maxframe 2.4.0rc1__cp312-cp312-win32.whl

maxframe/learn/contrib/models.py

@@ -0,0 +1,109 @@
+# Copyright 1999-2025 Alibaba Group Holding Ltd.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Type
+
+from ... import opcodes
+from ...core import ENTITY_TYPE, OutputType
+from ...core.operator import ObjectOperator, ObjectOperatorMixin
+from ...serialization.serializables import (
+    AnyField,
+    DictField,
+    FunctionField,
+    TupleField,
+)
+from ...udf import BuiltinFunction
+from ...utils import find_objects, replace_objects
+from ..core import Model, ModelData
+
+
+class ModelWithEvalData(ModelData):
+    __slots__ = ("_evals_result",)
+
+    _evals_result: dict
+
+    def __init__(self, *args, evals_result=None, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._evals_result = evals_result if evals_result is not None else dict()
+
+    def execute(self, session=None, **kw):
+        # The evals_result should be fetched when BoosterData.execute() is called.
+        result = super().execute(session=session, **kw)
+        if (
+            getattr(self.op, "has_evals_result", None)
+            and self.key == self.op.outputs[0].key
+        ):
+            self._evals_result.update(self.op.outputs[1].fetch(session=session))
+        return result
+
+
+class ModelWithEval(Model):
+    pass
+
+
+class ModelDataSource(ObjectOperator, ObjectOperatorMixin):
+    _op_type_ = opcodes.MODEL_DATA_SOURCE
+
+    data = AnyField("data")
+
+    def __call__(self, model_cls: Type[ModelWithEval]):
+        self._output_types = [OutputType.object]
+        return self.new_tileable(None, object_class=model_cls)
+
+
+class ModelApplyChunk(ObjectOperator, ObjectOperatorMixin):
+    _op_module_ = "maxframe.learn.contrib.models"
+    _op_type_ = opcodes.APPLY_CHUNK
+
+    func = FunctionField("func")
+    args = TupleField("args")
+    kwargs = DictField("kwargs")
+
+    def __init__(self, output_types=None, **kwargs):
+        if not isinstance(output_types, (tuple, list)):
+            output_types = [output_types]
+        self._output_types = list(output_types)
+        super().__init__(**kwargs)
+
+    def has_custom_code(self) -> bool:
+        return not isinstance(self.func, BuiltinFunction)
+
+    @classmethod
+    def _set_inputs(cls, op: "ModelApplyChunk", inputs):
+        super()._set_inputs(op, inputs)
+        old_inputs = find_objects(op.args, ENTITY_TYPE) + find_objects(
+            op.kwargs, ENTITY_TYPE
+        )
+        mapping = {o: n for o, n in zip(old_inputs, op._inputs[1:])}
+        op.args = replace_objects(op.args, mapping)
+        op.kwargs = replace_objects(op.kwargs, mapping)
+
+    @property
+    def output_limit(self) -> int:
+        return len(self._output_types)
+
+    def __call__(self, t, output_kws, args=None, **kwargs):
+        self.args = args or ()
+        self.kwargs = kwargs
+        inputs = (
+            [t]
+            + find_objects(self.args, ENTITY_TYPE)
+            + find_objects(self.kwargs, ENTITY_TYPE)
+        )
+        return self.new_tileables(inputs, kws=output_kws)
+
+
+def to_remote_model(model, model_cls: Type[ModelWithEval]) -> ModelWithEval:
+    op = ModelDataSource(data=model)
+    return op(model_cls)

maxframe/learn/contrib/pytorch/__init__.py

@@ -0,0 +1,16 @@
+# Copyright 1999-2025 Alibaba Group Holding Ltd.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .run_function import run_pytorch_function
+from .run_script import run_pytorch_script

maxframe/learn/contrib/pytorch/run_function.py

@@ -0,0 +1,110 @@
+# Copyright 1999-2025 Alibaba Group Holding Ltd.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import io
+from typing import Any, Callable, Dict
+
+from ....core import TILEABLE_TYPE
+from ....lib import wrapped_pickle as pickle
+from ....typing_ import SessionType
+from ....utils import find_objects, replace_objects
+from .run_script import run_pytorch_script
+
+_script_template = """
+import cloudpickle
+from maxframe.utils import replace_objects
+
+
+def main(**kwargs):
+    func = cloudpickle.loads(%(pickled_func)r)
+    nested = cloudpickle.loads(%(nested)r)
+    args, kw = replace_objects(nested, kwargs)
+    return func(*args, **kw)
+
+
+if __name__ == "__main__":
+    vars = dict()
+    for var_name in cloudpickle.loads(%(var_names)r):
+        vars[var_name] = globals()[var_name]
+    main(**vars)
+"""
+
+
+def run_pytorch_function(
+    func: Callable,
+    args: tuple = (),
+    kwargs: dict = None,
+    *,
+    n_workers: int = 1,
+    retry_when_fail: bool = False,
+    session: SessionType = None,
+    run_kwargs: Dict[str, Any] = None,
+    port: int = None,
+    execute: bool = True,
+):
+    """
+    Run PyTorch function in MaxFrame cluster.
+
+    Besides args and kwargs, the function will receive extra
+    environment variables from the caller:
+
+    * MASTER_ADDR, MASTER_PORT: the endpoint of the master node
+    * RANK: the index of the current worker
+
+    Parameters
+    ----------
+    func: Callable
+        Function or callable object to run
+    args: tuple
+        Args tuple to pass to the function
+    kwargs: dict
+    n_workers : int
+        Number of PyTorch workers
+    retry_when_fail : bool
+        If True, retry when function failed.
+    session
+        MaxFrame session, if not provided, will use default one.
+    run_kwargs : dict
+        Extra kwargs for `session.run`.
+    port : int
+        Port of PyTorch worker or ps, will automatically increase for the same worker
+
+    Returns
+    -------
+    status
+        return {'status': 'ok'} if succeeded, or error raised
+    """
+    kwargs = kwargs or {}
+    packed = [args, kwargs]
+    inputs = find_objects(packed, TILEABLE_TYPE)
+    input_to_var = {t: f"torch_var_{i}" for i, t in enumerate(inputs)}
+    var_to_input = {v: k for k, v in input_to_var.items()}
+    structure = replace_objects(packed, input_to_var)
+
+    replaces = {
+        "pickled_func": pickle.dumps(func),
+        "nested": pickle.dumps(structure),
+        "var_names": pickle.dumps(list(input_to_var.values())),
+    }
+    code = _script_template % replaces
+    return run_pytorch_script(
+        io.StringIO(code),
+        n_workers,
+        data=var_to_input,
+        retry_when_fail=retry_when_fail,
+        session=session,
+        run_kwargs=run_kwargs,
+        port=port,
+        execute=execute,
+    )

maxframe/learn/contrib/pytorch/run_script.py

@@ -0,0 +1,102 @@
+# Copyright 1999-2025 Alibaba Group Holding Ltd.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import os
+from typing import Any, BinaryIO, Dict, List, Optional, TextIO, Union
+
+from .... import opcodes
+from ....remote.run_script import RunScript, _extract_inputs
+from ....serialization.serializables import Int32Field, StringField
+from ....typing_ import SessionType, TileableType
+from ....utils import to_binary
+
+logger = logging.getLogger(__name__)
+
+
+class RunPyTorch(RunScript):
+    _op_type_ = opcodes.RUN_PYTORCH
+
+    # used for chunk op
+    master_port = Int32Field("master_port", default=None)
+    master_addr = StringField("master_addr", default=None)
+    init_method = StringField("init_method", default=None)
+    master_waiter_name = StringField("master_waiter_name", default=None)
+
+
+def run_pytorch_script(
+    script: Union[bytes, str, BinaryIO, TextIO],
+    n_workers: int,
+    data: Dict[str, TileableType] = None,
+    gpu: Optional[bool] = None,
+    command_argv: List[str] = None,
+    retry_when_fail: bool = False,
+    session: SessionType = None,
+    run_kwargs: Dict[str, Any] = None,
+    port: int = None,
+    execute: bool = True,
+):
+    """
+    Run PyTorch script in MaxFrame cluster.
+
+    Parameters
+    ----------
+    script: str or file-like object
+        Script to run
+    n_workers : int
+        Number of PyTorch workers
+    data : dict
+        Variable name to data.
+    gpu : bool
+        Run PyTorch script on GPU
+    command_argv : list
+        Extra command args for script
+    retry_when_fail : bool
+        If True, retry when function failed.
+    session
+        MaxFrame session, if not provided, will use default one.
+    run_kwargs : dict
+        Extra kwargs for `session.run`.
+    port : int
+        Port of PyTorch worker or ps, will automatically increase for the same worker
+
+    Returns
+    -------
+    status
+        return {'status': 'ok'} if succeeded, or error raised
+    """
+    if int(n_workers) <= 0:
+        raise ValueError("n_workers should be at least 1")
+    if hasattr(script, "read"):
+        code = script.read()
+    else:
+        with open(os.path.abspath(script), "rb") as f:
+            code = f.read()
+
+    inputs = _extract_inputs(data)
+    port = 29500 if port is None else port
+    op = RunPyTorch(
+        data=data,
+        code=to_binary(code),
+        world_size=int(n_workers),
+        retry_when_fail=retry_when_fail,
+        gpu=gpu,
+        master_port=port,
+        command_args=command_argv,
+    )
+    t = op(inputs)
+    if execute:
+        t.execute(session=session, **(run_kwargs or {}))
+    else:
+        return t

maxframe/learn/contrib/pytorch/tests/__init__.py

@@ -0,0 +1,13 @@
+# Copyright 1999-2025 Alibaba Group Holding Ltd.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

maxframe/learn/contrib/pytorch/tests/test_pytorch.py

@@ -0,0 +1,42 @@
+# Copyright 1999-2025 Alibaba Group Holding Ltd.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ..... import dataframe as md
+from ..... import tensor as mt
+from .....core import TILEABLE_TYPE
+from ..run_function import run_pytorch_function
+
+
+def test_run_function():
+    def test_func(a, b):
+        raise ValueError(a + b)
+
+    df = md.DataFrame(mt.random.rand(10, 10))
+    t = run_pytorch_function(test_func, args=(df, 1), execute=False)
+
+    global_dict = {"__name__": "__main__"}
+    global_dict.update(t.op.data)
+
+    try:
+        exec(t.op.code, global_dict)
+    except ValueError as ex:
+        result = ex.args[0]
+        assert isinstance(result, TILEABLE_TYPE)
+        assert result.op.lhs.key == df.key
+        assert result.op.rhs == 1
+    else:
+        raise AssertionError("Error not raised")
+
+    assert global_dict["torch_var_0"] is df
+    assert "torch_var_1" not in global_dict

maxframe/learn/contrib/utils.py

@@ -0,0 +1,108 @@
+# Copyright 1999-2025 Alibaba Group Holding Ltd.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import sys
+
+from ...serialization.serializables import Serializable
+
+
+def make_import_error_func(package_name):
+    def _func(*_, **__):  # pragma: no cover
+        raise ImportError(
+            f"Cannot import {package_name}, please reinstall that package."
+        )
+
+    return _func
+
+
+def config_mod_getattr(mod_dict, globals_):
+    def __getattr__(name):
+        import importlib
+
+        if name in mod_dict:
+            mod_name, cls_name = mod_dict[name].rsplit(".", 1)
+            mod = importlib.import_module(mod_name, globals_["__name__"])
+            cls = globals_[name] = getattr(mod, cls_name)
+            return cls
+        else:  # pragma: no cover
+            raise AttributeError(name)
+
+    if sys.version_info[:2] < (3, 7):
+        for _mod in mod_dict.keys():
+            __getattr__(_mod)
+
+    def __dir__():
+        return sorted([n for n in globals_ if not n.startswith("_")] + list(mod_dict))
+
+    globals_.update(
+        {
+            "__getattr__": __getattr__,
+            "__dir__": __dir__,
+            "__all__": list(__dir__()),
+            "__warningregistry__": dict(),
+        }
+    )
+
+
+class TrainingCallback(Serializable):
+    _local_cls = None
+
+    @classmethod
+    def _load_local_to_remote_mapping(cls, globals_dict):
+        if cls._local_to_remote:
+            return
+        for v in globals_dict.values():
+            if isinstance(v, type) and issubclass(v, cls) and v._local_cls is not None:
+                cls._local_to_remote[v._local_cls] = v
+
+    @classmethod
+    def from_local(cls, callback_obj):
+        if isinstance(callback_obj, (list, tuple)):
+            return [cls.from_local(x) for x in callback_obj]
+        if not type(callback_obj) in cls._local_to_remote:
+            return callback_obj
+
+        kw = {}
+        remote_cls = cls._local_to_remote[type(callback_obj)]
+        for attr in remote_cls._FIELDS:
+            try:
+                kw[attr] = getattr(callback_obj, attr)
+            except AttributeError:
+                pass
+        return remote_cls(**kw)
+
+    def has_custom_code(self) -> bool:
+        return False
+
+    @classmethod
+    def remote_to_local(cls, remote_obj):
+        if isinstance(remote_obj, (list, tuple)):
+            return [cls.remote_to_local(x) for x in remote_obj]
+        if not isinstance(remote_obj, TrainingCallback):
+            return remote_obj
+        return remote_obj.to_local()
+
+    def _extract_kw(self) -> dict:
+        kw = {}
+        for attr in type(self)._FIELDS:
+            val = getattr(self, attr, None)
+            if val is not None:
+                kw[attr] = val
+        return kw
+
+    def to_local(self):
+        return type(self)._local_cls(**self._extract_kw())
+
+    def __call__(self, *args, **kwargs):
+        return self.to_local()(*args, **kwargs)

maxframe/learn/contrib/xgboost/__init__.py

@@ -0,0 +1,33 @@
+# Copyright 1999-2025 Alibaba Group Holding Ltd.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .core import Booster
+from .dmatrix import DMatrix
+from .predict import predict
+from .train import train
+
+# isort: off
+from ..utils import config_mod_getattr as _config_mod_getattr
+
+_config_mod_getattr(
+    {
+        "XGBClassifier": ".classifier.XGBClassifier",
+        "XGBRegressor": ".regressor.XGBRegressor",
+    },
+    globals(),
+)
+
+del _config_mod_getattr
+
+__all__ = ["Booster", "DMatrix", "XGBClassifier", "XGBRegressor", "predict", "train"]

maxframe/learn/contrib/xgboost/callback.py

@@ -0,0 +1,86 @@
+# Copyright 1999-2025 Alibaba Group Holding Ltd.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, Optional, Sequence, Union
+
+from ....serialization.serializables import (
+    AnyField,
+    BoolField,
+    Float32Field,
+    Int32Field,
+    StringField,
+)
+from ....udf import BuiltinFunction
+from ..utils import TrainingCallback
+
+try:
+    from xgboost.callback import EarlyStopping as _EarlyStopping
+    from xgboost.callback import LearningRateScheduler as _LearningRateScheduler
+except ImportError:
+    _LearningRateScheduler = _EarlyStopping = None
+
+
+class XGBTrainingCallback(TrainingCallback):
+    _local_to_remote = {}
+
+    @classmethod
+    def from_local(cls, callback_obj):
+        cls._load_local_to_remote_mapping(globals())
+        return super().from_local(callback_obj)
+
+
+class LearningRateScheduler(XGBTrainingCallback):
+    _local_cls = _LearningRateScheduler
+
+    learning_rates = AnyField("learning_rates", default=None)
+
+    def __init__(
+        self, learning_rates: Union[Callable[[int], float], Sequence[float]], **kw
+    ) -> None:
+        super().__init__(learning_rates=learning_rates, **kw)
+
+    def has_custom_code(self) -> bool:
+        return not isinstance(self.learning_rates, (tuple, list, BuiltinFunction))
+
+
+class EarlyStopping(XGBTrainingCallback):
+    _local_cls = _EarlyStopping
+
+    rounds = Int32Field("rounds")
+    metric_name = StringField("metric_name", default=None)
+    data_name = StringField("data_name", default=None)
+    maximize = BoolField("maximize", default=None)
+    save_best = BoolField("save_best", default=None)
+    min_delta = Float32Field("min_delta", default=None)
+
+    def __init__(
+        self,
+        *,
+        rounds: int,
+        metric_name: Optional[str] = None,
+        data_name: Optional[str] = None,
+        maximize: Optional[bool] = None,
+        save_best: Optional[bool] = False,
+        min_delta: float = 0.0,
+        **kw
+    ) -> None:
+        super().__init__(
+            rounds=rounds,
+            metric_name=metric_name,
+            data_name=data_name,
+            maximize=maximize,
+            save_best=save_best,
+            min_delta=min_delta,
+            **kw
+        )