varbase_patch_methods.py 32.8 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
from .. import framework
C
crystal 已提交
23
from ..framework import convert_np_dtype_to_dtype_
24
from .. import core
25
from .. import unique_name
J
Jiabin Yang 已提交
26
from ..framework import Variable, Parameter, ParamBase, _getitem_impl_, _setitem_impl_, EagerParamBase
27
from .base import switch_to_static_graph
28
from .math_op_patch import monkey_patch_math_varbase
29
from .parallel import scale_loss
L
Leo Chen 已提交
30
from paddle.fluid.data_feeder import convert_dtype, _PADDLE_DTYPE_2_NUMPY_DTYPE
31
import paddle.utils.deprecated as deprecated
C
chenjian 已提交
32
import paddle.profiler as profiler
H
hong 已提交
33
from paddle import _C_ops
34

35 36
_grad_scalar = None

37

38 39 40
class TensorHookRemoveHelper(object):
    """
    A helper class that for removing Tensor gradient's hook.
41
    NOTE(wuweilong):the operation weakref.ref(tensor) will cause some unexpected errors in eager mode.
42 43 44
    """

    def __init__(self, tensor, hook_id):
J
Jiabin Yang 已提交
45 46
        self._tensor = tensor if framework._in_eager_mode_ else weakref.ref(
            tensor)
47 48 49 50 51 52 53 54 55
        self._hook_id = hook_id

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

        Returns:
            bool: Return True if removed successfully
        """
J
Jiabin Yang 已提交
56
        tensor = self._tensor if framework._in_eager_mode_ else self._tensor()
57 58 59 60 61 62 63 64 65 66 67
        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


68 69 70
_already_patch_repr = False


71
def monkey_patch_varbase():
72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98
    @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()

        """
99

100
        # Note: getattr(self, attr, None) will call x.grad=x.gradient(), but gradient() only available in dygraph.
101 102
        # 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']
J
Jiabin Yang 已提交
103
        if isinstance(self, (ParamBase, EagerParamBase)):
104 105
            attr_kwargs = self.__dict__.copy()
        else:
106 107
            attr_names = []
            for name in dir(self):
108 109 110 111
                if name not in attr_not_need_keys:
                    if not inspect.ismethod(getattr(
                            self, name)) and not name.startswith('_'):
                        attr_names.append(name)
112 113 114 115 116 117 118 119
            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)

J
Jiabin Yang 已提交
120
        if to_parameter or isinstance(self, (ParamBase, EagerParamBase)):
121
            del attr_kwargs['persistable']
122 123
            # NOTE(Aurelius84): All parameters should be placed into global block.
            attr_kwargs['block'] = attr_kwargs['block'].program.global_block()
124 125 126 127 128
            static_var = Parameter(**attr_kwargs)
        else:
            static_var = Variable(**attr_kwargs)
        return static_var

129 130 131 132 133
    # 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
134
            **This API is ONLY available in Dygraph mode**
135 136 137 138 139 140 141 142 143 144 145

        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
146
                from paddle.fluid.dygraph import Linear
147 148
                import numpy as np

149
                data = np.ones([3, 1024], dtype='float32')
150
                with fluid.dygraph.guard():
151
                    linear = fluid.dygraph.Linear(1024, 4)
152
                    t = to_variable(data)
153
                    linear(t)  # call with default weight
154
                    custom_weight = np.random.randn(1024, 4).astype("float32")
155 156
                    linear.weight.set_value(custom_weight)  # change existing weight
                    out = linear(t)  # call with different weight
157 158

        """
J
Jiabin Yang 已提交
159
        if framework._in_eager_mode_:
160
            base_tensor = core.eager.Tensor
161 162 163
        else:
            base_tensor = core.VarBase
        assert isinstance(value, (np.ndarray, base_tensor, dict, str)), \
S
Steffy-zxf 已提交
164 165 166 167 168 169 170 171 172 173 174 175
            "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:
C
crystal 已提交
176
            assert self.shape == list(value.shape),  \
S
Steffy-zxf 已提交
177
                "Variable Shape not match, Variable [ {} ] need tensor with shape {} but load set tensor with shape {}".format(
C
crystal 已提交
178 179 180 181 182 183
                    self.name, self.shape, value.shape)

            if isinstance(value, base_tensor):
                dtype = value.dtype
            else:
                dtype = convert_np_dtype_to_dtype_(value.dtype)
184

C
crystal 已提交
185
            assert self.dtype == dtype, \
S
Steffy-zxf 已提交
186
                "Variable dtype not match, Variable [ {} ] need tensor with dtype {}  but load tensor with dtype {}".format(
C
crystal 已提交
187
                    self.name, self.dtype, dtype)
188

189
            # NOTE(wuweilong): self could be VarBase or Tensor, the subsequent behavior are defined in different files
190
            # if self is VarBase, method value() return Variable that bindded in imperative.cc, get_tensor() bindded in pybind.cc
191
            # if self is Tensor, method value() return self that defined in this file, get_tensor() defined in eager_method.cc
192
            # this Interface behavior will be unifed in the future.
C
crystal 已提交
193
            self.value().get_tensor().set(value,
S
Steffy-zxf 已提交
194
                                          framework._current_expected_place())
195 196

    @framework.dygraph_only
197
    def backward(self, grad_tensor=None, retain_graph=False):
198
        """
199
        Run backward of current Graph which starts from current Tensor.
200

201 202 203 204
        The new gradient will accumulat on previous gradient.

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

205
        Args:
C
chenjian 已提交
206 207
            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;
208 209 210
            if `grad_tensor` is not None, it must have the same length as the current Tensor.
            Teh default value is None.

211
            retain_graph(bool, optional): If False, the graph used to compute grads will be freed. If you would
212 213 214
                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.
215 216 217 218 219 220
        Returns:
            NoneType: None

        Examples:
            .. code-block:: python

221
                import paddle
222 223 224 225 226 227 228 229 230 231 232 233 234 235
                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.
236

237 238 239 240 241 242 243 244 245 246 247
                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.]

248
        """
J
Jiabin Yang 已提交
249
        if framework._non_static_mode():
C
chenjian 已提交
250 251 252
            record_event = profiler.RecordEvent(
                "Gradient Backward", profiler.TracerEventType.Backward)
            record_event.begin()
253
            if grad_tensor is not None:
J
Jiabin Yang 已提交
254
                if framework._in_eager_mode_:
255
                    assert isinstance(
256 257
                        grad_tensor, core.eager.
                        Tensor), "The type of grad_tensor must be paddle.Tensor"
258 259 260 261
                else:
                    assert isinstance(
                        grad_tensor, paddle.
                        Tensor), "The type of grad_tensor must be paddle.Tensor"
262 263 264 265
                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)

J
Jiabin Yang 已提交
266
            if framework._in_eager_mode_:
267 268 269 270
                if grad_tensor is None:
                    grad_tensor = []
                else:
                    grad_tensor = [grad_tensor]
271 272 273
            if _grad_scalar:
                # When using amp with Fleet DistributedStrategy, we do loss scaling implicitly.
                self = _grad_scalar.scale(self)
K
kuizhiqing 已提交
274
            if paddle.is_compiled_with_xpu() or paddle.is_compiled_with_npu():
275
                # TODO(liuyuhui): Currently only for xpu. Will be removed in the future.
276
                scaled_loss = scale_loss(self)
J
Jiabin Yang 已提交
277
                if framework._in_eager_mode_:
278 279 280 281 282 283
                    core.eager.run_backward([scaled_loss], grad_tensor,
                                            retain_graph)
                else:
                    core.dygraph_run_backward([scaled_loss], [grad_tensor],
                                              retain_graph,
                                              framework._dygraph_tracer())
284
            else:
J
Jiabin Yang 已提交
285
                if framework._in_eager_mode_:
286 287 288 289 290
                    core.eager.run_backward([self], grad_tensor, retain_graph)
                else:
                    core.dygraph_run_backward([self], [grad_tensor],
                                              retain_graph,
                                              framework._dygraph_tracer())
C
chenjian 已提交
291
            record_event.end()
292 293
        else:
            raise ValueError(
T
tianshuo78520a 已提交
294
                "Variable.backward() is only available in DyGraph mode")
295 296

    @framework.dygraph_only
297 298
    @deprecated(
        since="2.1.0",
299 300
        level=1,
        reason="Please use tensor.grad, which returns the tensor value of the gradient."
301
    )
302 303
    def gradient(self):
        """
304 305 306 307
        .. 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.

308
        Get the Gradient of Current Tensor.
309 310

        Returns:
311
            ndarray: Numpy value of the gradient of current Tensor
312 313 314 315

        Examples:
            .. code-block:: python

316
                import paddle
317

318 319 320
                x = paddle.to_tensor(5., stop_gradient=False)
                y = paddle.pow(x, 4.0)
                y.backward()
321
                print("grad of x: {}".format(x.gradient()))
322
                # [500.]
323 324

        """
J
Jiabin Yang 已提交
325
        if framework._in_eager_mode_:
326
            if self.grad is None:
327
                return None
328 329
            if self.grad.is_selected_rows():
                return (np.array(self.grad.numpy()), np.array(self.grad.rows()))
330 331 332 333
            return self.grad.numpy()
        else:
            if self._grad_ivar() is None:
                return None
334

335 336 337 338
            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()),
339
                    np.array(new_ivar.value().get_selected_rows().rows()))
340 341
            else:
                return np.array(new_ivar.value().get_tensor())
342

343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409
    @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

410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438
    @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
439 440
            if type(dtype) is str:
                dtype = framework.convert_np_dtype_to_dtype_(dtype)
441 442 443

            # 1. gpu place need to determine whether the memory is sufficient for allocation.
            if t.place.is_gpu_place():
444
                size_dtype = core.size_of_dtype(dtype)
445 446 447 448 449
                # 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
450
                gpu_memory_available = core.gpu_memory_available()
451 452 453 454 455 456 457 458 459 460 461 462 463
                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:
464 465 466
                with paddle.fluid.framework._dygraph_place_guard(
                        place=t_used.place):
                    t_casted = t_used.cast(dtype=dtype)
467 468 469 470
            else:
                t_casted = t_used

            # 3. Copy casted Tensor(in CPU or GPU) to device
471 472 473 474
            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
475 476 477 478 479 480 481 482 483 484 485 486

            # 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)

487 488 489
    @property
    def grad(self):
        """
490
        .. warning::
C
chenjian 已提交
491
          This API will return the tensor value of the gradient. If you want
492 493 494 495 496 497 498 499 500 501 502
          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
503

504 505 506 507 508 509 510
                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.])

        """
511 512 513 514
        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()`'
515
        warning_msg = "\033[93m\nWarning:\n%s \033[0m" % (msg)
516 517 518
        # ensure ANSI escape sequences print correctly in cmd and powershell
        if sys.platform.lower() == 'win32':
            warning_msg = "\nWarning:\n%s " % (msg)
519
        warnings.warn(warning_msg)
520
        return self._grad_ivar()
521

522 523 524 525 526 527
    def clear_grad(self):
        """
        The alias of clear_gradient().
        """
        self.clear_gradient()

528 529
    def item(self, *args):
        """
C
chenjian 已提交
530
        Convert element at specific position in Tensor into Python scalars. If the position is not specified, the Tensor must be a
531
        single-element Tensor.
532 533 534 535 536 537 538 539 540

        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.
C
chenjian 已提交
541

542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569
        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()

570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590
    @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()

591 592
    def __str__(self):
        """
593
        Convert a VarBase object to a readable string.
594

595
        Returns(str): A readable string.
596 597 598 599

        Examples:
            .. code-block:: python

600
                import paddle
601
                x = paddle.rand([2, 5])
602
                print(x)
C
chenjian 已提交
603

604 605 606
                # 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]])
607
        """
J
Jiabin Yang 已提交
608
        if framework._in_eager_mode_:
609 610
            from paddle.tensor.to_string import tensor_to_string
            return tensor_to_string(self)
611 612 613
        else:
            from paddle.tensor.to_string import to_string
            return to_string(self)
614

615 616 617 618 619 620 621 622 623 624 625
    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)
C
chenjian 已提交
626

627 628 629 630 631 632 633 634 635 636 637 638 639
                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"
            )
J
Jiabin Yang 已提交
640
        if framework._in_eager_mode_:
641
            new_varbase = core.eager.Tensor()
642 643
        else:
            new_varbase = core.VarBase()
644 645 646 647 648
        new_varbase.name = self.name + unique_name.generate("_deepcopy")
        memo[id(self)] = new_varbase
        new_varbase.copy_(self, True)
        return new_varbase

649 650 651
    @property
    def block(self):
        return framework.default_main_program().global_block()
652

653 654 655
    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."
J
Jiabin Yang 已提交
656
        if framework._in_eager_mode_:
657 658 659 660 661 662
            assert self._is_initialized(), "tensor not initialized"
            return bool(np.all(self.numpy() > 0))
        else:
            tensor = self.value().get_tensor()
            assert tensor._is_initialized(), "tensor not initialized"
            return bool(np.all(tensor.__array__() > 0))
663 664 665 666

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

667
    def __array__(self, dtype=None):
668 669
        """
        Returns a numpy array shows the value of current Tensor.
C
chenjian 已提交
670

671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691
        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
692

W
WeiXin 已提交
693
    def contain_tensor(item):
694
        if not isinstance(item, (tuple, list)):
W
WeiXin 已提交
695 696 697 698 699 700 701 702 703
            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
704 705 706
                if isinstance(
                        slice_item,
                    (Variable, np.ndarray)) and Variable.dtype != paddle.bool:
W
WeiXin 已提交
707 708 709
                    return True
        return False

710
    def __getitem__(self, item):
W
WeiXin 已提交
711 712 713 714 715 716
        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
717 718 719
                else:
                    if type(item) != contain_type:
                        return False
W
WeiXin 已提交
720
                return True
721

W
WeiXin 已提交
722 723 724 725 726 727 728 729
            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):
730 731 732 733 734 735 736 737
            # 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
738
    def __setitem__(self, item, value):
Z
zyfncg 已提交
739 740 741
        def contain_tensor_or_list(item):
            if not isinstance(item, tuple):
                item = [item]
W
WeiXin 已提交
742

Z
zyfncg 已提交
743 744 745 746 747 748 749 750
            for slice_item in item:
                if isinstance(slice_item, list):
                    return True
                elif isinstance(slice_item, Variable):
                    return True

            return False

751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772
        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
773 774
            # To reuse code with static graph,
            # Call _setitem_impl_ when item contains tensor or list.
W
WeiXin 已提交
775 776 777
            return _setitem_impl_(self, item, value)

        else:
J
Jiabin Yang 已提交
778
            if framework._in_eager_mode_:
W
wanghuancoder 已提交
779 780 781 782
                return self.__setitem_eager_tensor__(item, value)
            else:
                # Call c++ func __setitem_varbase__ to speedup.
                return self.__setitem_varbase__(item, value)
W
WeiXin 已提交
783

784 785
    @framework.dygraph_only
    def _grad_ivar(self):
786 787 788 789
        if self.grad is not None:
            if self.grad._is_initialized():
                return self.grad
        return None
790

791 792 793 794 795 796 797 798 799 800
    @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")

    @framework.dygraph_only
    def clone(self):
H
hong 已提交
801
        return _C_ops.assign(self)
802

803 804 805 806
    @framework.dygraph_only
    def value(self):
        return self

J
Jiabin Yang 已提交
807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834
    @framework.dygraph_only
    def _slice(self, begin_idx, end_idx):
        return core.eager.Tensor(self.get_tensor()._slice(begin_idx, end_idx))

    @framework.dygraph_only
    def _numel(self):
        return self.get_tensor()._numel()

    @framework.dygraph_only
    def cpu(self):
        if self.place.is_cpu_place():
            return self
        else:
            res = self._copy_to(core.CPUPlace(), True)
            res.stop_gradient = self.stop_gradient
            res.persistable = self.persistable
            return res

    @framework.dygraph_only
    def cuda(self, device_id, blocking):
        if self.place.is_gpu_place():
            return self
        else:
            res = self._copy_to(core.CUDAPlace(device_id), True)
            res.stop_gradient = self.stop_gradient
            res.persistable = self.persistable
            return res

J
Jiabin Yang 已提交
835
    if framework._in_eager_mode_ and not hasattr(core, "eager"):
836 837
        return

838 839
    for method_name, method in (
        ("__bool__", __bool__), ("__nonzero__", __nonzero__),
840
        ("_to_static_var", _to_static_var), ("set_value", set_value),
841
        ("block", block), ("backward", backward), ("clear_grad", clear_grad),
842 843 844 845
        ("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
846
        ("__getitem__", __getitem__), ("item", item),
847
        ("__setitem__", __setitem__), ("_to", _to)):
J
Jiabin Yang 已提交
848
        if framework._in_eager_mode_:
849
            setattr(core.eager.Tensor, method_name, method)
L
Leo Chen 已提交
850
        else:
851 852
            setattr(core.VarBase, method_name, method)

J
Jiabin Yang 已提交
853
    if framework._in_eager_mode_:
854 855 856 857
        setattr(core.eager.Tensor, "_grad_ivar", _grad_ivar)
        setattr(core.eager.Tensor, "_set_grad_ivar", _set_grad_ivar)
        setattr(core.eager.Tensor, "clone", clone)
        setattr(core.eager.Tensor, "value", value)
J
Jiabin Yang 已提交
858 859 860 861
        setattr(core.eager.Tensor, "cpu", cpu)
        setattr(core.eager.Tensor, "cuda", cuda)
        setattr(core.eager.Tensor, "_slice", _slice)
        setattr(core.eager.Tensor, "_numel", _numel)
862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879
    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
880

881 882
        setattr(core.VarDesc.VarType, "__repr__", dtype_str)
        _already_patch_repr = True
L
Leo Chen 已提交
883

884 885
    # patch math methods for varbase
    monkey_patch_math_varbase()