norm.py 21.3 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
# TODO: define normalization api
16 17 18 19 20
import paddle
import paddle.fluid as fluid
from ...fluid.data_feeder import check_variable_and_dtype, check_type
from ...fluid.layer_helper import LayerHelper
from ...fluid.framework import in_dygraph_mode, core
21
from ...framework import create_parameter
22 23
from ..initializer import Constant
from ...framework import ParamAttr
24
from ...fluid import core, dygraph_utils
25
import numbers
W
wanghuancoder 已提交
26
from paddle import _C_ops
27

28 29
__all__ = []

30 31

def normalize(x, p=2, axis=1, epsilon=1e-12, name=None):
32
    r"""
33 34 35 36
    This op normalizes ``x`` along dimension ``axis`` using :math:`L_p` norm. This layer computes

    .. math::

37
        y = \frac{x}{ \max\left( \lvert \lvert x \rvert \rvert_p, epsilon\right) }
38

39
    .. math::
40
        \lvert \lvert x \rvert \rvert_p = \left( \sum_i {\lvert x_i \rvert^p}  \right)^{1/p}
41

42
    where, :math:`\sum_i{\lvert x_i \rvert^p}` is calculated along the ``axis`` dimension.
43 44


N
Noel 已提交
45
    Parameters:
46 47
        x (Tensor): The input tensor could be N-D tensor, and the input data type could be float32 or float64.
        p (float|int, optional): The exponent value in the norm formulation. Default: 2
48
        axis (int, optional): The axis on which to apply normalization. If `axis < 0`, the dimension to normalization is `x.ndim + axis`. -1 is the last dimension.
49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64
        epsilon (float, optional): Small float added to denominator to avoid dividing by zero. Default is 1e-12.
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Tensor, the output has the same shape and data type with ``x``.

    Examples:

        .. code-block:: python

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

            paddle.disable_static()
            x = np.arange(6, dtype=np.float32).reshape(2,3)
65
            x = paddle.to_tensor(x)
66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82
            y = F.normalize(x)
            print(y.numpy())
            # [[0.         0.4472136  0.8944272 ]
            # [0.42426404 0.5656854  0.7071067 ]]

            y = F.normalize(x, p=1.5)
            print(y.numpy())
            # [[0.         0.40862012 0.81724024]
            # [0.35684016 0.4757869  0.5947336 ]]

            y = F.normalize(x, axis=0)
            print(y.numpy())
            # [[0.         0.24253564 0.37139067]
            # [1.         0.97014254 0.9284767 ]]
    """
    if in_dygraph_mode():
        eps = fluid.dygraph.base.to_variable([epsilon], dtype=x.dtype)
W
wanghuancoder 已提交
83 84 85
        out = _C_ops.p_norm(x, 'axis', axis, 'porder',
                            float(p), 'keepdim', True, 'epsilon', epsilon)
        return x / _C_ops.elementwise_max(out, eps)
86 87 88

    check_type(p, 'p', (float, int), 'normalize')
    check_type(axis, 'axis', (int), 'normalize')
G
Guoxia Wang 已提交
89 90
    check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'],
                             'normalize')
91 92 93 94
    if len(x.shape) == 1 and axis != 0 and axis != -1:
        raise ValueError(
            "Axis must be 0 or -1 when x is a 1-D tensor, but received axis = {}".
            format(axis))
95 96 97 98 99 100 101 102 103 104 105 106

    attrs = {
        'axis': axis,
        'porder': float(p),
        'keepdim': True,
        'epsilon': epsilon,
    }
    helper = LayerHelper('p_norm', **locals())
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
        type='p_norm', inputs={'X': x}, outputs={'Out': out}, attrs=attrs)
    eps = out.block.create_var(dtype=out.dtype)
107
    paddle.fluid.layers.fill_constant([1], out.dtype, epsilon, out=eps)
108
    return paddle.divide(x, paddle.maximum(out, eps), name=name)
109 110 111 112 113 114 115 116 117 118 119


def batch_norm(x,
               running_mean,
               running_var,
               weight,
               bias,
               training=False,
               momentum=0.9,
               epsilon=1e-05,
               data_format="NCHW",
C
ceci3 已提交
120
               use_global_stats=None,
121 122 123 124
               name=None):
    """
    Applies Batch Normalization as described in the paper Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift .

C
cnn 已提交
125
    nn.functional.batch_norm is uesd for nn.BatchNorm1D, nn.BatchNorm2D, nn.BatchNorm3D. Please use above API for BatchNorm.
126

127 128 129 130 131
    Parameters:
        x(Tesnor): input value. It's data type should be float32, float64.
        running_mean(Tensor): running mean.
        running_var(Tensor): running variance.
        weight(Tensor): The weight tensor of batch_norm, can not be None.
132
        bias(Tensor): The bias tensor of batch_norm can not be None.
133 134 135
        epsilon(float, optional): The small value added to the variance to prevent division by zero. Default: 1e-5.
        momentum(float, optional): The value used for the moving_mean and moving_var computation. Default: 0.9.
        training(bool, optional): True means train mode which compute by batch data and track global mean and var during train period. False means inference mode which compute by global mean and var which calculated by train period. Defalut False.
F
Feiyu Chan 已提交
136
        data_format(str, optional): Specify the input data format, may be "NC", "NCL", "NCHW", "NCDHW", "NLC", "NHWC" or "NDHWC". Defalut "NCHW".
C
ceci3 已提交
137
        use_global_stats(bool|None, optional): Whether to use global mean and variance. If set to False, use the statistics of one mini-batch, if set to True, use the global statistics, if set to None, use global statistics in the test phase and use the statistics of one mini-batch in the training phase. Default: None.
138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160
        name(str, optional): Name for the BatchNorm, default is None. For more information, please refer to :ref:`api_guide_Name`..

    Returns:
        None

    Examples:
        .. code-block:: python

          import paddle
          import numpy as np

          x = np.random.seed(123)
          x = np.random.random(size=(2, 1, 2, 3)).astype('float32')
          running_mean = np.random.random(size=1).astype('float32')
          running_variance = np.random.random(size=1).astype('float32')
          weight_data = np.random.random(size=1).astype('float32')
          bias_data = np.random.random(size=1).astype('float32')
          x = paddle.to_tensor(x)
          rm = paddle.to_tensor(running_mean)
          rv = paddle.to_tensor(running_variance)
          w = paddle.to_tensor(weight_data)
          b = paddle.to_tensor(bias_data)
          batch_norm_out = paddle.nn.functional.batch_norm(x, rm, rv, w, b)
Z
zhang wenhui 已提交
161
          print(batch_norm_out)
162 163 164 165 166 167 168
    """
    assert len(x.shape) >= 2, "input dim must be larger than 1"

    # input ad out must share the memory
    mean_out = running_mean
    variance_out = running_var

F
Feiyu Chan 已提交
169
    true_data_format = ['NC', 'NCL', 'NCHW', 'NCDHW', 'NLC', 'NHWC', 'NDHWC']
170 171
    if data_format not in true_data_format:
        raise ValueError(
F
Feiyu Chan 已提交
172 173
            "data_format must be one of 'NC', 'NCL', 'NCHW', 'NCDHW', "
            "'NLC', 'NHWC', 'NDHWC' but receive {}".format(data_format))
174

F
Feiyu Chan 已提交
175
    data_format = 'NCHW' if data_format[1] == 'C' else 'NHWC'
176

C
ceci3 已提交
177 178 179 180 181 182
    if use_global_stats == None:
        use_global_stats = not training
        trainable_statistics = False
    else:
        trainable_statistics = not use_global_stats

183 184
    if in_dygraph_mode():
        # for dygraph need tuple
185 186 187 188
        attrs = ("momentum", momentum, "epsilon", epsilon, "is_test",
                 not training, "data_layout", data_format, "use_mkldnn", False,
                 "fuse_with_relu", False, "use_global_stats", use_global_stats,
                 "trainable_statistics", trainable_statistics)
W
wanghuancoder 已提交
189
        batch_norm_out, _, _, _, _, _ = _C_ops.batch_norm(
190 191 192 193 194 195 196 197 198 199 200 201
            x, weight, bias, running_mean, running_var, mean_out, variance_out,
            *attrs)
        return dygraph_utils._append_activation_in_dygraph(
            batch_norm_out, act=None)

    check_variable_and_dtype(x, 'input', ['float16', 'float32', 'float64'],
                             'BatchNorm')

    # for static need dict
    attrs = {
        "momentum": momentum,
        "epsilon": epsilon,
202
        "is_test": not training,
203 204 205 206
        "data_layout": data_format,
        "use_mkldnn": False,
        "fuse_with_relu": False,
        "use_global_stats": use_global_stats,
C
ceci3 已提交
207
        "trainable_statistics": trainable_statistics,
208 209 210 211 212 213 214 215 216 217 218 219
    }

    inputs = {
        "X": [x],
        "Scale": [weight],
        "Bias": [bias],
        "Mean": [running_mean],
        "Variance": [running_var]
    }

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

220
    param_dtype = x.dtype if x.dtype is not 'float16' else 'float32'
221
    saved_mean = helper.create_variable_for_type_inference(
222
        dtype=param_dtype, stop_gradient=True)
223
    saved_variance = helper.create_variable_for_type_inference(
224 225
        dtype=param_dtype, stop_gradient=True)
    batch_norm_out = helper.create_variable_for_type_inference(x.dtype)
226 227 228 229 230 231

    outputs = {
        "Y": [batch_norm_out],
        "MeanOut": [running_mean],
        "VarianceOut": [running_var],
        "SavedMean": [saved_mean],
232
        "SavedVariance": [saved_variance]
233 234
    }

235 236 237 238 239 240
    if training or trainable_statistics:
        # reserve_space is only used for training.
        reserve_space = helper.create_variable_for_type_inference(
            dtype=x.dtype, stop_gradient=True)
        outputs["ReserveSpace"] = [reserve_space]

241 242 243 244 245 246 247 248 249 250 251 252 253 254
    helper.append_op(
        type="batch_norm", inputs=inputs, outputs=outputs, attrs=attrs)

    return helper.append_activation(batch_norm_out)


def layer_norm(x,
               normalized_shape,
               weight=None,
               bias=None,
               epsilon=1e-05,
               name=None):
    """
    see more detail in paddle.nn.LayerNorm
255

256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279
    Parameters:
        x(Tensor): Input Tensor. It's data type should be float32, float64.
        normalized_shape(int|list|tuple): Input shape from an expected input of
            size :math:`[*, normalized_shape[0], normalized_shape[1], ..., normalized_shape[-1]]`.
            If it is a single integer, this module will normalize over the last dimension
            which is expected to be of that specific size.
        epsilon(float, optional): The small value added to the variance to prevent
            division by zero. Default: 1e-05.
        weight(Tensor, optional): The weight tensor of batch_norm. Default: None.
        bias(Tensor, optional): The bias tensor of batch_norm. Default: None.
        name(str, optional): Name for the LayerNorm, default is None. For more information, please refer to :ref:`api_guide_Name`..

    Returns:
        None

    Examples:

        .. code-block:: python

          import paddle
          import numpy as np

          np.random.seed(123)
          x_data = np.random.random(size=(2, 2, 2, 3)).astype('float32')
280
          x = paddle.to_tensor(x_data)
C
Chen Long 已提交
281
          layer_norm_out = paddle.nn.functional.layer_norm(x, x.shape[1:])
Z
zhang wenhui 已提交
282
          print(layer_norm_out)
283 284 285
    """
    input_shape = list(x.shape)
    input_ndim = len(input_shape)
286 287 288 289 290 291 292 293
    if isinstance(normalized_shape, numbers.Integral):
        normalized_shape = [normalized_shape]
    elif isinstance(normalized_shape, tuple):
        normalized_shape = list(normalized_shape)
    elif not isinstance(normalized_shape, list):
        raise ValueError(
            "`normalized_shape` should be int, list of ints or tuple of ints.")

294 295 296 297 298 299 300 301 302 303 304
    normalized_ndim = len(normalized_shape)
    begin_norm_axis = input_ndim - normalized_ndim
    if input_ndim < normalized_ndim or input_shape[
            begin_norm_axis:] != normalized_shape:
        str_normalized_shape = str(normalized_shape)
        raise ValueError('Given normalized_shape is ' + str_normalized_shape +
                         ', expected input with shape [*, ' +
                         str_normalized_shape[
                             1:] + ', but got input shape ' + str(input_shape))

    if in_dygraph_mode():
W
wanghuancoder 已提交
305 306
        pre_act, _, _ = _C_ops.layer_norm(x, weight, bias, 'epsilon', epsilon,
                                          'begin_norm_axis', begin_norm_axis)
307 308
        return dygraph_utils._append_activation_in_dygraph(pre_act, act=None)

F
furnace 已提交
309 310
    check_variable_and_dtype(x, 'input', ['float16', 'float32', 'float64'],
                             'LayerNorm')
311 312 313 314 315 316 317 318 319 320 321

    inputs = dict()
    inputs['X'] = [x]
    if weight:
        inputs['Scale'] = [weight]
    if bias:
        inputs['Bias'] = [bias]
    attrs = {"epsilon": epsilon, "begin_norm_axis": begin_norm_axis}

    # create output
    helper = LayerHelper('layer_norm', **locals())
F
furnace 已提交
322 323

    dtype = x.dtype
324
    mean_out = helper.create_variable_for_type_inference(
F
furnace 已提交
325
        dtype=dtype, stop_gradient=True)
326
    variance_out = helper.create_variable_for_type_inference(
F
furnace 已提交
327 328
        dtype=dtype, stop_gradient=True)
    layer_norm_out = helper.create_variable_for_type_inference(dtype)
329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354

    helper.append_op(
        type="layer_norm",
        inputs=inputs,
        outputs={
            "Y": layer_norm_out,
            "Mean": mean_out,
            "Variance": variance_out,
        },
        attrs={"epsilon": epsilon,
               "begin_norm_axis": begin_norm_axis})

    return helper.append_activation(layer_norm_out)


def instance_norm(x,
                  running_mean=None,
                  running_var=None,
                  weight=None,
                  bias=None,
                  use_input_stats=True,
                  momentum=0.9,
                  eps=1e-05,
                  data_format="NCHW",
                  name=None):
    """
C
cnn 已提交
355
    See more detail in nn.layer.InstanceNorm2D.
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

    Parameters:
        x(Tensor): Input Tensor. It's data type should be float32, float64.
        running_mean(Tensor): running mean. Default None.
        running_var(Tensor): running variance. Default None.
        weight(Tensor, optional): The weight tensor of instance_norm. Default: None.
        bias(Tensor, optional): The bias tensor of instance_norm. Default: None.
        eps(float, optional): A value added to the denominator for numerical stability. Default is 1e-5.
        momentum(float, optional): The value used for the moving_mean and moving_var computation. Default: 0.9.
        use_input_stats(bool): Default True.
        data_format(str, optional): Specify the input data format, may be "NC", "NCL", "NCHW" or "NCDHW". Defalut "NCHW".
        name(str, optional): Name for the InstanceNorm, default is None. For more information, please refer to :ref:`api_guide_Name`..

    Returns:
        None.

    Examples:

        .. code-block:: python

          import paddle
          import numpy as np

          np.random.seed(123)
          x_data = np.random.random(size=(2, 2, 2, 3)).astype('float32')
381
          x = paddle.to_tensor(x_data)
C
Chen Long 已提交
382
          instance_norm_out = paddle.nn.functional.instance_norm(x)
383

Z
zhang wenhui 已提交
384
          print(instance_norm_out)
385 386 387 388

    """

    if in_dygraph_mode():
W
wanghuancoder 已提交
389 390 391
        out, _, _ = _C_ops.instance_norm(x, weight, bias, "epsilon", eps,
                                         "momentum", momentum, "data_format",
                                         data_format)
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
        return out

    check_variable_and_dtype(x, 'input', ['float32', 'float64'], "InstanceNorm")

    attrs = {"epsilon": eps, "momentum": momentum, "data_format": data_format}

    if weight and bias:
        inputs = {"X": [x], "Scale": [weight], "Bias": [bias]}
    else:
        inputs = {"X": [x]}

    helper = LayerHelper('instance_norm', **locals())
    saved_mean = helper.create_variable_for_type_inference(
        dtype=x.dtype, stop_gradient=True)
    saved_variance = helper.create_variable_for_type_inference(
        dtype=x.dtype, stop_gradient=True)
    instance_norm_out = helper.create_variable_for_type_inference(x.dtype)

    outputs = {
        "Y": [instance_norm_out],
        "SavedMean": [saved_mean],
        "SavedVariance": [saved_variance]
    }

    helper.append_op(
        type="instance_norm", inputs=inputs, outputs=outputs, attrs=attrs)
    return instance_norm_out
419 420 421 422 423 424 425 426 427


def local_response_norm(x,
                        size,
                        alpha=1e-4,
                        beta=0.75,
                        k=1.,
                        data_format="NCHW",
                        name=None):
428
    r"""
429 430 431 432 433 434 435
        Local Response Normalization performs a type of "lateral inhibition" by normalizing over local input regions.
        For more information, please refer to `ImageNet Classification with Deep Convolutional Neural Networks <https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf>`_

        The formula is as follows:

        .. math::

436
            Output(i, x, y) = Input(i, x, y) / \left(k + \alpha \sum\limits^{\min(C-1, i + size/2)}_{j = \max(0, i - size/2)}(Input(j, x, y))^2\right)^{\beta}
437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490

        In the above equation:

        - :math:`size` : The number of channels to sum over.
        - :math:`k` : The offset (avoid being divided by 0).
        - :math:`\\alpha` : The scaling parameter.
        - :math:`\\beta` : The exponent parameter.


        Args:
            x (Tensor): The input 3-D/4-D/5-D tensor. The data type is float32.
            size (int): The number of channels to sum over.
            alpha (float, optional): The scaling parameter, positive. Default:1e-4
            beta (float, optional): The exponent, positive. Default:0.75
            k (float, optional): An offset, positive. Default: 1.0
            data_format (str, optional): Specify the data format of the input, and the data format of the output
                will be consistent with that of the input. An optional string from:
                If x is 3-D Tensor, the string could be `"NCL"` or `"NLC"` . When it is `"NCL"`,
                the data is stored in the order of: `[batch_size, input_channels, feature_length]`.
                If x is 4-D Tensor, the string could be  `"NCHW"`, `"NHWC"`. When it is `"NCHW"`,
                the data is stored in the order of: `[batch_size, input_channels, input_height, input_width]`.
                If x is 5-D Tensor, the string could be  `"NCDHW"`, `"NDHWC"` . When it is `"NCDHW"`,
                the data is stored in the order of: `[batch_size, input_channels, input_depth, input_height, input_width]`.
            name (str, optional): Name for the operation (optional, default is None). For more information,
                please refer to :ref:`api_guide_Name`.

        Returns:
            A tensor storing the transformation result with the same shape and data type as input.


        Examples:

        .. code-block:: python

            import paddle

            x = paddle.rand(shape=(3, 3, 112, 112), dtype="float32")
            y = paddle.nn.functional.local_response_norm(x, size=5)
            print(y.shape)  # [3, 3, 112, 112]
        """
    if not in_dygraph_mode():
        check_variable_and_dtype(x, 'x', ['float32'], 'local_response_norm')
    if data_format not in ['NCL', 'NLC', 'NCHW', 'NHWC', 'NCDHW', 'NDHWC']:
        raise ValueError(
            "data_format should be in one of [NCL, NCHW, NCDHW, NLC, NHWC, NDHWC], " \
            "but got {}".format(data_format))

    sizes = x.shape
    dim = len(sizes)
    if dim < 3:
        raise ValueError(
            'Expected 3D or higher dimensionality input, but got {} dimensions'.
            format(dim))

H
huangjun12 已提交
491 492
    for i, sz in enumerate(sizes):
        if not sz > 0:
H
huangjun12 已提交
493 494
            raise ValueError("Expected every dim's size to be larger than 0, "
                             "but the size of the {}-th dim is {}".format(i,
H
huangjun12 已提交
495
                                                                          sz))
H
huangjun12 已提交
496

497 498
    channel_last = True if data_format[-1] == "C" else False

499 500 501
    from functools import reduce
    sum_sizes = reduce(lambda x, y: x * y, sizes[1:])

502 503 504 505
    div = paddle.unsqueeze(paddle.multiply(x, x), axis=1)
    if not channel_last:
        pad4d_shape = [0, 0, size // 2, (size - 1) // 2]
        pool2d_shape = (size, 1)
506 507 508 509
        reshape_shape = [
            sizes[0], 1, sizes[1], sizes[2],
            int(sum_sizes / (sizes[1] * sizes[2]))
        ]
510 511 512 513 514
        pad5d_shape = [0, 0, 0, 0, size // 2, (size - 1) // 2]
        pool3d_shape = (size, 1, 1)
    else:
        pad4d_shape = [size // 2, (size - 1) // 2, 0, 0]
        pool2d_shape = (1, size)
515 516 517 518
        reshape_shape = [
            sizes[0], 1, sizes[1], int(sum_sizes / (sizes[1] * sizes[-1])),
            sizes[-1]
        ]
519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539
        pad5d_shape = [size // 2, (size - 1) // 2, 0, 0, 0, 0]
        pool3d_shape = (1, 1, size)

    if dim == 3:
        div = paddle.nn.functional.pad(div, pad=pad4d_shape)
        div = paddle.nn.functional.avg_pool2d(
            div, kernel_size=pool2d_shape, stride=1)
        div = paddle.squeeze(div, axis=1)
    else:
        div = paddle.reshape(div, shape=reshape_shape)
        div = paddle.nn.functional.pad(div,
                                       pad=pad5d_shape,
                                       data_format='NCDHW')
        div = paddle.nn.functional.avg_pool3d(
            div, kernel_size=pool3d_shape, stride=1)
        div = paddle.reshape(paddle.squeeze(div, axis=1), sizes)

    div = paddle.scale(div, scale=alpha, bias=k)
    div = paddle.pow(div, beta)
    res = paddle.divide(x, div, name=name)
    return res