varbase_patch_methods.py 30.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# 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.

15
import inspect
16
import numpy as np
17 18
import warnings
import weakref
19
import sys
20 21

import paddle
22 23
from .. import framework
from .. import core
24
from .. import unique_name
25
from ..framework import Variable, Parameter, ParamBase, _getitem_impl_, _setitem_impl_, _in_eager_mode, EagerParamBase
26
from .base import switch_to_static_graph
27
from .math_op_patch import monkey_patch_math_varbase
28
from .parallel import scale_loss
L
Leo Chen 已提交
29
from paddle.fluid.data_feeder import convert_dtype, _PADDLE_DTYPE_2_NUMPY_DTYPE
30
import paddle.utils.deprecated as deprecated
31 32


33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
class TensorHookRemoveHelper(object):
    """
    A helper class that for removing Tensor gradient's hook.
    """

    def __init__(self, tensor, hook_id):
        self._tensor_ref = weakref.ref(tensor)
        self._hook_id = hook_id

    def remove(self):
        """
        Remove reference Tensor's hook.

        Returns:
            bool: Return True if removed successfully
        """
        tensor = self._tensor_ref()
        if tensor is not None:
            res = tensor._remove_grad_hook(self._hook_id)
            if res is True:
                return True
            else:
                warnings.warn(
                    "The backward hook (ID: %d) of Tensor `%s` you want to remove does not exist or has been removed."
                    % (self._hook_id, tensor.name), RuntimeWarning)
        return False


61 62 63
_already_patch_repr = False


64
def monkey_patch_varbase():
65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91
    @switch_to_static_graph
    def _to_static_var(self, to_parameter=False, **kwargs):
        """
        **Notes**:
            **This API is ONLY available in Dygraph mode**

        Transform a VarBase into static Variable with same attributes. It's a low level interface used
        in dy2static and shall not be called directly.

        Args:
            to_parameter (bool): It takes effect only if the input a VarBase. If set True,
                                 the VarBase will be converted into framework.Parameters. Otherwise, it will
                                 be converted into framework.Variable. Default False.

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                from paddle.fluid.dygraph.base import to_variable
                import numpy as np

                data = np.ones([3, 1024], dtype='float32')
                with fluid.dygraph.guard():
                    var_base = to_variable(data)
                    static_var = var_base._to_static_var()

        """
92

93
        # Note: getattr(self, attr, None) will call x.grad=x.gradient(), but gradient() only available in dygraph.
94 95
        # It will fail. So, for propery that different between dynamic and static graph, should not getattr(self, attr, None).
        attr_not_need_keys = ['grad', 'T']
96 97 98
        if isinstance(self, ParamBase):
            attr_kwargs = self.__dict__.copy()
        else:
99 100
            attr_names = []
            for name in dir(self):
101 102 103 104
                if name not in attr_not_need_keys:
                    if not inspect.ismethod(getattr(
                            self, name)) and not name.startswith('_'):
                        attr_names.append(name)
105 106 107 108 109 110 111 112 113 114
            attr_kwargs = {name: getattr(self, name) for name in attr_names}

        attr_keys = ['block', 'shape', 'dtype', 'type', 'name', 'persistable']
        for attr in attr_keys:
            attr_kwargs[attr] = getattr(self, attr, None)

        attr_kwargs.update(kwargs)

        if to_parameter or isinstance(self, ParamBase):
            del attr_kwargs['persistable']
115 116
            # NOTE(Aurelius84): All parameters should be placed into global block.
            attr_kwargs['block'] = attr_kwargs['block'].program.global_block()
117 118 119 120 121
            static_var = Parameter(**attr_kwargs)
        else:
            static_var = Variable(**attr_kwargs)
        return static_var

122 123 124 125 126
    # TODO(jiabin): move this to cplusplus end if we find some performance issue on it
    @framework.dygraph_only
    def set_value(self, value):
        """
        **Notes**:
T
tianshuo78520a 已提交
127
            **This API is ONLY available in Dygraph mode**
128 129 130 131 132 133 134 135 136 137 138

        Set a new value for this Variable.

        Args:
            value (Variable|np.ndarray): the new value.

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                from paddle.fluid.dygraph.base import to_variable
139
                from paddle.fluid.dygraph import Linear
140 141
                import numpy as np

142
                data = np.ones([3, 1024], dtype='float32')
143
                with fluid.dygraph.guard():
144
                    linear = fluid.dygraph.Linear(1024, 4)
145
                    t = to_variable(data)
146
                    linear(t)  # call with default weight
147
                    custom_weight = np.random.randn(1024, 4).astype("float32")
148 149
                    linear.weight.set_value(custom_weight)  # change existing weight
                    out = linear(t)  # call with different weight
150 151

        """
152
        if core._in_eager_mode():
153 154 155 156
            base_tensor = core.eager.EagerTensor
        else:
            base_tensor = core.VarBase
        assert isinstance(value, (np.ndarray, base_tensor, dict, str)), \
S
Steffy-zxf 已提交
157 158 159 160 161 162 163 164 165 166 167 168 169
            "Variable set_value function, arguments type only support Variable, numpy, VarBase, dict, string."

        if isinstance(value, (dict, str)):
            assert len(self) == len(
                value
            ), "Variable length not match, Variable [ {} ] need tensor with length {} but load set tensor with length {}".format(
                self.name, len(self), len(value))
            if isinstance(value, dict):
                self.value().set_vocab(value)
            else:
                self.value().set_string_list(value)
        else:
            value_np = value
170
            if isinstance(value, base_tensor):
S
Steffy-zxf 已提交
171
                value_np = value.numpy()
172

S
Steffy-zxf 已提交
173
            self_tensor_np = self.numpy()
174

S
Steffy-zxf 已提交
175 176 177
            assert self_tensor_np.shape == value_np.shape, \
                "Variable Shape not match, Variable [ {} ] need tensor with shape {} but load set tensor with shape {}".format(
                    self.name, self_tensor_np.shape, value_np.shape)
178

S
Steffy-zxf 已提交
179 180 181
            assert self_tensor_np.dtype == value_np.dtype, \
                "Variable dtype not match, Variable [ {} ] need tensor with dtype {}  but load tensor with dtype {}".format(
                    self.name, self_tensor_np.dtype, value_np.dtype)
182

S
Steffy-zxf 已提交
183 184
            self.value().get_tensor().set(value_np,
                                          framework._current_expected_place())
185 186

    @framework.dygraph_only
187
    def backward(self, grad_tensor=None, retain_graph=False):
188
        """
189
        Run backward of current Graph which starts from current Tensor.
190

191 192 193 194
        The new gradient will accumulat on previous gradient.

        You can clear gradient by ``Tensor.clear_grad()`` .

195
        Args:
196 197 198 199 200
            grad_tensor(Tensor, optional): initial gradient values of the current Tensor. If `grad_tensor` is None, 
            the initial gradient values of the current Tensor would be Tensor filled with 1.0; 
            if `grad_tensor` is not None, it must have the same length as the current Tensor.
            Teh default value is None.

201
            retain_graph(bool, optional): If False, the graph used to compute grads will be freed. If you would
202 203 204
                like to add more ops to the built graph after calling this method( :code:`backward` ), set the parameter
                :code:`retain_graph` to True, then the grads will be retained. Thus, seting it to False is much more memory-efficient.
                Defaults to False.
205 206 207 208 209 210
        Returns:
            NoneType: None

        Examples:
            .. code-block:: python

211
                import paddle
212 213 214 215 216 217 218 219 220 221 222 223 224 225
                x = paddle.to_tensor(5., stop_gradient=False)
                for i in range(5):
                    y = paddle.pow(x, 4.0)
                    y.backward()
                    print("{}: {}".format(i, x.grad))
                # 0: [500.]
                # 1: [1000.]
                # 2: [1500.]
                # 3: [2000.]
                # 4: [2500.]

                x.clear_grad()
                print("{}".format(x.grad))
                # 0.
226

227 228 229 230 231 232 233 234 235 236 237
                grad_tensor=paddle.to_tensor(2.)
                for i in range(5):
                    y = paddle.pow(x, 4.0)
                    y.backward(grad_tensor)
                    print("{}: {}".format(i, x.grad))
                # 0: [1000.]
                # 1: [2000.]
                # 2: [3000.]
                # 3: [4000.]
                # 4: [5000.]

238 239
        """
        if framework.in_dygraph_mode():
240
            if grad_tensor is not None:
241
                if core._in_eager_mode():
242 243 244 245 246 247 248
                    assert isinstance(
                        grad_tensor, core.eager.EagerTensor
                    ), "The type of grad_tensor must be paddle.Tensor"
                else:
                    assert isinstance(
                        grad_tensor, paddle.
                        Tensor), "The type of grad_tensor must be paddle.Tensor"
249 250 251 252
                assert grad_tensor.shape == self.shape, \
                    "Tensor shape not match, Tensor of grad_tensor [ {} ] with shape {} mismatch Tensor [ {} ] with shape {}".format(
                    grad_tensor.name, grad_tensor.shape, self.name, self.shape)

253
            if core._in_eager_mode():
254 255 256 257
                if grad_tensor is None:
                    grad_tensor = []
                else:
                    grad_tensor = [grad_tensor]
K
kuizhiqing 已提交
258
            if paddle.is_compiled_with_xpu() or paddle.is_compiled_with_npu():
259
                # TODO(liuyuhui): Currently only for xpu. Will be removed in the future.
260
                scaled_loss = scale_loss(self)
261
                if core._in_eager_mode():
262 263 264 265 266 267
                    core.eager.run_backward([scaled_loss], grad_tensor,
                                            retain_graph)
                else:
                    core.dygraph_run_backward([scaled_loss], [grad_tensor],
                                              retain_graph,
                                              framework._dygraph_tracer())
268
            else:
269
                if core._in_eager_mode():
270 271 272 273 274
                    core.eager.run_backward([self], grad_tensor, retain_graph)
                else:
                    core.dygraph_run_backward([self], [grad_tensor],
                                              retain_graph,
                                              framework._dygraph_tracer())
275 276
        else:
            raise ValueError(
T
tianshuo78520a 已提交
277
                "Variable.backward() is only available in DyGraph mode")
278 279

    @framework.dygraph_only
280 281
    @deprecated(
        since="2.1.0",
282 283
        level=1,
        reason="Please use tensor.grad, which returns the tensor value of the gradient."
284
    )
285 286
    def gradient(self):
        """
287 288 289 290
        .. warning::
          This API will be deprecated in the future, it is recommended to use
          :code:`x.grad` which returns the tensor value of the gradient.

291
        Get the Gradient of Current Tensor.
292 293

        Returns:
294
            ndarray: Numpy value of the gradient of current Tensor
295 296 297 298

        Examples:
            .. code-block:: python

299
                import paddle
300

301 302 303
                x = paddle.to_tensor(5., stop_gradient=False)
                y = paddle.pow(x, 4.0)
                y.backward()
304
                print("grad of x: {}".format(x.gradient()))
305
                # [500.]
306 307

        """
308
        if core._in_eager_mode():
309 310 311 312 313 314 315
            if not self.grad._is_initialized():
                return None
            # TODO(wanghuancoder) support SELECTED_ROWS
            return self.grad.numpy()
        else:
            if self._grad_ivar() is None:
                return None
316

317 318 319 320
            new_ivar = self._grad_ivar()._copy_to(core.CPUPlace(), True)
            if self._grad_ivar().type == core.VarDesc.VarType.SELECTED_ROWS:
                return (
                    np.array(new_ivar.value().get_selected_rows().get_tensor()),
321
                    np.array(new_ivar.value().get_selected_rows().rows()))
322 323
            else:
                return np.array(new_ivar.value().get_tensor())
324

325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391
    @framework.dygraph_only
    def register_hook(self, hook):
        """
        Registers a backward hook for current Tensor.

        The hook will be called every time the gradient Tensor of current Tensor is computed.

        The hook should not modify the input gradient Tensor, but it can optionally return
        a new gradient Tensor which will be used in place of current Tensor's gradient.

        The hook should have the following signature:

            hook(grad) -> Tensor or None

        Args:
            hook(function): A backward hook to be registered for Tensor.grad

        Returns:
            TensorHookRemoveHelper: A helper object that can be used to remove the registered hook by calling `remove()` method.

        Examples:
            .. code-block:: python

                import paddle

                # hook function return None
                def print_hook_fn(grad):
                    print(grad)

                # hook function return Tensor
                def double_hook_fn(grad):
                    grad = grad * 2
                    return grad

                x = paddle.to_tensor([0., 1., 2., 3.], stop_gradient=False)
                y = paddle.to_tensor([4., 5., 6., 7.], stop_gradient=False)
                z = paddle.to_tensor([1., 2., 3., 4.])

                # one Tensor can register multiple hooks
                h = x.register_hook(print_hook_fn)
                x.register_hook(double_hook_fn)

                w = x + y
                # register hook by lambda function
                w.register_hook(lambda grad: grad * 2)

                o = z.matmul(w)
                o.backward()
                # print_hook_fn print content in backward
                # Tensor(shape=[4], dtype=float32, place=CUDAPlace(0), stop_gradient=False,
                #        [2., 4., 6., 8.])

                print("w.grad:", w.grad) # w.grad: [1. 2. 3. 4.]
                print("x.grad:", x.grad) # x.grad: [ 4.  8. 12. 16.]
                print("y.grad:", y.grad) # y.grad: [2. 4. 6. 8.]

                # remove hook
                h.remove()
        """
        if self.stop_gradient is True:
            raise RuntimeError(
                "Cannot register hook on a tensor that stop gradient.")

        hook_id = self._register_grad_hook(hook)
        helper = TensorHookRemoveHelper(self, hook_id)
        return helper

392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420
    @framework.dygraph_only
    def _to(self, device=None, dtype=None, blocking=None):

        if device is None and dtype is None and blocking is None:
            return self

        if device is not None:
            if isinstance(device, str):
                device = paddle.device._convert_to_place(device)
            elif isinstance(device, (core.CPUPlace, core.CUDAPlace,
                                     core.CUDAPinnedPlace, core.XPUPlace)):
                pass
            else:
                raise ValueError(
                    "device value error, must be str, paddle.CPUPlace(), paddle.CUDAPlace(), paddle.CUDAPinnedPlace() or paddle.XPUPlace(), but the type of device is "
                    + type(device).__name__)

        if blocking is None:
            blocking = True
        else:
            assert isinstance(
                blocking,
                bool), "blocking value error, must be the True, False or None"

        def transform(t, device, dtype, blocking):
            if device is None:
                device = t.place
            if dtype is None:
                dtype = t.dtype
421 422
            if type(dtype) is str:
                dtype = framework.convert_np_dtype_to_dtype_(dtype)
423 424 425

            # 1. gpu place need to determine whether the memory is sufficient for allocation.
            if t.place.is_gpu_place():
426
                size_dtype = core.size_of_dtype(dtype)
427 428 429 430 431
                # Note(weilong wu): Paddle GPU minimum memory allocation unit is 256 bytes,
                # waiting_alloc_memory will compute the memory space occupied by 't'.
                # Coefficient 1.2 is used to avoid OOM that may occur in this critical state when the memory is just enough.
                waiting_alloc_memory = (
                    (t._numel() * size_dtype) / 256 + 1) * 256 * 1.2
432
                gpu_memory_available = core.gpu_memory_available()
433 434 435 436 437 438 439 440 441 442 443 444 445
                if gpu_memory_available < waiting_alloc_memory:
                    # Copy Tensor to cpu
                    t_used = t._copy_to(paddle.CPUPlace(), blocking)
                    # Release memory of t
                    t._clear()
                else:
                    # Tensor still in GPU
                    t_used = t
            else:
                t_used = t

            # 2. cast Tensor to dtype
            if dtype is not None and dtype != t_used.dtype:
446 447 448
                with paddle.fluid.framework._dygraph_place_guard(
                        place=t_used.place):
                    t_casted = t_used.cast(dtype=dtype)
449 450 451 452
            else:
                t_casted = t_used

            # 3. Copy casted Tensor(in CPU or GPU) to device
453 454 455 456
            if device is not None and not t_casted.place._equals(device):
                new_t = t_casted._copy_to(device, blocking)
            else:
                new_t = t_casted
457 458 459 460 461 462 463 464 465 466 467 468

            # 4. Share Tensor to origin Tensor
            dst_tensor = t.value().get_tensor()
            src_tensor = new_t.value().get_tensor()
            dst_tensor._share_data_with(src_tensor)

            return t

        with warnings.catch_warnings():
            warnings.filterwarnings("ignore", category=UserWarning)
            return transform(self, device, dtype, blocking)

469 470 471
    @property
    def grad(self):
        """
472 473 474 475 476 477 478 479 480 481 482 483 484
        .. warning::
          This API will return the tensor value of the gradient. If you want 
          to get the numpy value of the gradient, you can use :code:`x.grad.numpy()`.

        Get the Gradient of Current Tensor.

        Returns:
            Tensor: the gradient of current Tensor

        Examples:
            .. code-block:: python

                import paddle
485

486 487 488 489 490 491 492
                x = paddle.to_tensor(5., stop_gradient=False)
                y = paddle.pow(x, 4.0)
                y.backward()
                print("grad of x: {}".format(x.grad))
                # Tensor(shape=[1], dtype=float32, place=CUDAPlace(0), stop_gradient=False, [500.])

        """
493 494 495 496
        msg = 'tensor.grad will return the tensor value of the gradient.' \
            ' This is an incompatible upgrade for tensor.grad API. ' \
            ' It\'s return type changes from numpy.ndarray in version 2.0 to paddle.Tensor in version 2.1.0. ' \
            ' If you want to get the numpy value of the gradient, you can use :code:`x.grad.numpy()`'
497
        warning_msg = "\033[93m\nWarning:\n%s \033[0m" % (msg)
498 499 500
        # ensure ANSI escape sequences print correctly in cmd and powershell
        if sys.platform.lower() == 'win32':
            warning_msg = "\nWarning:\n%s " % (msg)
501
        warnings.warn(warning_msg)
502
        return self._grad_ivar()
503

504 505 506 507 508 509
    def clear_grad(self):
        """
        The alias of clear_gradient().
        """
        self.clear_gradient()

510 511
    def item(self, *args):
        """
512 513
        Convert element at specific position in Tensor into Python scalars. If the position is not specified, the Tensor must be a 
        single-element Tensor.
514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551

        Args:
            *args(int): The input coordinates. If it's single int, the data in the corresponding order of flattened Tensor will be returned.
                Default: None, and it must be in the case where Tensor has only one element.

        Returns(Python scalar): A Python scalar, whose dtype is corresponds to the dtype of Tensor.

        Raises:
            ValueError: If the Tensor has more than one element, there must be coordinates.
        
        Examples:
            .. code-block:: python

                import paddle

                x = paddle.to_tensor(1)
                print(x.item())             #1
                print(type(x.item()))       #<class 'int'>

                x = paddle.to_tensor(1.0)
                print(x.item())             #1.0
                print(type(x.item()))       #<class 'float'>

                x = paddle.to_tensor(True)
                print(x.item())             #True
                print(type(x.item()))       #<class 'bool'>

                x = paddle.to_tensor(1+1j)
                print(x.item())             #(1+1j)
                print(type(x.item()))       #<class 'complex'>

                x = paddle.to_tensor([[1.1, 2.2, 3.3]])
                print(x.item(2))            #3.3
                print(x.item(0, 2))         #3.3

        """
        return self._getitem_from_offset(*args).item()

552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572
    @property
    def inplace_version(self):
        """
        The inplace version of current Tensor.
        The version number is incremented whenever the current Tensor is modified through an inplace operation.

        **Notes: This is a read-only property**

        Examples:
          .. code-block:: python

            import paddle
            var = paddle.ones(shape=[4, 2, 3], dtype="float32")
            print(var.inplace_version)  # 0

            var[1] = 2.2
            print(var.inplace_version)  # 1

        """
        return self._inplace_version()

573 574
    def __str__(self):
        """
575
        Convert a VarBase object to a readable string.
576

577
        Returns(str): A readable string.
578 579 580 581

        Examples:
            .. code-block:: python

582
                import paddle
583
                x = paddle.rand([2, 5])
584
                print(x)
585 586 587 588
                
                # Tensor(shape=[2, 5], dtype=float32, place=CPUPlace,
                #        [[0.30574632, 0.55739117, 0.30902600, 0.39413780, 0.44830436],
                #         [0.79010487, 0.53972793, 0.09495186, 0.44267157, 0.72112119]])
589
        """
590
        if core._in_eager_mode():
591 592 593 594 595
            from paddle.tensor.to_string import eager_tensor_to_string
            return eager_tensor_to_string(self)
        else:
            from paddle.tensor.to_string import to_string
            return to_string(self)
596

597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621
    def __deepcopy__(self, memo):
        """
        Deep copy Tensor, it will always performs Tensor copy.

        Examples:
            .. code-block:: python

                import paddle
                import copy
                x = paddle.to_tensor(2.)
                y = copy.deepcopy(x)
                
                print(x)
                # Tensor(shape=[1], dtype=float32, place=CPUPlace, stop_gradient=True,
                #        [2.])

                print(y)
                # Tensor(shape=[1], dtype=float32, place=CPUPlace, stop_gradient=True,
                #        [2.])

        """
        if not self.is_leaf:
            raise RuntimeError(
                "Only Leaf Tensor support the deepcopy at the moment, non-Leaf Tensors contains graph information that does't support deepcopy"
            )
622
        if core._in_eager_mode():
623 624 625
            new_varbase = core.eager.EagerTensor()
        else:
            new_varbase = core.VarBase()
626 627 628 629 630
        new_varbase.name = self.name + unique_name.generate("_deepcopy")
        memo[id(self)] = new_varbase
        new_varbase.copy_(self, True)
        return new_varbase

631 632 633
    @property
    def block(self):
        return framework.default_main_program().global_block()
634

635 636 637 638 639 640 641 642 643 644
    def __nonzero__(self):
        numel = np.prod(self.shape)
        assert numel == 1, "When Variable is used as the condition of if/while , Variable can only contain one element."
        tensor = self.value().get_tensor()
        assert tensor._is_initialized(), "tensor not initialized"
        return bool(np.all(tensor.__array__() > 0))

    def __bool__(self):
        return self.__nonzero__()

645
    def __array__(self, dtype=None):
646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669
        """
        Returns a numpy array shows the value of current Tensor.
        
        Returns:
            ndarray: The numpy value of current Tensor.

        Returns type:
            ndarray: dtype is same as current Tensor

        Examples:
            .. code-block:: python

                import paddle
                import numpy as np
                x = paddle.randn([2, 2])
                x_array = np.array(x)

                print(type(x_array))      #<class 'numpy.ndarray'>
                print(x_array.shape)      #(2, 2)
        """
        array = self.numpy()
        if dtype:
            array = array.astype(dtype)
        return array
670

W
WeiXin 已提交
671
    def contain_tensor(item):
672
        if not isinstance(item, (tuple, list)):
W
WeiXin 已提交
673 674 675 676 677 678 679 680 681
            item = [item]

        for slice_item in item:
            if isinstance(slice_item, slice):
                if isinstance(slice_item.start, Variable)  \
                    or isinstance(slice_item.stop, Variable) \
                        or isinstance(slice_item.step, Variable):
                    return True
            else:
W
WeiXin 已提交
682 683 684
                if isinstance(
                        slice_item,
                    (Variable, np.ndarray)) and Variable.dtype != paddle.bool:
W
WeiXin 已提交
685 686 687
                    return True
        return False

688
    def __getitem__(self, item):
W
WeiXin 已提交
689 690 691 692 693 694
        def is_list_tuple(index, contain_type):
            def _is_list_tuple(item):
                if isinstance(item, (tuple, list)):
                    for s in item:
                        if not _is_list_tuple(s):
                            return False
695 696 697
                else:
                    if type(item) != contain_type:
                        return False
W
WeiXin 已提交
698
                return True
699

W
WeiXin 已提交
700 701 702 703 704 705 706 707
            if not isinstance(index, (tuple, list)):
                return False
            for s in index:
                if not _is_list_tuple(s):
                    return False
            return True

        if contain_tensor(item) or is_list_tuple(item, int):
708 709 710 711 712 713 714 715
            # 1. Call _getitem_impl_ when item contains tensor.
            # Why not call a c++ function ? Because item can't be parsed when it contains tensor.
            return _getitem_impl_(self, item)

        else:
            # 2. Call c++ func getitem_index_not_tensor to speedup.
            return self._getitem_index_not_tensor(item)

W
WeiXin 已提交
716
    def __setitem__(self, item, value):
Z
zyfncg 已提交
717 718 719
        def contain_tensor_or_list(item):
            if not isinstance(item, tuple):
                item = [item]
W
WeiXin 已提交
720

Z
zyfncg 已提交
721 722 723 724 725 726 727 728
            for slice_item in item:
                if isinstance(slice_item, list):
                    return True
                elif isinstance(slice_item, Variable):
                    return True

            return False

729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750
        def is_combine_index(item):
            var_type = None
            item_type = None
            if isinstance(item, (tuple, list)):
                for slice_item in item:
                    if item_type is None:
                        item_type = type(slice_item)
                    else:
                        if type(slice_item) != item_type:
                            return True

                    if isinstance(slice_item, Variable):
                        if var_type is None:
                            var_type = slice_item.dtype
                        else:
                            if var_type != slice_item.dtype:
                                return True
                return False

            return False

        if contain_tensor_or_list(item) and not is_combine_index(item):
Z
zyfncg 已提交
751 752
            # To reuse code with static graph,
            # Call _setitem_impl_ when item contains tensor or list.
W
WeiXin 已提交
753 754 755
            return _setitem_impl_(self, item, value)

        else:
Z
zyfncg 已提交
756
            # Call c++ func __setitem_varbase__ to speedup.
W
WeiXin 已提交
757 758
            return self.__setitem_varbase__(item, value)

759 760 761 762 763 764 765
    @framework.dygraph_only
    def _grad_ivar(self):
        if self.grad._is_initialized():
            return self.grad
        else:
            return None

766 767 768 769 770 771 772 773
    @framework.dygraph_only
    def _set_grad_ivar(self, value):
        if isinstance(self, EagerParamBase):
            self.grad = value
        else:
            raise TypeError(
                "_set_grad_ivar is only supported for Parameter Tensor")

774 775 776 777 778 779 780
    @framework.dygraph_only
    def clear_gradient(self, set_to_zero=True):
        if set_to_zero:
            self._zero_grads()
        else:
            self._clear_gradient()

781 782 783 784
    @framework.dygraph_only
    def clone(self):
        return _C_ops_.assign(self)

785 786 787
    if core._in_eager_mode() and not hasattr(core, "eager"):
        return

788 789
    for method_name, method in (
        ("__bool__", __bool__), ("__nonzero__", __nonzero__),
790
        ("_to_static_var", _to_static_var), ("set_value", set_value),
791
        ("block", block), ("backward", backward), ("clear_grad", clear_grad),
792 793 794 795
        ("inplace_version", inplace_version), ("gradient", gradient),
        ("register_hook", register_hook), ("__str__", __str__),
        ("__repr__", __str__), ("__deepcopy__", __deepcopy__),
        ("__module__", "paddle"), ("__array__", __array__),
W
WeiXin 已提交
796
        ("__getitem__", __getitem__), ("item", item),
797
        ("__setitem__", __setitem__), ("_to", _to)):
798 799
        if core._in_eager_mode():
            setattr(core.eager.EagerTensor, method_name, method)
L
Leo Chen 已提交
800
        else:
801 802 803 804
            setattr(core.VarBase, method_name, method)

    if core._in_eager_mode():
        setattr(core.eager.EagerTensor, "_grad_ivar", _grad_ivar)
805
        setattr(core.eager.EagerTensor, "_set_grad_ivar", _set_grad_ivar)
806
        setattr(core.eager.EagerTensor, "clear_gradient", clear_gradient)
807
        setattr(core.eager.EagerTensor, "clone", clone)
808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825
    else:
        setattr(core.VarBase, "__name__", "Tensor")
        setattr(core.VarBase, "grad", grad)

    global _already_patch_repr
    if not _already_patch_repr:
        # NOTE(zhiqiu): pybind11 will set a default __str__ method of enum class.
        # So, we need to overwrite it to a more readable one.
        # See details in https://github.com/pybind/pybind11/issues/2537.
        origin = getattr(core.VarDesc.VarType, "__repr__")

        def dtype_str(dtype):
            if dtype in _PADDLE_DTYPE_2_NUMPY_DTYPE:
                prefix = 'paddle.'
                return prefix + _PADDLE_DTYPE_2_NUMPY_DTYPE[dtype]
            else:
                # for example, paddle.fluid.core.VarDesc.VarType.LOD_TENSOR
                return origin(dtype)
L
Leo Chen 已提交
826

827 828
        setattr(core.VarDesc.VarType, "__repr__", dtype_str)
        _already_patch_repr = True
L
Leo Chen 已提交
829

830 831
    # patch math methods for varbase
    monkey_patch_math_varbase()