment(
metric=metrics.TestCaseResult(
interpreter=interpreter,
metric=metrics.use_environment(
metric=metrics.Iou(),
in_keys=["actual", "expected"],
value_key="actual",
)
),
in_keys=["test_cases", "actual"],
value_key="actual",
transform=Threshold(threshold=0.9, dtype="float"),
),
collate_fn,
1, 1,
Epoch(10), evaluation_interval=Epoch(10),
snapshot_interval=Epoch(10),
use_pretrained_model=True,
use_pretrained_optimizer=True,
threshold=1.0)
def pretrain(self, output_dir):
dataset = Dataset(4, 1, 2, 1, 45, seed=0)
"""
"""
train_dataset = ListDataset([
Environment(
{"ground_truth": Circle(1)},
set(["ground_truth"]),
),
Environment(
{"ground_truth": Rectangle(1, 2)},
set(["ground_truth"]),
),
Environment(
{"ground_truth": Rectangle(1, 1)},
set(["ground_truth"]),
),
Environment(
{"ground_truth": Rotation(45, Rectangle(1, 1))},
set(["ground_truth"]),
),
Environment(
{"ground_truth": Translation(1, 1, Rectangle(1, 1))},
set(["ground_truth"]),
),
Environment(
{"ground_truth": Difference(Circle(1), Circle(1))},
set(["ground_truth"]),
),
Environment(
{"ground_truth": Union(Rectangle(1, 2), Circle(1))},
set(["ground_truth"]),
),
Environment(
{"ground_truth": Difference(Rectangle(1, 1), Circle(1))},
set(["ground_truth"]),
),
])
with tempfile.TemporaryDirectory() as tmpdir:
interpreter = self.interpreter()
train_dataset = data_transform(
train_dataset,
Apply(
module=AddTestCases(interpreter),
in_keys=["ground_truth"],
out_key="test_cases",
is_out_supervision=False,
))
encoder = self.prepare_encoder(dataset, Parser())
collate = Collate(
test_case_tensor=CollateOptions(False, 0, 0),
variables_tensor=CollateOptions(True, 0, 0),
previous_actions=CollateOptions(True, 0, -1),
hidden_state=CollateOptions(False, 0, 0),
state=CollateOptions(False, 0, 0),
ground_truth_actions=CollateOptions(True, 0, -1)
)
collate_fn = Sequence(OrderedDict([
("to_episode", Map(self.to_episode(encoder,
interpreter))),
("flatten", Flatten()),
("transform", Map(self.transform(
encoder, interpreter, Parser()))),
("collate", collate.collate)
]))
model = self.prepare_model(encoder)
optimizer = self.prepare_optimizer(model)
train_supervised(
tmpdir, output_dir,
train_dataset, model, optimizer,
torch.nn.Sequential(OrderedDict([
("loss",
Apply(
module=Loss(
reduction="sum",
),
in_keys=[
"rule_probs",
"token_probs",
"reference_probs",
"ground_truth_actions",
],
out_key="action_sequence_loss",
)),
("normalize", # divided by batch_size
Apply(
[("action_sequence_loss", "lhs")],
"loss",
mlprogram.nn.Function(Div()),
constants={"rhs": 1})),
("pick",
mlprogram.nn.Function(
Pick("loss")))
])),
None, "score",
collate_fn,
1, Epoch(100), evaluation_interval=Epoch(10),
snapshot_interval=Epoch(100)
)
return encoder, train_dataset
def prepare_synthesizer(self, model, encoder, interpreter, rollout=True):
collate = Collate(
test_case_tensor=CollateOptions(False, 0, 0),
input_feature=CollateOptions(False, 0, 0),
test_case_feature=CollateOptions(False, 0, 0),
reference_features=CollateOptions(True, 0, 0),
variables_tensor=CollateOptions(True, 0, 0),
previous_actions=CollateOptions(True, 0, -1),
hidden_state=CollateOptions(False, 0, 0),
state=CollateOptions(False, 0, 0),
ground_truth_actions=CollateOptions(True, 0, -1)
)
subsampler = ActionSequenceSampler(
encoder, IsSubtype(),
Sequence(OrderedDict([
("tinput",
Apply(
module=TransformInputs(),
in_keys=["test_cases"],
out_key="test_case_tensor",
)),
("tvariable",
Apply(
module=TransformVariables(),
in_keys=["variables", "test_case_tensor"],
out_key="variables_tensor"
)),
])),
Compose(OrderedDict([
("add_previous_actions",
Apply(
module=AddPreviousActions(encoder, n_dependent=1),
in_keys=["action_sequence", "reference"],
out_key="previous_actions",
constants={"train": False},
)),
("add_state", AddState("state")),
("add_hidden_state", AddState("hidden_state"))
])),
collate, model,
rng=np.random.RandomState(0))
subsampler = mlprogram.samplers.transform(
subsampler,
Parser().unparse
)
subsynthesizer = SMC(
5, 1,
subsampler,
max_try_num=1,
to_key=Pick("action_sequence"),
rng=np.random.RandomState(0)
)
sampler = SequentialProgramSampler(
subsynthesizer,
Apply(
module=TransformInputs(),
in_keys=["test_cases"],
out_key="test_case_tensor",
),
collate,
model.encode_input,
interpreter=interpreter,
expander=Expander(),
rng=np.random.RandomState(0))
if rollout:
sampler = FilteredSampler(
sampler,
metrics.use_environment(
metric=metrics.TestCaseResult(
interpreter,
metric=metrics.use_environment(
metric=metrics.Iou(),
in_keys=["actual", "expected"],
value_key="actual",
)
),
in_keys=["test_cases", "actual"],
value_key="actual"
),
1.0
)
return SMC(3, 1, sampler, rng=np.random.RandomState(0),
to_key=Pick("interpreter_state"), max_try_num=1)
else:
sampler = SamplerWithValueNetwork(
sampler,
Sequence(OrderedDict([
("tinput",
Apply(
module=TransformInputs(),
in_keys=["test_cases"],
out_key="test_case_tensor",
)),
("tvariable",
Apply(
module=TransformVariables(),
in_keys=["variables", "test_case_tensor"],
out_key="variables_tensor"
)),
])),
collate,
torch.nn.Sequential(OrderedDict([
("encoder", model.encoder),
("value", model.value),
("pick",
mlprogram.nn.Function(
Pick("value")))
])))
synthesizer = SynthesizerWithTimeout(
SMC(3, 1, sampler, rng=np.random.RandomState(0),
to_key=Pick("interpreter_state"),
max_try_num=1),
1
)
return FilteredSynthesizer(
synthesizer,
metrics.use_environment(
metric=metrics.TestCaseResult(
interpreter,
metric=metrics.use_environment(
metric=metrics.Iou(),
in_keys=["actual", "expected"],
value_key="actual",
)
),
in_keys=["test_cases", "actual"],
value_key="actual"
),
1.0
)
if arg0 == arg1:
return True
if arg0 == "Ysub" and arg1 == "Y":
return True
return False
def create_encoder():
return ActionSequenceEncoder(Samples(
[Root2X, Root2Y, X2Y_list, Ysub2Str],
[R, X, Y, Ysub, Y_list, Str],
[("Str", "x"),
("Int", "1")]), 0)
collate = Collate(input=CollateOptions(False, 0, -1),
length=CollateOptions(False, 0, -1))
def create_transform_input(reference: List[Token[str, str]]):
def transform_input(env):
env["reference"] = reference
env["input"] = torch.zeros((1,))
return env
return transform_input
def transform_action_sequence(kwargs):
kwargs["length"] = \
torch.tensor(len(kwargs["action_sequence"].action_sequence))
return kwargs
def __call__(self, input, n_required_output=None):
n_required_output = n_required_output or 1
for _ in range(n_required_output):
yield Result(input["value"], 0, True, 1)
@pytest.fixture
def synthesizer(model):
return MockSynthesizer(model)
def reward(sample, output):
return sample["value"] == output
collate = Collate(output=CollateOptions(False, 0, 0),
value=CollateOptions(False, 0, 0),
reward=CollateOptions(False, 0, 0),
ground_truth=CollateOptions(False, 0, 0))
class DummyModel(nn.Module):
def __init__(self):
super().__init__()
self.m = nn.Linear(1, 1)
def forward(self, kwargs):
kwargs["value"] = self.m(kwargs["value"].float())
return kwargs