loss.py 61.7 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
from ...fluid.layers import dice_loss  #DEFINE_ALIAS
from ...fluid.layers import log_loss  #DEFINE_ALIAS
from ...fluid.layers import npair_loss  #DEFINE_ALIAS
29
from ...fluid.layers import reshape
30 31 32
from ...fluid.layers import softmax_with_cross_entropy  #DEFINE_ALIAS
from ...fluid.layers import square_error_cost  #DEFINE_ALIAS

33 34
from ...fluid.layers import edit_distance  #DEFINE_ALIAS
from ...fluid.layers import sampled_softmax_with_cross_entropy  #DEFINE_ALIAS
35
from ...fluid.layers import huber_loss
36
from ...fluid.layer_helper import LayerHelper
37
from ...fluid.framework import in_dygraph_mode
38
from ...fluid.framework import _varbase_creator
39
from ...fluid.framework import Variable
40

41
__all__ = [
42
    'binary_cross_entropy',
43
    'binary_cross_entropy_with_logits',
44 45
    'cross_entropy',
    'dice_loss',
46
    'hsigmoid_loss',
47
    'kl_div',
48
    'l1_loss',
49 50
    'log_loss',
    'mse_loss',
51
    'margin_ranking_loss',
52
    #       'nce',
53
    'nll_loss',
54 55
    'npair_loss',
    'sigmoid_focal_loss',
56
    'smooth_l1_loss',
57 58
    'softmax_with_cross_entropy',
    'square_error_cost',
59
    'ctc_loss',
60
]
61 62


63 64 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 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
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

123 124
            input = paddle.to_tensor([0.5, 0.6, 0.7], 'float32')
            label = paddle.to_tensor([1.0, 0.0, 1.0], 'float32')
125
            output = paddle.nn.functional.binary_cross_entropy(input, label)
N
Noel 已提交
126
            print(output)  # [0.65537095]
127 128 129 130 131 132 133 134 135

    """
    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():
136
        out = core.ops.bce_loss(input, label)
137 138 139 140 141 142 143
        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
144
            return core.ops.mean(out)
145 146 147 148 149 150 151 152 153
        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
154 155 156 157 158 159 160 161 162
    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]})
163 164

    if weight is not None:
165
        if isinstance(weight, paddle.static.Variable):
166
            weight_name = name if reduction is 'none' else None
167
            out = paddle.multiply(out, weight, name=weight_name)
168 169 170 171 172 173 174 175 176 177 178 179
        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


180 181 182 183 184 185
def binary_cross_entropy_with_logits(logit,
                                     label,
                                     weight=None,
                                     reduction='mean',
                                     pos_weight=None,
                                     name=None):
186
    r"""
187 188 189 190 191 192 193 194 195 196 197 198 199 200 201
    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))

N
Noel 已提交
202
    We know that :math:`\\sigma(Logit) = \\frac{1}{1 + e^{-Logit}}`. By substituting this we get:
203 204

    .. math::
N
Noel 已提交
205
           Out = Logit - Logit * Labels + \\log(1 + e^{-Logit})
206

N
Noel 已提交
207
    For stability and to prevent overflow of :math:`e^{-Logit}` when Logit < 0,
208 209 210
    we reformulate the loss as follows:

    .. math::
N
Noel 已提交
211
           Out = \\max(Logit, 0) - Logit * Labels + \\log(1 + e^{-\|Logit\|})
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

    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
N
Noel 已提交
256

257 258
            logit = paddle.to_tensor([5.0, 1.0, 3.0])
            label = paddle.to_tensor([1.0, 0.0, 1.0])
259
            output = paddle.nn.functional.binary_cross_entropy_with_logits(logit, label)
N
Noel 已提交
260
            print(output)  # [0.45618808]
261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299

    """
    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

300
    out = paddle.fluid.layers.sigmoid_cross_entropy_with_logits(
301 302
        logit, label, name=sigmoid_name)

303 304
    one = paddle.fluid.layers.fill_constant(
        shape=[1], value=1.0, dtype=logit.dtype)
305 306 307 308 309
    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(
310 311
            paddle.multiply(
                label, paddle.fluid.layers.elementwise_sub(pos_weight, one)),
312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329
            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


330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461
def hsigmoid_loss(input,
                  label,
                  num_classes,
                  weight,
                  bias=None,
                  path_table=None,
                  path_code=None,
                  is_sparse=False,
                  name=None):
    """
    The hierarchical sigmoid organizes the classes into a complete binary tree to reduce the computational complexity
    and speed up the model training, especially the training of language model.
    Each leaf node of the complete binary tree represents a class(word) and each non-leaf node acts as a binary classifier.
    For each class(word), there's a unique path from root to itself, hsigmoid calculate the cost for each non-leaf node on
    the path, and sum them to get a total cost.
    Comparing to softmax, the OP can reduce the computational complexity from :math:`O(N)` to :math:`O(logN)`, where :math:`N`
    represents the number of classes or the size of word dict.

    The OP supports default tree and custom tree. For the default tree, you can refer to `Hierarchical Probabilistic Neural
    Network Language Model <http://www.iro.umontreal.ca/~lisa/pointeurs/hierarchical-nnlm-aistats05.pdf>`_. For the custom
    tree, you need to set :attr:`is_custom` to True, and do the following steps (take the language model as an example):

    1. Using a custom word dict to build a binary tree, each leaf node should be an word in the word dict.
    2. Creating a dict map word_id -> path that from the word to the root node, we call it path_table.
    3. Creating a dict map word_id -> code of path that from the word to the root node, we call it path_code.
       Code means the label of each binary classifier, 1 indicate true, 0 indicate false.
    4. Now, each word should has its path and code along the path, you can pass a batch of path and code related
       to the same batch of inputs.

    Parameters:
        input (Tensor): A tensor with the shape [N, D], where N is the size of mini-batch,
            and D is the feature size. Its data type supports float32 or float64.
        label (Tensor): A tensor contains the labels of training data. Its shape is [N, 1]
            and data type is int64.
        num_classes (int): The number of classes or the size of word dict, must be greater than 2.
            If the default tree is used (path_code and path_table is None are None), `num_classes`
            should not be None. If the custom tree is used (path_code and path_table is None are not None),
            `num_classes` should be the number of non-leaf nodes, which indicates the num of
            classes using by the binary classifier.
        weight (Tensor): A tensor with shape (num_classes - 1, D), with the same data type as `input`.
        bias (Tensor, optional): A tensor with shape (num_classes - 1, 1), with the same data type as `input`.
            If `bias` is None, no bias will be add. Default is None.
        path_table (Tensor, optional): A tensor that stores each batch of samples' path from leaf to root
            node, its shape is [N, L] and data type is int64, where L is the length of path. For each sample i,
            path_table[i] is a np.array like structure and each element in this array is the indexes in parent
            nodes' weight matrix. If `path_table` and `path_code` are None, the default tree will be used.
            Default is None.
        path_code (Tensor, optional): A tensor that stores each batch of samples' code of path from leaf
            to root node, its shape is [N, L] and data type is int64, which is the same as :attr:`path_table`.
            Each code of path is consisted with the code of nodes from leaf to root node. If `path_table` and
            `path_code` are None, the default tree will be used. Default is None.
        is_sparse (bool, optional): Whether use sparse updating instead of dense updating. If `is_sparse` is True,
            the gradient of `weight` and `input` will be sparse. Default is False.
        name (str, optional): Name for the operation (optional, default is None).
            For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        A tensor with the cost of hierarchical sigmoid, its shape is [N, 1] and data type is the same as `input`.

    Examples:
        .. code-block:: python

            import paddle
            import paddle.nn.functional as F

            paddle.set_device('cpu')

            input = paddle.uniform([2, 3])
            # [[-0.8018668   0.8736385  -0.9064771 ] # random
            #  [-0.10228515 -0.87188244 -0.8783718 ]] # random
            label = paddle.to_tensor([0, 1, 4, 5])
            num_classes = 5
            weight=paddle.uniform([num_classes-1, 3])
            # [[-0.24148715  0.8449961  -0.7399121 ] # random
            #  [-0.9800559   0.43509364  0.9091208 ] # random
            #  [ 0.60194826  0.10430074 -0.4521166 ] # random
            #  [-0.4469818  -0.01536179 -0.604454  ]] # random

            out=F.hsigmoid_loss(input, label, num_classes, weight)
            # [[3.0159328]
            #  [2.2407534]]
    """

    if in_dygraph_mode():
        out, _, _ = core.ops.hierarchical_sigmoid(
            input, weight, label, path_table, path_code, bias, 'num_classes',
            num_classes, 'is_sparse', is_sparse, 'remote_prefetch', is_sparse)
        return out

    check_variable_and_dtype(input, 'input', ['float32', 'float64'],
                             'hsigmoid_loss')
    check_variable_and_dtype(label, 'label', ['int64'], 'hsigmoid_loss')
    check_variable_and_dtype(weight, 'weight', ['float32', 'float64'],
                             'hsigmoid_loss')
    if bias is not None:
        check_variable_and_dtype(bias, 'bias', ['float32', 'float64'],
                                 'hsigmoid_loss')
    if path_table is not None:
        check_variable_and_dtype(path_table, 'path_table', ['int64'],
                                 'hsigmoid_loss')
    if path_code is not None:
        check_variable_and_dtype(path_code, 'path_code', ['int64'],
                                 'hsigmoid_loss')

    attrs = {
        "num_classes": num_classes,
        "is_sparse": is_sparse,
        "remote_prefetch": is_sparse
    }

    inputs = {
        "X": input,
        "W": weight,
        "Bias": bias,
        "PathTable": path_table,
        "PathCode": path_code,
        "Label": label
    }

    helper = LayerHelper('hsigmoid_loss', **locals())
    out = helper.create_variable_for_type_inference(input.dtype)
    pre_out = helper.create_variable_for_type_inference(input.dtype)
    outputs = {"Out": out, "PreOut": pre_out, "W_Out": weight}

    helper.append_op(
        type="hierarchical_sigmoid",
        inputs=inputs,
        outputs=outputs,
        attrs=attrs)
    return out


462
def smooth_l1_loss(input, label, reduction='mean', delta=1.0, name=None):
463
    r"""
464 465 466 467 468 469 470
    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::

G
Guanghua Yu 已提交
471
         loss(x,y) = \\frac{1}{n}\\sum_{i}z_i
472 473 474 475 476 477


    where z_i is given by:

    .. math::

G
Guanghua Yu 已提交
478
         \\mathop{z_i} = \\left\\{\\begin{array}{rcl}
479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494
        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'``.
495
        delta (float, optional): Specifies the hyperparameter delta to be used.
496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517
            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

            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)
G
Guanghua Yu 已提交
518
            print(output)
519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538
    """
    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)


539 540
def margin_ranking_loss(input,
                        other,
541
                        label,
542 543 544
                        margin=0.0,
                        reduction='mean',
                        name=None):
545
    r"""
546

547
    This op the calcluate the the margin rank loss between the input, other and label, use the math function as follows.
548

549
    .. math::
550
        margin\_rank\_loss = max(0, -label * (input - other) + margin)
551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566

    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.
567
        label(Tensor): the label value corresponding to input, it's data type should be float32, float64.
568 569 570 571 572 573 574 575 576 577
        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

578 579
            import paddle

Z
Zhong Hui 已提交
580 581 582
            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')
583
            loss = paddle.nn.functional.margin_ranking_loss(input, other, label)
N
Noel 已提交
584
            print(loss) # [0.75]
585
    """
586 587 588 589
    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)
590 591
    if fluid.framework.in_dygraph_mode():
        out = core.ops.elementwise_sub(other, input)
592
        out = core.ops.elementwise_mul(out, label)
593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608
        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(
609
        label, 'label', ['float32', 'float64'], 'margin_rank_loss')
610

611
    out = paddle.fluid.layers.elementwise_sub(other, input)
612
    out = paddle.multiply(out, label)
613 614 615

    if margin != 0.0:
        margin_var = out.block.create_var(dtype=out.dtype)
616 617
        paddle.fluid.layers.fill_constant(
            [1], out.dtype, margin, out=margin_var)
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
        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


645
def l1_loss(input, label, reduction='mean', name=None):
646
    r"""
647
    This operator computes the L1 Loss of Tensor ``input`` and ``label`` as follows.
648

649
    If `reduction` set to ``'none'``, the loss is:
650 651

    .. math::
N
Noel 已提交
652
        Out = \\lvert input - label \\rvert
653

654
    If `reduction` set to ``'mean'``, the loss is:
655 656

    .. math::
N
Noel 已提交
657
        Out = MEAN(\\lvert input - label \\rvert)
658

659
    If `reduction` set to ``'sum'``, the loss is:
660 661

    .. math::
N
Noel 已提交
662
        Out = SUM(\\lvert input - label\\rvert)
663

664

665
    Parameters:
N
Noel 已提交
666 667
        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.
668
        reduction (str, optional): Indicate the reduction to apply to the loss,
669
            the candicates are ``'none'`` | ``'mean'`` | ``'sum'``.
670 671 672
            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.
673 674
            Default is ``'mean'``.
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.
N
Noel 已提交
675

676
    Returns:
677 678 679
        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].
N
Noel 已提交
680

681 682
    Examples:
        .. code-block:: python
N
Noel 已提交
683

684
            import paddle
685

686
            paddle.disable_static()
687 688
            input = paddle.to_tensor([[1.5, 0.8], [0.2, 1.3]])
            label = paddle.to_tensor([[1.7, 1], [0.4, 0.5]])
689

690
            l1_loss = paddle.nn.functional.l1_loss(input, label)
691
            print(l1_loss.numpy())
692 693
            # [0.35]

694
            l1_loss = paddle.nn.functional.l1_loss(input, label, reduction='none')
695
            print(l1_loss.numpy())
696 697 698
            # [[0.20000005 0.19999999]
            # [0.2        0.79999995]]

699
            l1_loss = paddle.nn.functional.l1_loss(input, label, reduction='sum')
700
            print(l1_loss.numpy())
701 702 703 704 705 706 707 708 709
            # [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(
710
            input, label, axis=-1, act='abs', op_name='elementwise_sub')
711 712 713 714 715 716 717 718 719
        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(
720
        input, 'input', ['float32', 'float64', 'int32', 'int64'], 'l1_loss')
721 722 723 724
    fluid.data_feeder.check_variable_and_dtype(
        label, 'label', ['float32', 'float64', 'int32', 'int64'], 'l1_loss')

    if reduction == 'sum':
725
        unreduced = paddle.fluid.layers.elementwise_sub(input, label, act='abs')
726 727
        return paddle.sum(unreduced, name=name)
    elif reduction == 'mean':
728
        unreduced = paddle.fluid.layers.elementwise_sub(input, label, act='abs')
729 730
        return paddle.mean(unreduced, name=name)
    else:
731 732
        return paddle.fluid.layers.elementwise_sub(
            input, label, act='abs', name=name)
733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782


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)

Z
Zhong Hui 已提交
783
                input = paddle.to_tensor(input_np)
784
                log_out = log_softmax(input)
Z
Zhong Hui 已提交
785
                label = paddle.to_tensor(label_np)
786
                result = nll_loss(log_out, label)
787
                print(result) # [1.0720209]
788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839
    """
    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
840 841


842
def kl_div(input, label, reduction='mean', name=None):
843
    r"""
844 845 846 847 848 849 850 851 852 853 854
    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
855
    the same shape as input, loss in each point is calculated
856
    seperately and no reduction is applied.
857

858 859
    While :attr:`reduction` is :attr:`mean`, output loss is in
    shape of [1] and loss value is the mean value of all losses.
860

861 862
    While :attr:`reduction` is :attr:`sum`, output loss is in
    shape of [1] and loss value is the sum value of all losses.
863 864

    While :attr:`reduction` is :attr:`batchmean`, output loss is
865 866 867 868
    in shape of [1] and loss value is the sum value of all losses
    divided by batch size.

    Args:
869
        input (Tensor): The input tensor. The shapes is [N, *], where N is batch size and `*` means
870 871 872 873 874 875 876 877 878
             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'``.
879
        name(str, optional): Name for the operation (optional, default is None). For more information,
880 881 882 883 884 885 886 887 888 889 890
            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
891

892 893 894 895
            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
896
            # 'batchmean' reduction, loss shape will be [1]
897 898
            pred_loss = F.kl_div(paddle.to_tensor(input),
                                 paddle.to_tensor(target), reduction='batchmean')
L
LielinJiang 已提交
899
            # shape=[1]
900

901
            # 'mean' reduction, loss shape will be [1]
902 903
            pred_loss = F.kl_div(paddle.to_tensor(input),
                                 paddle.to_tensor(target), reduction='mean')
904 905 906
            # shape=[1]

            # 'sum' reduction, loss shape will be [1]
907 908
            pred_loss = F.kl_div(paddle.to_tensor(input),
                                 paddle.to_tensor(target), reduction='sum')
909 910 911
            # shape=[1]

            # 'none' reduction, loss shape is same with input shape
912 913
            pred_loss = F.kl_div(paddle.to_tensor(input),
                                 paddle.to_tensor(target), reduction='none')
914 915 916
            # shape=[5, 20]

    """
L
LielinJiang 已提交
917 918 919 920 921 922 923 924 925 926
    # ugly type promotion
    if fluid.data_feeder.convert_dtype(
            input.dtype) == 'float32' and fluid.data_feeder.convert_dtype(
                label.dtype) == 'float64':
        input = fluid.layers.cast(input, 'float64')
    elif fluid.data_feeder.convert_dtype(
            input.dtype) == 'float64' and fluid.data_feeder.convert_dtype(
                label.dtype) == 'float32':
        label = fluid.layers.cast(label, 'float64')

927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948
    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


949
def mse_loss(input, label, reduction='mean', name=None):
950
    r"""
951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983
    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.
984

985 986 987
    Examples:

        .. code-block:: python
988

989 990
            import paddle
            mse_loss = paddle.nn.loss.MSELoss()
991 992
            input = paddle.to_tensor(1.5)
            label = paddle.to_tensor(1.7)
993
            output = mse_loss(input, label)
B
Bai Yifan 已提交
994
            print(output)
995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021
            # [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)
1022 1023


1024 1025 1026 1027 1028 1029 1030 1031
def ctc_loss(log_probs,
             labels,
             input_lengths,
             label_lengths,
             blank=0,
             reduction='mean'):
    """

1032 1033 1034
    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
1035 1036 1037
    is interated to the Warp-CTC library to normalize values for each row of the input tensor.

    Parameters:
1038
        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.
1039 1040 1041 1042 1043 1044 1045 1046
        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``.
1047

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 1080 1081 1082 1083 1084 1085
    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")

1086 1087 1088 1089
            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)
1090

1091 1092 1093 1094
            loss = F.ctc_loss(log_probs, labels,
                input_lengths,
                label_lengths,
                blank=0,
1095
                reduction='none')
1096
            print(loss)  #[3.9179852 2.9076521]
1097

1098 1099 1100 1101 1102
            loss = F.ctc_loss(log_probs, labels,
                input_lengths,
                label_lengths,
                blank=0,
                reduction='mean')
1103
            print(loss)  #[1.1376063]
1104 1105 1106 1107 1108 1109 1110 1111 1112

    """

    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
1113
        loss_out = paddle.mean(loss_out / label_lengths)
1114 1115 1116 1117 1118
    elif reduction == 'sum':
        loss_out = paddle.sum(loss_out)
    return loss_out


1119 1120 1121 1122
def cross_entropy(input,
                  label,
                  weight=None,
                  ignore_index=-100,
1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133
                  reduction='mean',
                  soft_label=False,
                  axis=-1,
                  name=None):
    """
    This operator implements the cross entropy loss function with softmax. This function 
    combines the calculation of the softmax operation and the cross entropy loss function 
    to provide a more numerically stable gradient.
    Because this operator performs a softmax on logits internally, it expects
    unscaled logits. This operator should not be used with the output of
    softmax operator since that would produce incorrect results.
1134

1135 1136 1137 1138
    When the attribute :attr:`soft_label` is set :attr:`False`, this operators 
    expects mutually exclusive hard labels, each sample in a batch is in exactly 
    one class with a probability of 1.0. Each sample in the batch will have a 
    single label.
1139

1140 1141 1142
    The equation is as follows:

    1) Hard label (one-hot label, so every sample has exactly one class)
1143 1144 1145

    .. math::

1146 1147
        loss_j =  -\\text{logits}_{label_j} +
        \\log\\left(\\sum_{i=0}^{K}\\exp(\\text{logits}_i)\\right), j = 1,..., K
1148

1149
    2) Soft label (each sample can have a distribution over all classes)
1150 1151 1152

    .. math::

1153 1154 1155 1156 1157 1158 1159
        loss_j =  -\\sum_{i=0}^{K}\\text{label}_i
        \\left(\\text{logits}_i - \\log\\left(\\sum_{i=0}^{K}
        \\exp(\\text{logits}_i)\\right)\\right), j = 1,...,K

 
    It is useful when training a classification problem with ``C`` classes.

1160 1161 1162 1163

    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
1164
	    is (N, D1, D2,..., Dk, C), k >= 1.
1165
        label (Tensor): Label tensor, the data type is int64. Shape is (N), where each
1166 1167
	    value is 0 <= label[i] <= C-1, and if shape is more than 2D, this is
	    (N, D1, D2,..., Dk), k >= 1.
1168 1169 1170
        weight (Tensor, optional):a manual rescaling weight given to each class. 
            If given, has to be a Tensor of size C and the data type is float32, float64. 
            Default is ``'None'``.
1171 1172 1173 1174 1175 1176 1177 1178
        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``.
1179 1180 1181 1182 1183 1184
        soft_label (bool): indicate whether label is soft. Default False, meaning that
                the label is hard. If soft_label=True, the label is soft.
        axis (int, optional): The index of dimension to perform softmax calculations. It 
                              should be in range :math:`[-1, rank - 1]`, while :math:`rank`
                              is the rank of input :attr:`logits`. Default: -1.

1185 1186

    Returns:
1187
        Tensor.The tensor storing the cross_entropy_loss of input and label.
1188 1189 1190 1191


    Examples:
        .. code-block:: python
1192

1193
            import paddle
1194 1195 1196 1197 1198 1199 1200 1201 1202
            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)
            # [4.28546723]
1203 1204 1205 1206
    """

    if reduction not in ['sum', 'mean', 'none']:
        raise ValueError(
1207 1208 1209 1210 1211 1212
            "The value of 'reduction' in softmax_cross_entropy"
            "should be 'sum', 'mean' or 'none', but received %s, which is not allowed."
            % reduction)
    input_dims = len(list(input.shape))
    label_dims = len(list(label.shape))
    if input_dims - 1 != label_dims and input_dims != label_dims:
1213
        raise ValueError(
1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230
            'Expected nput_dims - 1 = label_dims or input_dims == label_dims\
             (got nput_dims{}, label_dims{})'.format(input_dims, label_dims))
    if input_dims - 1 == label_dims:
        label = paddle.unsqueeze(label, axis=axis)
    if in_dygraph_mode():
        out = softmax_with_cross_entropy(
            input,
            label,
            soft_label=soft_label,
            ignore_index=ignore_index,
            axis=axis)
        if weight is not None:
            weight_gather = core.ops.gather_nd(weight, label)  #trans to sample
            input_shape = list(label.shape)
            weight_gather_reshape, _ = core.ops.reshape2(weight_gather, 'shape',
                                                         input_shape)
            out = core.ops.elementwise_mul(out, weight_gather_reshape)
1231

1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243
        if reduction == "sum":
            return core.ops.reduce_sum(out, 'reduce_all', True)
        elif reduction == "mean":
            if weight is not None:
                out_sum = core.ops.reduce_sum(out, 'reduce_all', True)
                total_weight = core.ops.reduce_sum(weight_gather_reshape,
                                                   'reduce_all', True)
                return out_sum / total_weight
            else:
                return core.ops.mean(out)
        else:
            return out
1244

1245 1246 1247 1248 1249 1250 1251 1252 1253 1254
    fluid.data_feeder.check_variable_and_dtype(
        input, 'input', ['float32', 'float64'], 'softmax_cross_entropy')
    fluid.data_feeder.check_variable_and_dtype(
        label, 'label', ['int32', 'int64'], 'softmax_cross_entropy')
    out = softmax_with_cross_entropy(
        input,
        label,
        soft_label=soft_label,
        ignore_index=ignore_index,
        axis=axis)
1255
    if weight is not None:
1256 1257 1258 1259 1260 1261 1262
        fluid.data_feeder.check_variable_and_dtype(
            weight, 'weight', ['float32', 'float64'], 'softmax_cross_entropy')
        weight_name = name if reduction == 'none' else None
        weight_gather = paddle.gather_nd(weight, label)  #trans to sample
        input_shape = list(label.shape)
        weight_gather_reshape = reshape(weight_gather, shape=input_shape)
        out = paddle.multiply(out, weight_gather_reshape, name=weight_name)
1263

1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274
    if reduction == "sum":
        return paddle.sum(out, name=name)
    elif reduction == "mean":
        if weight is not None:
            out_sum = paddle.sum(out, name=name)
            total_weight = paddle.sum(weight_gather_reshape)
            return out_sum / total_weight
        else:
            return paddle.mean(out, name=name)
    else:
        return out
1275 1276 1277 1278 1279 1280 1281 1282 1283


def sigmoid_focal_loss(logit,
                       label,
                       normalizer=None,
                       alpha=0.25,
                       gamma=2.0,
                       reduction='sum',
                       name=None):
1284
    r"""
1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347
    `Focal Loss <https://arxiv.org/abs/1708.02002>`_ is proposed to address the
    foreground-background class imbalance for classification tasks. It down-weights
    easily-classified examples and thus focuses training on hard examples. For example,
    it is used in one-stage object detection where the foreground-background class
    imbalance is extremely high.

    This operator measures focal loss function as follows: 

    .. math::
           Out = -Labels * alpha * {(1 - \\sigma(Logit))}^{gamma}\\log(\\sigma(Logit)) - (1 - Labels) * (1 - alpha) * {\\sigma(Logit)}^{gamma}\\log(1 - \\sigma(Logit))

    We know that :math:`\\sigma(Logit) = \\frac{1}{1 + \\exp(-Logit)}`. 

    Then, if :attr:`normalizer` is not None, this operator divides the
    normalizer tensor on the loss `Out`:

    .. math::
           Out = \\frac{Out}{normalizer}

    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 ``label`` is 0 for the negative class and is 1 for the positive class.

    Args:
        logit (Tensor): The input logit tensor. The shape is [N, *], where N is batch_size,
            `*` means any number of additional dimensions. The ``logit`` is usually the
            output of a convolution layer. Available dtype is float32, float64.
        label (Tensor): The target label tensor with the same shape as
            ``logit``. The target label whose value should be numbers between 0 and 1.
            Available dtype is float32, float64.
        normalizer (Tensor, optional): The number normalizes the focal loss. It has to be
            a 1-D Tensor whose shape is `[1, ]`. The data type is float32, float64.
            For object detection task, it is the the number of positive samples.
            If set to None, the focal loss will not be normalized. Default is None.
        alpha(int|float, optional): Hyper-parameter to balance the positive and negative example,
            it should be between 0 and 1.  Default value is set to 0.25. 
        gamma(int|float, optional): Hyper-parameter to modulate the easy and hard examples.
            Default value is set to 2.0.
        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 ``'sum'``.
        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 ``logit``. The same dtype as ``logit`` tensor.

    Examples:

        .. code-block:: python

            import paddle

            logit = paddle.to_tensor([[0.97, 0.91, 0.03], [0.55, 0.43, 0.71]], dtype='float32')
            label = paddle.to_tensor([[1.0, 0.0, 0.0], [0.0, 1.0, 0.0]], dtype='float32')
            one = paddle.to_tensor([1.], dtype='float32')
            fg_label = paddle.greater_equal(label, one)
1348
            fg_num = paddle.sum(paddle.cast(fg_label, dtype='float32'))
1349
            output = paddle.nn.functional.sigmoid_focal_loss(logit, label, normalizer=fg_num)
1350
            print(output)  # [0.65782464]
1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436

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

    if normalizer is not None:
        fluid.data_feeder.check_variable_and_dtype(normalizer, 'normalizer',
                                                   ['float32', 'float64'],
                                                   'sigmoid_focal_loss')
        normalizer_shape = list(normalizer.shape)
        normalizer_dims = len(normalizer_shape)
        if normalizer_dims > 1:
            raise ValueError(
                "Expected one dimension of normalizer in sigmoid_focal_loss but got {}.".
                format(normalizer_dims))

    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',
                               logit.shape)
        loss = core.ops.sigmoid_cross_entropy_with_logits(logit, label)
        pred = core.ops.sigmoid(logit)
        p_t = core.ops.elementwise_add(
            core.ops.elementwise_mul(pred, label),
            core.ops.elementwise_mul(
                core.ops.elementwise_sub(one, pred),
                core.ops.elementwise_sub(one, label)))

        alpha = fluid.dygraph.base.to_variable([alpha], dtype=loss.dtype)
        alpha_t = core.ops.elementwise_add(
            core.ops.elementwise_mul(alpha, label),
            core.ops.elementwise_mul(
                core.ops.elementwise_sub(one, alpha),
                core.ops.elementwise_sub(one, label)))
        loss = core.ops.elementwise_mul(alpha_t, loss)

        gamma = fluid.dygraph.base.to_variable([gamma], dtype=loss.dtype)
        gamma_t = core.ops.elementwise_pow(
            core.ops.elementwise_sub(one, p_t), gamma)
        loss = core.ops.elementwise_mul(gamma_t, loss)

        if normalizer is not None:
            loss = core.ops.elementwise_div(loss, normalizer)

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

        return loss

    fluid.data_feeder.check_variable_and_dtype(
        logit, 'logit', ['float32', 'float64'], 'sigmoid_focal_loss')
    fluid.data_feeder.check_variable_and_dtype(
        label, 'label', ['float32', 'float64'], 'sigmoid_focal_loss')

    bce_name = None
    if reduction == 'none' and normalizer is None:
        bce_name = name
    loss = paddle.nn.functional.binary_cross_entropy_with_logits(
        logit, label, reduction='none', name=bce_name)

    pred = fluid.layers.sigmoid(logit)
    p_t = pred * label + (1 - pred) * (1 - label)

    alpha_t = alpha * label + (1 - alpha) * (1 - label)
    loss = paddle.multiply(alpha_t, loss)

    gamma_t = paddle.pow((1 - p_t), gamma)
    loss = paddle.multiply(gamma_t, loss)

    if normalizer is not None:
        normalizer_name = name if reduction == 'none' else None
        loss = paddle.divide(loss, normalizer, name=normalizer_name)

    if reduction == 'mean':
        loss = paddle.mean(loss, name=name)
    elif reduction == 'sum':
        loss = paddle.sum(loss, name=name)

    return loss