loss.py 47.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
#   Copyright (c) 2020 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 paddle
16 17 18
from ...fluid.layer_helper import LayerHelper
from ...fluid.data_feeder import check_variable_and_dtype
import paddle.fluid as fluid
19

20
# TODO: define loss functions of neural network
21
import numpy as np
22 23 24 25
import paddle
import paddle.fluid as fluid
from ...fluid.framework import core, in_dygraph_mode
from ...fluid.layers.nn import _elementwise_op_in_dygraph
26 27 28 29 30 31 32
from ...fluid.layers import bpr_loss  #DEFINE_ALIAS
from ...fluid.layers import center_loss  #DEFINE_ALIAS
from ...fluid.layers import dice_loss  #DEFINE_ALIAS
from ...fluid.layers import iou_similarity  #DEFINE_ALIAS
from ...fluid.layers import log_loss  #DEFINE_ALIAS
from ...fluid.layers import npair_loss  #DEFINE_ALIAS
from ...fluid.layers import rank_loss  #DEFINE_ALIAS
33
from ...fluid.layers import reshape
34 35 36 37 38 39 40 41
from ...fluid.layers import sigmoid_cross_entropy_with_logits  #DEFINE_ALIAS
from ...fluid.layers import sigmoid_focal_loss  #DEFINE_ALIAS
from ...fluid.layers import smooth_l1  #DEFINE_ALIAS
from ...fluid.layers import softmax_with_cross_entropy  #DEFINE_ALIAS
from ...fluid.layers import square_error_cost  #DEFINE_ALIAS
from ...fluid.layers import ssd_loss  #DEFINE_ALIAS
from ...fluid.layers import teacher_student_sigmoid_loss  #DEFINE_ALIAS

42 43
from ...fluid.layers import edit_distance  #DEFINE_ALIAS
from ...fluid.layers import sampled_softmax_with_cross_entropy  #DEFINE_ALIAS
44
from ...fluid.layer_helper import LayerHelper
45
from ...fluid.framework import in_dygraph_mode
46
from ...fluid.framework import _varbase_creator
47
from ...fluid.framework import Variable
48

49
__all__ = [
50
    'binary_cross_entropy',
51
    'binary_cross_entropy_with_logits',
52 53 54 55
    'bpr_loss',
    'center_loss',
    'cross_entropy',
    'dice_loss',
56
    'edit_distance',
57
    'iou_similarity',
58
    'kl_div',
59
    'l1_loss',
60 61
    'log_loss',
    'mse_loss',
62
    'margin_ranking_loss',
63
    #       'nce',
64
    'nll_loss',
65 66
    'npair_loss',
    'rank_loss',
67
    'sampled_softmax_with_cross_entropy',
68 69 70
    'sigmoid_cross_entropy_with_logits',
    'sigmoid_focal_loss',
    'smooth_l1',
71
    'smooth_l1_loss',
72 73 74
    'softmax_with_cross_entropy',
    'square_error_cost',
    'ssd_loss',
75 76
    'teacher_student_sigmoid_loss',
    'ctc_loss',
77
]
78 79


80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140
def binary_cross_entropy(input, label, weight=None, reduction='mean',
                         name=None):
    """
    This op measures the binary_cross_entropy loss between input predictions ``input``
    and target labels ``label`` . The binary_cross_entropy loss can be described as:

    If :attr:`weight` is set, the loss is:

    .. math::
        Out = -1 * weight * (label * log(input) + (1 - label) * log(1 - input))

    If :attr:`weight` is None, the loss is:

    .. math::
        Out = -1 * (label * log(input) + (1 - label) * log(1 - input))

    If :attr:`reduction` set to ``'none'``, the interface will return the original loss `Out`.

    If :attr:`reduction` set to ``'mean'``, the reduced mean loss is:

    .. math::
        Out = MEAN(Out)

    If :attr:`reduction` set to ``'sum'``, the reduced sum loss is:

    .. math::
        Out = SUM(Out)

    Note that the input predictions ``input`` always be the output of sigmoid, and the target labels ``label``
    should be numbers between 0 and 1.

    Parameters:
        input (Tensor): The input predications tensor. 2-D tensor with shape: [N, *],
            N is batch_size, `*` means number of additional dimensions. The ``input``
            should always be the output of sigmod.  Available dtype is float32, float64.
        label (Tensor): The target labels tensor. 2-D tensor with the same shape as
            ``input``. The target labels which values should be numbers between 0 and 1.
            Available dtype is float32, float64.
        weight (Tensor, optional): A manual rescaling weight given to the loss of each
            batch element. If given, has to be a Tensor of size nbatch and the data type
            is float32, float64. Default is ``'None'``.
        reduction (str, optional): Indicate how to average the loss by batch_size,
            the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
            If :attr:`reduction` is ``'none'``, the unreduced loss is returned;
            If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned;
            If :attr:`reduction` is ``'sum'``, the summed loss is returned.
            Default is ``'mean'``.
        name (str, optional): Name for the operation (optional, default is None).
            For more information, please refer to :ref:`api_guide_Name`.


    Returns:
        output (Tensor): If ``reduction`` is ``'none'``, the shape of output is
            same as ``input`` , else the shape of output is scalar.

    Examples:
        .. code-block:: python

            import paddle

            paddle.disable_static()
141 142
            input = paddle.to_tensor([0.5, 0.6, 0.7], 'float32')
            label = paddle.to_tensor([1.0, 0.0, 1.0], 'float32')
143 144 145 146 147 148 149 150 151 152 153
            output = paddle.nn.functional.binary_cross_entropy(input, label)
            print(output.numpy())  # [0.65537095]

    """
    if reduction not in ['sum', 'mean', 'none']:
        raise ValueError(
            "The value of 'reduction' in binary_cross_entropy should be 'sum', "
            "'mean' or 'none', but received %s, which is not allowed." %
            reduction)

    if in_dygraph_mode():
154
        out = core.ops.bce_loss(input, label)
155 156 157 158 159 160 161
        if weight is not None:
            out = core.ops.elementwise_mul(out, weight, 'axis', -1)

        if reduction == 'sum':
            return core.ops.reduce_sum(out, 'dim', [0], 'keep_dim', False,
                                       "reduce_all", True)
        elif reduction == 'mean':
Z
Zhong Hui 已提交
162
            return core.ops.mean(out)
163 164 165 166 167 168 169 170 171
        else:
            return out

    fluid.data_feeder.check_variable_and_dtype(
        input, 'input', ['float32', 'float64'], 'binary_cross_entropy')
    fluid.data_feeder.check_variable_and_dtype(
        label, 'label', ['float32', 'float64'], 'binary_cross_entropy')

    sub_name = name if weight is None and reduction is 'none' else None
172 173 174 175 176 177 178 179 180
    helper = LayerHelper("binary_cross_entropy", name=sub_name)
    out = helper.create_variable_for_type_inference(dtype=input.dtype)
    helper.append_op(
        type='bce_loss',
        inputs={
            'X': [input],
            'Label': [label],
        },
        outputs={'Out': [out]})
181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197

    if weight is not None:
        if isinstance(weight, paddle.framework.Variable):
            weight_name = name if reduction is 'none' else None
            out = paddle.multiply(out, weight, axis=-1, name=weight_name)
        else:
            raise ValueError(
                "The weight is not a Tensor, please convert to Tensor.")

    if reduction == 'sum':
        return paddle.sum(out, name=name)
    elif reduction == 'mean':
        return paddle.mean(out, name=name)
    else:
        return out


198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274
def binary_cross_entropy_with_logits(logit,
                                     label,
                                     weight=None,
                                     reduction='mean',
                                     pos_weight=None,
                                     name=None):
    """
    This operator combines the sigmoid layer and the :ref:`api_nn_loss_BCELoss` layer.
    Also, we can see it as the combine of ``sigmoid_cross_entropy_with_logits``
    layer and some reduce operations.

    This measures the element-wise probability error in classification tasks
    in which each class is independent.
    This can be thought of as predicting labels for a data-point, where labels
    are not mutually exclusive. For example, a news article can be about
    politics, technology or sports at the same time or none of these.

    First this operator calculate loss function as follows:

    .. math::
           Out = -Labels * \\log(\\sigma(Logit)) - (1 - Labels) * \\log(1 - \\sigma(Logit))

    We know that :math:`\\sigma(Logit) = \\frac{1}{1 + \\e^{-Logit}}`. By substituting this we get:

    .. math::
           Out = Logit - Logit * Labels + \\log(1 + \\e^{-Logit})

    For stability and to prevent overflow of :math:`\\e^{-Logit}` when Logit < 0,
    we reformulate the loss as follows:

    .. math::
           Out = \\max(Logit, 0) - Logit * Labels + \\log(1 + \\e^{-\|Logit\|})

    Then, if ``weight`` or ``pos_weight`` is not None, this operator multiply the
    weight tensor on the loss `Out`. The ``weight`` tensor will attach different
    weight on every items in the batch. The ``pos_weight`` will attach different
    weight on the positive label of each class.

    Finally, this operator applies reduce operation on the loss.
    If :attr:`reduction` set to ``'none'``, the operator will return the original loss `Out`.
    If :attr:`reduction` set to ``'mean'``, the reduced mean loss is :math:`Out = MEAN(Out)`.
    If :attr:`reduction` set to ``'sum'``, the reduced sum loss is :math:`Out = SUM(Out)`.

    Note that the target labels ``label`` should be numbers between 0 and 1.

    Args:
        logit (Tensor): The input predications tensor. 2-D tensor with shape: [N, *],
            N is batch_size, `*` means number of additional dimensions. The ``logit``
            is usually the output of Linear layer. Available dtype is float32, float64.
        label (Tensor): The target labels tensor. 2-D tensor with the same shape as
            ``logit``. The target labels which values should be numbers between 0 and 1.
            Available dtype is float32, float64.
        weight (Tensor, optional): A manual rescaling weight given to the loss of each
            batch element. If given, it has to be a 1D Tensor whose size is `[N, ]`,
            The data type is float32, float64. Default is ``'None'``.
        reduction (str, optional): Indicate how to average the loss by batch_size,
            the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
            If :attr:`reduction` is ``'none'``, the unreduced loss is returned;
            If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned;
            If :attr:`reduction` is ``'sum'``, the summed loss is returned.
            Default is ``'mean'``.
        pos_weight (Tensor, optional): A weight of positive examples. Must be a vector
            with length equal to the number of classes. The data type is float32, float64.
            Default is ``'None'``.
        name (str, optional): Name for the operation (optional, default is None).
            For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        output (Tensor): If ``reduction`` is ``'none'``, the shape of output is
            same as ``logit`` , else the shape of output is scalar.

    Examples:

        .. code-block:: python

            import paddle
            paddle.disable_static()
275 276
            logit = paddle.to_tensor([5.0, 1.0, 3.0])
            label = paddle.to_tensor([1.0, 0.0, 1.0])
277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345
            output = paddle.nn.functional.binary_cross_entropy_with_logits(logit, label)
            print(output.numpy())  # [0.45618808]

    """
    if reduction not in ['sum', 'mean', 'none']:
        raise ValueError(
            "The value of 'reduction' in binary_cross_entropy_with_logits "
            "should be 'sum', 'mean' or 'none', but received %s, which is not allowed."
            % reduction)

    if in_dygraph_mode():
        one = _varbase_creator(dtype=logit.dtype)
        core.ops.fill_constant(one, 'value',
                               float(1.0), 'force_cpu', False, 'dtype',
                               one.dtype, 'str_value', '1.0', 'shape', [1])
        out = core.ops.sigmoid_cross_entropy_with_logits(logit, label)
        if pos_weight is not None:
            log_weight = core.ops.elementwise_add(
                core.ops.elementwise_mul(
                    label, core.ops.elementwise_sub(pos_weight, one)), one)
            out = core.ops.elementwise_mul(out, log_weight)
        if weight is not None:
            out = core.ops.elementwise_mul(out, weight)

        if reduction == "sum":
            return core.ops.reduce_sum(out, 'reduce_all', True)
        elif reduction == "mean":
            return core.ops.mean(out)
        else:
            return out

    fluid.data_feeder.check_variable_and_dtype(
        logit, 'logit', ['float32', 'float64'],
        'binary_cross_entropy_with_logits')
    fluid.data_feeder.check_variable_and_dtype(
        label, 'label', ['float32', 'float64'],
        'binary_cross_entropy_with_logits')
    sigmoid_name = None
    if reduction == 'none' and pos_weight is None and weight is None:
        sigmoid_name = name

    out = paddle.nn.functional.sigmoid_cross_entropy_with_logits(
        logit, label, name=sigmoid_name)

    one = paddle.fill_constant(shape=[1], value=1.0, dtype=logit.dtype)
    if pos_weight is not None:
        fluid.data_feeder.check_variable_and_dtype(
            pos_weight, 'pos_weight', ['float32', 'float64'],
            'binary_cross_entropy_with_logits')
        log_weight = paddle.add(
            paddle.multiply(label, paddle.elementwise_sub(pos_weight, one)),
            one)
        pos_weight_name = name if reduction == 'none' and weight is None else None
        out = paddle.multiply(out, log_weight, name=pos_weight_name)

    if weight is not None:
        fluid.data_feeder.check_variable_and_dtype(
            weight, 'weight', ['float32', 'float64'],
            'binary_cross_entropy_with_logits')
        weight_name = name if reduction == 'none' else None
        out = paddle.multiply(out, weight, name=weight_name)

    if reduction == "sum":
        return paddle.sum(out, name=name)
    elif reduction == "mean":
        return paddle.mean(out, name=name)
    return out


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
def smooth_l1_loss(input, label, reduction='mean', delta=1.0, name=None):
    """
    This operator calculates smooth_l1_loss. Creates a criterion that uses a squared
    term if the absolute element-wise error falls below 1 and an L1 term otherwise.
    In some cases it can prevent exploding gradients and it is more robust and less
    sensitivity to outliers. Also known as the Huber loss:

    .. math::

         loss(x,y)=\\frac{1}{n}\\sum_{i}z_i


    where z_i is given by:

    .. math::

         \\mathop{z_i}=\\left\\{\\begin{array}{rcl}
        0.5(x_i - y_i)^2 & & {if |x_i - y_i| < delta} \\\\
        delta * |x_i - y_i| - 0.5 * delta^2 & & {otherwise}
        \\end{array} \\right.

    Parameters:
        input (Tensor): Input tensor, the data type is float32 or float64. Shape is
            (N, C), where C is number of classes, and if shape is more than 2D, this
            is (N, C, D1, D2,..., Dk), k >= 1.
        label (Tensor): Label tensor, the data type is float32 or float64. The shape of label
            is the same as the shape of input.
        reduction (str, optional): Indicate how to average the loss by batch_size,
            the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
            If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned;
            If :attr:`reduction` is ``'sum'``, the reduced sum loss is returned.
            If :attr:`reduction` is ``'none'``, the unreduced loss is returned.
            Default is ``'mean'``.
379
        delta (float, optional): Specifies the hyperparameter delta to be used.
380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423
            The value determines how large the errors need to be to use L1. Errors
            smaller than delta are minimized with L2. Parameter is ignored for
            negative/zero values. Default = 1.0
        name (str, optional): Name for the operation (optional, default is
            None). For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        The tensor variable storing the smooth_l1_loss of input and label.

    Return type: Tensor.

    Examples:
        .. code-block:: python

            import paddle
            import numpy as np

            paddle.disable_static()
            input_data = np.random.rand(3,3).astype("float32")
            label_data = np.random.rand(3,3).astype("float32")
            input = paddle.to_tensor(input_data)
            label = paddle.to_tensor(label_data)
            output = paddle.nn.functioanl.smooth_l1_loss(input, label)
            print(output.numpy())
    """
    fluid.data_feeder.check_variable_and_dtype(
        input, 'input', ['float32', 'float64'], 'smooth_l1_loss')
    fluid.data_feeder.check_variable_and_dtype(
        label, 'label', ['float32', 'float64'], 'smooth_l1_loss')

    out = huber_loss(input=input, label=label, delta=delta)

    if reduction not in ['sum', 'mean', 'none']:
        raise ValueError(
            "The value of 'reduction' in smooth_l1_loss should be 'sum', 'mean' or"
            " 'none', but received %s, which is not allowed." % reduction)
    if reduction == 'none':
        return out
    elif reduction == 'mean':
        return fluid.layers.reduce_mean(out)
    elif reduction == 'sum':
        return fluid.layers.reduce_sum(out)


424 425
def margin_ranking_loss(input,
                        other,
426
                        label,
427 428 429 430 431
                        margin=0.0,
                        reduction='mean',
                        name=None):
    """

432
    This op the calcluate the the margin rank loss between the input, other and label, use the math function as follows.
433

434
    .. math::
435
        margin\_rank\_loss = max(0, -label * (input - other) + margin)
436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451

    If :attr:`reduction` set to ``'mean'``, the reduced mean loss is:

    .. math::
        Out = MEAN(margin\_rank\_loss)

    If :attr:`reduction` set to ``'sum'``, the reduced sum loss is:

    .. math::
        Out = SUM(margin\_rank\_loss)

    If :attr:`reduction` set to ``'none'``, just return the origin ``margin_rank_loss``.

    Parameters:
        input(Tensor): the first input tensor, it's data type should be float32, float64.
        other(Tensor): the second input tensor, it's data type should be float32, float64.
452
        label(Tensor): the label value corresponding to input, it's data type should be float32, float64.
453 454 455 456 457 458 459 460 461 462
        margin (float, optional): The margin value to add, default value is 0;
        reduction (str, optional): Indicate the reduction to apply to the loss, the candicates are ``'none'``, ``'mean'``, ``'sum'``.If :attr:`reduction` is ``'none'``, the unreduced loss is returned; If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned. If :attr:`reduction` is ``'sum'``, the reduced sum loss is returned. Default is ``'mean'``.
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.

    Returns: Tensor, if :attr:`reduction` is ``'mean'`` or ``'sum'``, the out shape is :math:`[1]`, otherwise the shape is the same as `input` .The same dtype as input tensor.

    Examples:

        .. code-block:: python

463
            import paddle
464
            paddle.disable_static()
465

Z
Zhong Hui 已提交
466 467 468
            input = paddle.to_tensor([[1, 2], [3, 4]], dtype='float32')
            other = paddle.to_tensor([[2, 1], [2, 4]], dtype='float32')
            label = paddle.to_tensor([[1, -1], [-1, -1]], dtype='float32')
469
            loss = paddle.nn.functional.margin_ranking_loss(input, other, label)
470 471
            print(loss.numpy()) # [0.75]
    """
472 473 474 475
    if reduction not in ['sum', 'mean', 'none']:
        raise ValueError(
            "The value of 'reduction' in MarginRankingLoss should be 'sum', 'mean' or 'none', but "
            "received %s, which is not allowed." % reduction)
476 477
    if fluid.framework.in_dygraph_mode():
        out = core.ops.elementwise_sub(other, input)
478
        out = core.ops.elementwise_mul(out, label)
479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494
        if margin != 0.0:
            margin = fluid.dygraph.base.to_variable([margin], dtype=out.dtype)
            out = core.ops.elementwise_add(out, margin)
        out = core.ops.relu(out)
        if reduction == 'sum':
            return core.ops.reduce_sum(out, 'reduce_all', True)
        elif reduction == 'mean':
            return core.ops.mean(out)
        return out

    helper = LayerHelper("margin_ranking_loss", **locals())
    fluid.data_feeder.check_variable_and_dtype(
        input, 'input', ['float32', 'float64'], 'margin_rank_loss')
    fluid.data_feeder.check_variable_and_dtype(
        other, 'other', ['float32', 'float64'], 'margin_rank_loss')
    fluid.data_feeder.check_variable_and_dtype(
495
        label, 'label', ['float32', 'float64'], 'margin_rank_loss')
496 497

    out = paddle.elementwise_sub(other, input)
498
    out = paddle.multiply(out, label)
499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529

    if margin != 0.0:
        margin_var = out.block.create_var(dtype=out.dtype)
        paddle.fill_constant([1], out.dtype, margin, out=margin_var)
        out = paddle.add(out, margin_var)

    result_out = helper.create_variable_for_type_inference(input.dtype)

    if reduction == 'none':
        helper.append_op(
            type="relu", inputs={"X": out}, outputs={"Out": result_out})
        return result_out
    elif reduction == 'sum':
        out = paddle.nn.functional.relu(out)
        attrs = {"dim": [0], "keep_dim": False, "reduce_all": True}
        helper.append_op(
            type="reduce_sum",
            inputs={"X": out},
            outputs={"Out": result_out},
            attrs=attrs)
        return result_out
    elif reduction == 'mean':
        out = paddle.nn.functional.relu(out)
        helper.append_op(
            type="mean",
            inputs={"X": out},
            outputs={"Out": result_out},
            attrs={})
        return result_out


530
def l1_loss(input, label, reduction='mean', name=None):
531
    """
532
    This operator computes the L1 Loss of Tensor ``input`` and ``label`` as follows.
533

534
    If `reduction` set to ``'none'``, the loss is:
535 536

    .. math::
537
        Out = \lvert input - label\rvert
538

539
    If `reduction` set to ``'mean'``, the loss is:
540 541

    .. math::
542
        Out = MEAN(\lvert input - label\rvert)
543

544
    If `reduction` set to ``'sum'``, the loss is:
545 546

    .. math::
547
        Out = SUM(\lvert input - label\rvert)
548

549

550
    Parameters:
551 552
        input (Tensor): The input tensor. The shapes is [N, *], where N is batch size and `*` means any number of additional dimensions. It's data type should be float32, float64, int32, int64.
        label (Tensor): label. The shapes is [N, *], same shape as ``input`` . It's data type should be float32, float64, int32, int64.
553
        reduction (str, optional): Indicate the reduction to apply to the loss,
554
            the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
555 556 557
            If `reduction` is ``'none'``, the unreduced loss is returned;
            If `reduction` is ``'mean'``, the reduced mean loss is returned.
            If `reduction` is ``'sum'``, the reduced sum loss is returned.
558 559 560
            Default is ``'mean'``.
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.
    Returns:
561 562 563
        Tensor, the L1 Loss of Tensor ``input`` and ``label``.
            If `reduction` is ``'none'``, the shape of output loss is [N, *], the same as ``input`` .
            If `reduction` is ``'mean'`` or ``'sum'``, the shape of output loss is [1].
564 565 566
    Examples:
        .. code-block:: python
            import paddle
567

568
            paddle.disable_static()
569 570
            input = paddle.to_tensor([[1.5, 0.8], [0.2, 1.3]])
            label = paddle.to_tensor([[1.7, 1], [0.4, 0.5]])
571

572
            l1_loss = paddle.nn.functional.l1_loss(input, label)
573
            print(l1_loss.numpy())
574 575
            # [0.35]

576
            l1_loss = paddle.nn.functional.l1_loss(input, label, reduction='none')
577
            print(l1_loss.numpy())
578 579 580
            # [[0.20000005 0.19999999]
            # [0.2        0.79999995]]

581
            l1_loss = paddle.nn.functional.l1_loss(input, label, reduction='sum')
582
            print(l1_loss.numpy())
583 584 585 586 587 588 589 590 591
            # [1.4]
    """
    if reduction not in ['sum', 'mean', 'none']:
        raise ValueError(
            "The value of 'reduction' in L1Loss should be 'sum', 'mean' or 'none', but "
            "received %s, which is not allowed." % reduction)

    if in_dygraph_mode():
        unreduced = _elementwise_op_in_dygraph(
592
            input, label, axis=-1, act='abs', op_name='elementwise_sub')
593 594 595 596 597 598 599 600 601
        if reduction == 'mean':
            return core.ops.mean(unreduced)
        elif reduction == 'sum':
            return core.ops.reduce_sum(unreduced, 'dim', [0], 'keep_dim', False,
                                       'reduce_all', True)
        else:
            return unreduced

    fluid.data_feeder.check_variable_and_dtype(
602
        input, 'input', ['float32', 'float64', 'int32', 'int64'], 'l1_loss')
603 604 605 606
    fluid.data_feeder.check_variable_and_dtype(
        label, 'label', ['float32', 'float64', 'int32', 'int64'], 'l1_loss')

    if reduction == 'sum':
607
        unreduced = paddle.elementwise_sub(input, label, act='abs')
608 609
        return paddle.sum(unreduced, name=name)
    elif reduction == 'mean':
610
        unreduced = paddle.elementwise_sub(input, label, act='abs')
611 612
        return paddle.mean(unreduced, name=name)
    else:
613
        return paddle.elementwise_sub(input, label, act='abs', name=name)
614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665


def nll_loss(input,
             label,
             weight=None,
             ignore_index=-100,
             reduction='mean',
             name=None):
    """
    This api returns negative log likelihood.
    See more detail in :ref:`api_nn_loss_NLLLoss` .

    Parameters:
         input (Tensor): Input tensor, the shape is :math:`[N, C]`, `C` is the number of classes.
             But in K-dimension situation, the shape is :math:`[N, C, d_1, d_2, ..., d_K]`.
             The data type is float32, float64.
         label (Tensor): Label tensor, the shape is :math:`[N,]` or :math:`[N, d_1, d_2, ..., d_K]`.
             The data type is int64.
         weight (Tensor, optional): Weight tensor, a manual rescaling weight given
             to each class. If given, it has to be a 1D Tensor whose size is `[C, ]`. Otherwise,
             it treated as if having all ones. the data type is
             float32, float64, Default is ``'None'``.
         ignore_index (int64, optional): Specifies a target value that is ignored
             and does not contribute to the input gradient.
         reduction (str, optional): Indicate how to average the loss,
             the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
             If `reduction` is ``'mean'``, the reduced mean loss is returned;
             if `reduction` is ``'sum'``, the reduced sum loss is returned;
             if `reduction` is ``'none'``, no reduction will be apllied.
             Default is ``'mean'``.
         name (str, optional): Name for the operation (optional, default is None).
             For more information, please refer to :ref:`api_guide_Name`.

    Returns:
         `Tensor`, the value of negative log likelihood loss.

    Examples:
        .. code-block:: python
                import paddle
                import numpy as np
                from paddle.nn.functional import nll_loss
                log_softmax = paddle.nn.LogSoftmax(axis=1)

                input_np = np.array([[0.88103855, 0.9908683 , 0.6226845 ],
                                     [0.53331435, 0.07999352, 0.8549948 ],
                                     [0.25879037, 0.39530203, 0.698465  ],
                                     [0.73427284, 0.63575995, 0.18827209],
                                     [0.05689114, 0.0862954 , 0.6325046 ]]).astype(np.float32)
                label_np = np.array([0, 2, 1, 1, 0]).astype(np.int64)

                place = paddle.CPUPlace()
                paddle.disable_static(place)
Z
Zhong Hui 已提交
666
                input = paddle.to_tensor(input_np)
667
                log_out = log_softmax(input)
Z
Zhong Hui 已提交
668
                label = paddle.to_tensor(label_np)
669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722
                result = nll_loss(log_out, label)
                print(result.numpy()) # [1.0720209]
    """
    if reduction not in ['sum', 'mean', 'none']:
        raise ValueError(
            "The value of 'reduction' in nll_loss should be 'sum', 'mean' or "
            "'none', but received %s, which is not allowed." % reduction)

    input_shape = list(input.shape)
    input_dims = len(input_shape)
    if input_dims < 2:
        raise ValueError('Expected 2 or more dimensions (got {})'.format(
            input_dims))
    n = input_shape[0]
    c = input_shape[1]
    if in_dygraph_mode():
        if input_dims != 2 and input_dims != 4:
            input, _ = core.ops.reshape2(input, 'shape', [n, c, 1, -1])
            label, _ = core.ops.reshape2(label, 'shape', [n, 1, -1])
            out_shape = [n] + input_shape[2:]
        out, total_weight = core.ops.nll_loss(input, label, weight,
                                              'ignore_index', ignore_index,
                                              'reduction', reduction)
        if input_dims != 2 and input_dims != 4 and reduction == 'none':
            out, _ = core.ops.reshape2(out, 'shape', out_shape)
        return out

    helper = LayerHelper('nll_loss', **locals())

    if input_dims != 2 and input_dims != 4:
        input = reshape(input, shape=[n, c, 1, -1])
        label = reshape(label, shape=[n, 1, -1])
        out_shape = [n] + input_shape[2:]

    fluid.data_feeder.check_variable_and_dtype(
        input, 'input', ['float32', 'float64'], 'nll_loss')
    fluid.data_feeder.check_variable_and_dtype(label, 'label', ['int64'],
                                               'nll_loss')
    inputs = {'X': input, 'Label': label}
    attrs = {'reduction': reduction, 'ignore_index': ignore_index}
    if weight is not None:
        if isinstance(weight, Variable):
            inputs['Weight'] = weight

    out = helper.create_variable_for_type_inference(dtype=input.dtype)
    total_weight = helper.create_variable_for_type_inference(dtype=input.dtype)
    outputs = {'Out': out, 'Total_weight': total_weight}

    helper.append_op(
        type='nll_loss', inputs=inputs, outputs=outputs, attrs=attrs)
    if input_dims != 2 and input_dims != 4 and reduction == 'none':
        out = reshape(out, shape=out_shape)

    return out
723 724


725 726 727 728 729 730 731 732 733 734 735 736 737
def kl_div(input, label, reduction='mean', name=None):
    """
    This operator calculates the Kullback-Leibler divergence loss
    between Input(X) and Input(Target). Notes that Input(X) is the
    log-probability and Input(Target) is the probability.

    KL divergence loss is calculated as follows:

    $$l(x, y) = y * (\log(y) - x)$$

    While :math:`x` is input and :math:`y` is label.

    While :attr:`reduction` is :attr:`none`, output loss is in
738
    the same shape as input, loss in each point is calculated
739
    seperately and no reduction is applied.
740

741 742
    While :attr:`reduction` is :attr:`mean`, output loss is in
    shape of [1] and loss value is the mean value of all losses.
743

744 745
    While :attr:`reduction` is :attr:`sum`, output loss is in
    shape of [1] and loss value is the sum value of all losses.
746 747

    While :attr:`reduction` is :attr:`batchmean`, output loss is
748 749 750 751
    in shape of [1] and loss value is the sum value of all losses
    divided by batch size.

    Args:
752
        input (Tensor): The input tensor. The shapes is [N, *], where N is batch size and `*` means
753 754 755 756 757 758 759 760 761
             any number of additional dimensions. It's data type should be float32, float64.
        label (Tensor): label. The shapes is [N, *], same shape as ``input`` . It's data type should be float32, float64.
        reduction (Tensor): Indicate how to average the loss,
             the candicates are ``'none'`` | ``'batchmean'`` | ``'mean'`` | ``'sum'``.
             If `reduction` is ``'mean'``, the reduced mean loss is returned;
             If `reduction` is ``'batchmean'``, the sum loss divided by batch size is returned;
             if `reduction` is ``'sum'``, the reduced sum loss is returned;
             if `reduction` is ``'none'``, no reduction will be apllied.
             Default is ``'mean'``.
762
        name(str, optional): Name for the operation (optional, default is None). For more information,
763 764 765 766 767 768 769 770 771 772 773
            please refer to :ref:`api_guide_Name`.

    Returns:
        Tensor: The KL divergence loss. The data type is same as input tensor

    Examples:
        .. code-block:: python

            import paddle
            import numpy as np
            import paddle.nn.functional as F
774

775
            paddle.disable_static()
776

777 778 779 780
            shape = (5, 20)
            input = np.random.uniform(-10, 10, shape).astype('float32')
            target = np.random.uniform(-10, 10, shape).astype('float32')

L
LielinJiang 已提交
781
            # 'batchmean' reduction, loss shape will be [1]
782 783
            pred_loss = F.kl_div(paddle.to_tensor(input),
                                 paddle.to_tensor(target), reduction='batchmean')
L
LielinJiang 已提交
784
            # shape=[1]
785

786
            # 'mean' reduction, loss shape will be [1]
787 788
            pred_loss = F.kl_div(paddle.to_tensor(input),
                                 paddle.to_tensor(target), reduction='mean')
789 790 791
            # shape=[1]

            # 'sum' reduction, loss shape will be [1]
792 793
            pred_loss = F.kl_div(paddle.to_tensor(input),
                                 paddle.to_tensor(target), reduction='sum')
794 795 796
            # shape=[1]

            # 'none' reduction, loss shape is same with input shape
797 798
            pred_loss = F.kl_div(paddle.to_tensor(input),
                                 paddle.to_tensor(target), reduction='none')
799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823
            # shape=[5, 20]

    """
    if paddle.in_dynamic_mode():
        out = core.ops.kldiv_loss(input, label, 'reduction', reduction)
        return out

    helper = LayerHelper('kl_div', **locals())

    fluid.data_feeder.check_variable_and_dtype(input, 'input',
                                               ['float32', 'float64'], 'kl_div')
    fluid.data_feeder.check_variable_and_dtype(label, 'label',
                                               ['float32', 'float64'], 'kl_div')
    fluid.data_feeder.check_type(reduction, 'reduction', str, 'kl_div')

    loss = helper.create_variable_for_type_inference(dtype=input.dtype)
    helper.append_op(
        type='kldiv_loss',
        inputs={'X': input,
                'Target': label},
        outputs={'Loss': loss},
        attrs={'reduction': reduction})
    return loss


824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858
def mse_loss(input, label, reduction='mean', name=None):
    """
    This op accepts input predications and label and returns the mean square error.

    If :attr:`reduction` is set to ``'none'``, loss is calculated as:

    .. math::
        Out = (input - label)^2

    If :attr:`reduction` is set to ``'mean'``, loss is calculated as:

    .. math::
        Out = \operatorname{mean}((input - label)^2)

    If :attr:`reduction` is set to ``'sum'``, loss is calculated as:

    .. math::
        Out = \operatorname{sum}((input - label)^2)

    Parameters:
        input (Tensor): Input tensor, the data type should be float32 or float64.
        label (Tensor): Label tensor, the data type should be float32 or float64.
        reduction (string, optional): The reduction method for the output,
            could be 'none' | 'mean' | 'sum'.
            If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned.
            If :attr:`reduction` is ``'sum'``, the reduced sum loss is returned.
            If :attr:`reduction` is ``'none'``, the unreduced loss is returned.
            Default is ``'mean'``.
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.


    Returns:
        Tensor: The tensor tensor storing the mean square error difference of input and label.

    Return type: Tensor.
859

860 861 862
    Examples:

        .. code-block:: python
863

864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886
            import paddle


            # static graph mode
            paddle.enable_static()
            mse_loss = paddle.nn.loss.MSELoss()
            input = paddle.data(name="input", shape=[1])
            label = paddle.data(name="label", shape=[1])
            place = paddle.CPUPlace()

            output = mse_loss(input,label)
            exe = paddle.static.Executor(place)
            exe.run(paddle.static.default_startup_program())
            output_data = exe.run(
                paddle.static.default_main_program(),
                feed={"input":input_data, "label":label_data},
                fetch_list=[output],
                return_numpy=True)
            print(output_data)
            # [array([0.04000002], dtype=float32)]

            # dynamic graph mode
            paddle.disable_static()
887 888
            input = paddle.to_tensor(1.5)
            label = paddle.to_tensor(1.7)
889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917
            output = mse_loss(input, label)
            print(output.numpy())
            # [0.04000002]

    """

    if reduction not in ['sum', 'mean', 'none']:
        raise ValueError(
            "'reduction' in 'mse_loss' should be 'sum', 'mean' or 'none', "
            "but received {}.".format(reduction))

    if not paddle.fluid.framework.in_dygraph_mode():
        paddle.fluid.data_feeder.check_variable_and_dtype(
            input, 'input', ['float32', 'float64'], 'mse_loss')
        paddle.fluid.data_feeder.check_variable_and_dtype(
            label, 'label', ['float32', 'float64'], 'mse_loss')

    if reduction == 'none':
        return paddle.fluid.layers.square(
            paddle.fluid.layers.elementwise_sub(input, label), name=name)
    elif reduction == 'mean':
        return paddle.mean(
            paddle.fluid.layers.square(
                paddle.fluid.layers.elementwise_sub(input, label)),
            name=name)
    else:
        return paddle.sum(paddle.fluid.layers.square(
            paddle.fluid.layers.elementwise_sub(input, label)),
                          name=name)
918 919


920 921 922 923 924 925 926 927
def ctc_loss(log_probs,
             labels,
             input_lengths,
             label_lengths,
             blank=0,
             reduction='mean'):
    """

928 929 930
    An operator integrating the open source Warp-CTC library (https://github.com/baidu-research/warp-ctc)
    to compute Connectionist Temporal Classification (CTC) loss.
    It can be aliased as softmax with CTC, since a native softmax activation
931 932 933
    is interated to the Warp-CTC library to normalize values for each row of the input tensor.

    Parameters:
934
        log_probs (Tensor): The unscaled probability sequence with padding, which is a 3-D Tensor. The tensor shape is [max_logit_length, batch_size, num_classes + 1], where max_logit_length is the longest length of input logit sequence. The data type should be float32 or float64.
935 936 937 938 939 940 941 942
        labels (Tensor): The ground truth sequence with padding, which must be a 3-D Tensor. The tensor shape is [batch_size, max_label_length], where max_label_length is the longest length of label sequence. The data type must be int32.
        input_lengths (Tensor): The length for each input sequence, it should have shape [batch_size] and dtype int64.
        label_lengths (Tensor): The length for each label sequence, it should have shape [batch_size] and dtype int64.
        blank (int, optional): The blank label index of Connectionist Temporal Classification (CTC) loss, which is in the half-opened interval [0, num_classes + 1). The data type must be int32. Default is 0.
        reduction (string, optional): Indicate how to average the loss, the candicates are ``'none'`` | ``'mean'`` | ``'sum'``. If :attr:`reduction` is ``'mean'``, the output loss will be divided by the label_lengths, and then return the mean of quotient; If :attr:`reduction` is ``'sum'``, return the sum of loss; If :attr:`reduction` is ``'none'``, no reduction will be applied. Default is ``'mean'``.

    Returns:
        Tensor, The Connectionist Temporal Classification (CTC) loss between ``log_probs`` and  ``labels``. If attr:`reduction` is ``'none'``, the shape of loss is [batch_size], otherwise, the shape of loss is [1]. Data type is the same as ``log_probs``.
943

944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982
    Examples:

        .. code-block:: python

            # declarative mode
            import paddle.nn.functional as F
            import numpy as np
            import paddle

            # length of the longest logit sequence
            max_seq_length = 4
            #length of the longest label sequence
            max_label_length = 3
            # number of logit sequences
            batch_size = 2
            # class num
            class_num = 3

            np.random.seed(1)
            log_probs = np.array([[[4.17021990e-01, 7.20324516e-01, 1.14374816e-04],
                                    [3.02332580e-01, 1.46755889e-01, 9.23385918e-02]],

                                    [[1.86260208e-01, 3.45560730e-01, 3.96767467e-01],
                                    [5.38816750e-01, 4.19194520e-01, 6.85219526e-01]],

                                    [[2.04452246e-01, 8.78117442e-01, 2.73875929e-02],
                                    [6.70467496e-01, 4.17304814e-01, 5.58689833e-01]],

                                    [[1.40386939e-01, 1.98101491e-01, 8.00744593e-01],
                                    [9.68261600e-01, 3.13424170e-01, 6.92322612e-01]],

                                    [[8.76389146e-01, 8.94606650e-01, 8.50442126e-02],
                                    [3.90547849e-02, 1.69830427e-01, 8.78142476e-01]]]).astype("float32")
            labels = np.array([[1, 2, 2],
                            [1, 2, 2]]).astype("int32")
            input_lengths = np.array([5, 5]).astype("int64")
            label_lengths = np.array([3, 3]).astype("int64")

            paddle.disable_static()
983 984 985 986
            log_probs = paddle.to_tensor(log_probs)
            labels = paddle.to_tensor(labels)
            input_lengths = paddle.to_tensor(input_lengths)
            label_lengths = paddle.to_tensor(label_lengths)
987

988 989 990 991
            loss = F.ctc_loss(log_probs, labels,
                input_lengths,
                label_lengths,
                blank=0,
992 993 994
                reduction='none')
            print(loss.numpy())  #[3.9179852 2.9076521]

995 996 997 998 999
            loss = F.ctc_loss(log_probs, labels,
                input_lengths,
                label_lengths,
                blank=0,
                reduction='mean')
1000 1001 1002 1003 1004 1005 1006 1007 1008 1009
            print(loss.numpy())  #[1.1376063]

    """

    loss_out = fluid.layers.warpctc(log_probs, labels, blank, False,
                                    input_lengths, label_lengths)

    loss_out = fluid.layers.squeeze(loss_out, [-1])
    assert reduction in ['mean', 'sum', 'none']
    if reduction == 'mean':
S
ShenLiang 已提交
1010
        loss_out = paddle.mean(loss_out / label_lengths)
1011 1012 1013 1014 1015
    elif reduction == 'sum':
        loss_out = paddle.sum(loss_out)
    return loss_out


1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045
def cross_entropy(input,
                  label,
                  weight=None,
                  ignore_index=-100,
                  reduction='mean'):
    """
    This operator implements the cross entropy loss function. This OP combines ``LogSoftmax``,
    and ``NLLLoss`` together.

    It is useful when training a classification problem with ``C`` classes.
    If provided, the optional argument ``weight`` should be a 1D Variable assigning
    weight to each of the classes.

    For predictions label, and target label, the loss is calculated as follows.

    .. math::

        loss_j =  -\\text{input[class]} +
        \\log\\left(\\sum_{i=0}^{K}\\exp(\\text{input}_i)\\right), j = 1,..., K

    If weight is not ``None``:

    .. math::

        loss_j =  \\text{weight[class]}(-\\text{input[class]} +
        \\log\\left(\\sum_{i=0}^{K}\\exp(\\text{input}_i)\\right)), j = 1,..., K

    Parameters:
        input (Tensor): Input tensor, the data type is float32, float64. Shape is
	    (N, C), where C is number of classes, and if shape is more than 2D, this
1046 1047
	    is (N, C, D1, D2,..., Dk), k >= 1.
        label (Tensor): Label tensor, the data type is int64. Shape is (N), where each
1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079
	    value is 0 <= label[i] <= C-1, and if shape is more than 2D, this is
	    (N, D1, D2,..., Dk), k >= 1.
        weight (Tensor, optional): Weight tensor, a manual rescaling weight given
            to each class and the shape is (C). It has the same dimensions as class
	    number and the data type is float32, float64. Default is ``'None'``.
        reduction (str, optional): Indicate how to average the loss by batch_size,
            the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
            If :attr:`reduction` is ``'mean'``, the reduced mean loss is returned;
            If :attr:`size_average` is ``'sum'``, the reduced sum loss is returned.
            If :attr:`reduction` is ``'none'``, the unreduced loss is returned.
            Default is ``'mean'``.
        ignore_index (int64, optional): Specifies a target value that is ignored
            and does not contribute to the input gradient. Default is ``-100``.

    Returns:
        The tensor variable storing the cross_entropy_loss of input and label.

    Return type: Tensor.

    Examples:
        .. code-block:: python

            import paddle
            paddle.disable_static()
            input_data = np.random.random([5, 100]).astype("float64")
            label_data = np.random.randint(0, 100, size=(5)).astype(np.int64)
            weight_data = np.random.random([100]).astype("float64")
            input =  paddle.to_tensor(input_data)
            label =  paddle.to_tensor(label_data)
            weight = paddle.to_tensor(weight_data)
            loss = paddle.nn.functional.cross_entropy(input=input, label=label, weight=weight)
            print(loss.numpy())
1080

1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093
    """
    if not in_dygraph_mode():
        fluid.data_feeder.check_variable_and_dtype(
            input, 'input', ['float32', 'float64'], 'cross_entropy_loss')
        fluid.data_feeder.check_variable_and_dtype(label, 'label', ['int64'],
                                                   'cross_entropy_loss')

    if reduction not in ['sum', 'mean', 'none']:
        raise ValueError(
            "The value of 'reduction' in cross_entropy_loss should be 'sum', 'mean' or"
            " 'none', but received %s, which is not allowed." % reduction)

    #step 1. log_softmax
1094
    log_softmax_out = paddle.nn.functional.log_softmax(input, axis=1)
1095 1096 1097 1098
    if weight is not None and not isinstance(weight, Variable):
        raise ValueError(
            "The weight' is not a Variable, please convert to Variable.")

1099
    #step 2. nll_loss
1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142
    input = log_softmax_out
    helper = LayerHelper('nll_loss', **locals())
    dtype = helper.input_dtype(input)

    if not in_dygraph_mode():
        fluid.data_feeder.check_variable_and_dtype(
            input, 'input', ['float32', 'float64'], 'nll_loss')
        fluid.data_feeder.check_variable_and_dtype(label, 'label', ['int64'],
                                                   'nll_loss')

    x_shape = list(input.shape)
    n = x_shape[0]
    c = x_shape[1]
    x_dims = len(x_shape)
    if x_dims < 2:
        raise ValueError('Expected 2 or more dimensions (got {})'.format(
            x_dims))
    if x_dims != 2 and x_dims != 4:
        input = reshape(input, shape=[n, c, 1, -1])
        label = reshape(label, shape=[n, 1, -1])
        out_shape = [n] + x_shape[2:]

    if not in_dygraph_mode():
        fluid.data_feeder.check_variable_and_dtype(
            input, 'input', ['float32', 'float64'], 'nll_loss')
        fluid.data_feeder.check_variable_and_dtype(label, 'label', ['int64'],
                                                   'nll_loss')
    inputs = {'X': input, 'Label': label}
    attrs = {'reduction': reduction, 'ignore_index': ignore_index}
    if weight is not None:
        if isinstance(weight, Variable):
            inputs['Weight'] = weight

    out = helper.create_variable_for_type_inference(dtype=input.dtype)
    total_weight = helper.create_variable_for_type_inference(dtype=input.dtype)
    outputs = {'Out': out, 'Total_weight': total_weight}

    helper.append_op(
        type='nll_loss', inputs=inputs, outputs=outputs, attrs=attrs)
    if x_dims != 2 and x_dims != 4 and reduction == 'none':
        out = reshape(out, shape=out_shape)

    return out