clip.py 16.3 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
F
fengjiayi 已提交
3 4 5
# 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
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
F
fengjiayi 已提交
9 10 11 12 13
# 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.
F
update  
fengjiayi 已提交
14

15 16
from __future__ import print_function

F
fengjiayi 已提交
17
import copy
18
import six
F
fengjiayi 已提交
19

Y
Yu Yang 已提交
20
import functools
21 22
from . import layers
from . import framework
F
fengjiayi 已提交
23
from . import core
J
Jiabin Yang 已提交
24
from .dygraph.base import _not_support
Y
Yu Yang 已提交
25

F
fengjiayi 已提交
26
__all__ = [
27
    'set_gradient_clip',
28
    'ErrorClipByValue',
F
fengjiayi 已提交
29 30 31
    'GradientClipByValue',
    'GradientClipByNorm',
    'GradientClipByGlobalNorm',
F
fengjiayi 已提交
32
]
Y
Yu Yang 已提交
33 34


F
fengjiayi 已提交
35
class BaseErrorClipAttr(object):
F
fengjiayi 已提交
36 37 38
    def __str__(self):
        raise NotImplementedError()

Y
yuyang18 已提交
39
    def _append_clip_op(self, block, grad_name):
F
fengjiayi 已提交
40 41 42 43
        raise NotImplementedError()


class ErrorClipByValue(BaseErrorClipAttr):
44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59
    """
    Clips tensor values to the range [min, max].

    Given a tensor t, this operation clips its value to min and max inplace.

    - Any values less than min are set to min.
    - Any values greater than max are set to max.

    Args:
        max (float): The maximum value to clip by.
        min (float, optional): The minimum value to clip by. if not set by user, \
        will be set to -max by framework.

    Examples:
        .. code-block:: python

60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76
            import paddle.fluid as fluid
            BATCH_SIZE = 128
            CLIP_MAX = 2e-6
            CLIP_MIN = -1e-6
            prog = fluid.framework.Program()
            with fluid.program_guard(main_program=prog):
                image = fluid.layers.data(name='x', shape=[784], dtype='float32')
                hidden1 = fluid.layers.fc(input=image, size=128, act='relu')
                hidden2 = fluid.layers.fc(input=hidden1, size=64, act='relu')
                predict = fluid.layers.fc(input=hidden2, size=10, act='softmax')
                label = fluid.layers.data(name='y', shape=[1], dtype='int64')
                cost = fluid.layers.cross_entropy(input=predict, label=label)
                avg_cost = fluid.layers.mean(cost)
            prog_clip = prog.clone()
            prog_clip.block(0).var(hidden1.name)._set_error_clip(
                fluid.clip.ErrorClipByValue(
                    max=CLIP_MAX, min=CLIP_MIN)
77 78
    """

F
fengjiayi 已提交
79 80 81 82 83 84 85 86 87
    def __init__(self, max, min=None):
        max = float(max)
        if min is None:
            min = -max
        else:
            min = float(min)
        self.max = max
        self.min = min

F
fengjiayi 已提交
88 89 90
    def __str__(self):
        return "ByValue, min=%f, max=%f" % (self.min, self.max)

Y
yuyang18 已提交
91
    def _append_clip_op(self, block, grad_name):
92 93 94 95
        clip_op_desc = block.desc.append_op()
        clip_op_desc.set_type("clip")
        clip_op_desc.set_input("X", [grad_name])
        clip_op_desc.set_output("Out", [grad_name])
W
Wu Yi 已提交
96 97
        clip_op_desc._set_attr("min", self.min)
        clip_op_desc._set_attr("max", self.max)
F
fengjiayi 已提交
98 99 100 101 102 103


def error_clip_callback(block, context):
    # the context is a grad_to_var map
    grad_to_var = context
    op_desc = block.desc.op(block.desc.op_size() - 1)
104
    for grad_n in [n for n in op_desc.output_arg_names() if n in grad_to_var]:
W
Wu Yi 已提交
105
        fwd_var = block._var_recursive(grad_to_var[grad_n])
F
fengjiayi 已提交
106
        error_clip = getattr(fwd_var, "error_clip", None)
F
fengjiayi 已提交
107 108 109 110 111
        if not (error_clip is None or isinstance(error_clip,
                                                 BaseErrorClipAttr)):
            raise TypeError(
                "Variable's error_clip should be an instance of BaseErrorClipAttr or None."
            )
F
fengjiayi 已提交
112
        if error_clip is not None:
Y
yuyang18 已提交
113
            error_clip._append_clip_op(block, grad_n)
F
fengjiayi 已提交
114 115


Y
Yu Yang 已提交
116
class BaseGradientClipAttr(object):
F
fengjiayi 已提交
117 118 119
    def __str__(self):
        raise NotImplementedError()

Y
yuyang18 已提交
120
    def _process_context(self, context, param, grad):
Y
Yu Yang 已提交
121 122
        raise NotImplementedError()

Y
yuyang18 已提交
123
    def _create_operators(self, param, grad):
Y
Yu Yang 已提交
124 125 126 127
        raise NotImplementedError()


class NullGradientClipAttr(BaseGradientClipAttr):
F
fengjiayi 已提交
128 129 130
    def __str__(self):
        return "Null"

Y
yuyang18 已提交
131
    def _process_context(self, context, param, grad):
Y
Yu Yang 已提交
132 133
        pass

Y
yuyang18 已提交
134
    def _create_operators(self, param, grad):
Y
Yu Yang 已提交
135 136 137 138
        return param, grad


class GradientClipByValue(BaseGradientClipAttr):
139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154
    """
    Clips gradient values to the range [min, max].

    Given a tensor t, this operation clips its value to min and max inplace.

    - Any values less than min are set to min.
    - Any values greater than max are set to max.

    Args:
        max (float): The maximum value to clip by.
        min (float, optional): The minimum value to clip by. if not set by user, \
        will be set to -max by framework.

    Examples:
        .. code-block:: python

155
            import paddle.fluid as fluid
T
Tink_Y 已提交
156 157
            w_param_attrs = fluid.ParamAttr(name=None,
              initializer=fluid.initializer.UniformInitializer(low=-1.0, high=1.0, seed=0),
158
              learning_rate=1.0,
T
Tink_Y 已提交
159
              regularizer=fluid.regularizer.L1Decay(1.0),
160
              trainable=True,
161 162
              gradient_clip=fluid.clip.GradientClipByValue(-1.0, 1.0))
            x = fluid.layers.data(name='x', shape=[10], dtype='float32')
163 164 165
            y_predict = fluid.layers.fc(input=x, size=1, param_attr=w_param_attrs)
    """

Y
Yu Yang 已提交
166 167 168 169 170 171 172 173 174
    def __init__(self, max, min=None):
        max = float(max)
        if min is None:
            min = -max
        else:
            min = float(min)
        self.max = max
        self.min = min

F
fengjiayi 已提交
175 176 177
    def __str__(self):
        return "ByValue, min=%f, max=%f" % (self.min, self.max)

Y
yuyang18 已提交
178
    def _process_context(self, context, param, grad):
Y
Yu Yang 已提交
179 180
        pass

Y
yuyang18 已提交
181
    def _create_operators(self, param, grad):
Y
Yu Yang 已提交
182 183 184 185
        new_grad = layers.clip(x=grad, min=self.min, max=self.max)
        return param, new_grad


F
fengjiayi 已提交
186
class GradientClipByNorm(BaseGradientClipAttr):
187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207
    """
    Clips tensor values to a maximum L2-norm.

    This operator limits the L2 norm of the input :math:`X` within :math:`max\_norm`.
    If the L2 norm of :math:`X` is less than or equal to :math:`max\_norm`, :math:`Out`
    will be the same as :math:`X`. If the L2 norm of :math:`X` is greater than
    :math:`max\_norm`, :math:`X` will be linearly scaled to make the L2 norm of
    :math:`Out` equal to :math:`max\_norm`, as shown in the following formula:

    .. math::

        Out = \\frac{max\_norm * X}{norm(X)},

    where :math:`norm(X)` represents the L2 norm of :math:`X`.

    Args:
        clip_norm (float): The maximum norm value

    Examples:
        .. code-block:: python

208 209
            import paddle.fluid as fluid
            w_param_attrs = fluid.ParamAttr(name=None,
T
Tink_Y 已提交
210
              initializer=fluid.initializer.UniformInitializer(low=-1.0, high=1.0, seed=0),
211
              learning_rate=1.0,
T
Tink_Y 已提交
212
              regularizer=fluid.regularizer.L1Decay(1.0),
213
              trainable=True,
214 215
              gradient_clip=fluid.clip.GradientClipByNorm(clip_norm=2.0))
            x = fluid.layers.data(name='x', shape=[10], dtype='float32')
216 217 218 219
            y_predict = fluid.layers.fc(input=x, size=1, param_attr=w_param_attrs)

    """

F
fengjiayi 已提交
220 221 222
    def __init__(self, clip_norm):
        self.clip_norm = clip_norm

F
fengjiayi 已提交
223 224 225
    def __str__(self):
        return "ByNorm, clip_norm=%f" % self.clip_norm

Y
yuyang18 已提交
226
    def _process_context(self, context, param, grad):
F
fengjiayi 已提交
227 228
        pass

Y
yuyang18 已提交
229
    def _create_operators(self, param, grad):
F
fengjiayi 已提交
230 231 232 233
        new_grad = layers.clip_by_norm(x=grad, max_norm=self.clip_norm)
        return param, new_grad


F
fengjiayi 已提交
234
class GradientClipByGlobalNorm(BaseGradientClipAttr):
235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263
    """
    Clips values of multiple tensors by the ratio of the sum of their norms.

    Given a list of tensors t_list, and a clipping ratio clip_norm, this
    operation returns a list of clipped tensors list_clipped and the global
    norm (global_norm) of all tensors in t_list.

    To perform the clipping, the values :math:`t\_list[i]` are set to:

    .. math::

        t\_list[i] = t\_list[i] * \\frac{clip\_norm}{\max(global\_norm, clip\_norm)}

    where:

    .. math::

        global\_norm = \sqrt{\sum_{i=0}^{N-1}(l2norm(t\_list[i]))^2}

    If :math:`clip\_norm > global\_norm` then the entries in t_list remain as they are,
    otherwise they're all shrunk by the global ratio.

    Args:
        clip_norm (float): The maximum norm value
        group_name (str, optional): The group name for this clip.

    Examples:
        .. code-block:: python

264 265 266 267 268 269 270 271 272 273 274 275 276 277
            import paddle.fluid as fluid
            prog = fluid.framework.Program()
            startup_program = fluid.framework.Program()
            with fluid.program_guard(
                    main_program=prog, startup_program=startup_program):
                image = fluid.layers.data(name='x', shape=[784], dtype='float32')
                label = fluid.layers.data(name='y', shape=[1], dtype='int64')
                hidden1 = fluid.layers.fc(input=image, size=128, act='relu')
                hidden2 = fluid.layers.fc(input=hidden1, size=64, act='relu')
                predict = fluid.layers.fc(input=hidden2, size=10, act='softmax')
                cost = fluid.layers.cross_entropy(input=predict, label=label)
                avg_cost = fluid.layers.mean(cost)
            prog_clip = prog.clone()
            avg_cost_clip = prog_clip.block(0).var(avg_cost.name)
278 279 280 281 282 283 284 285 286
            p_g_clip = fluid.backward.append_backward(loss=avg_cost_clip)

            with fluid.program_guard(main_program=prog_clip):
                fluid.clip.set_gradient_clip(
                    fluid.clip.GradientClipByGlobalNorm(clip_norm=2.0))
                p_g_clip = fluid.clip.append_gradient_clip_ops(p_g_clip)

    """

F
update  
fengjiayi 已提交
287
    def __init__(self, clip_norm, group_name="default_group"):
288 289
        if not isinstance(group_name, six.string_types):
            raise TypeError("'group_name' must be a %s." % (six.string_types))
F
update  
fengjiayi 已提交
290 291 292

        self.clip_norm = clip_norm
        self.group_name = group_name
293

F
fengjiayi 已提交
294 295 296 297
    def __str__(self):
        return "ByGlobalNorm, group_name=%s, clip_norm=%f" % (self.group_name,
                                                              self.clip_norm)

Y
yuyang18 已提交
298
    def _process_context(self, context, param, grad):
F
update  
fengjiayi 已提交
299 300 301 302 303 304 305 306 307 308
        if self.group_name not in context:
            context[self.group_name] = []
            context[self.group_name + "_clip_value"] = self.clip_norm
            context[self.group_name + "_clip"] = layers.fill_constant(
                shape=[1], dtype="float32", value=self.clip_norm)
        else:
            if not self.clip_norm == context[self.group_name + "_clip_value"]:
                raise ValueError(
                    "All parameters' 'clip_norm' of a same group should be the same"
                )
F
fengjiayi 已提交
309

C
chengduo 已提交
310 311 312 313 314 315
        merge_grad = grad
        if grad.type == core.VarDesc.VarType.SELECTED_ROWS:
            merge_grad = layers.merge_selected_rows(grad)
            merge_grad = layers.get_tensor_from_selected_rows(merge_grad)

        square = layers.square(merge_grad)
P
phlrain 已提交
316
        local_norm_var = layers.reduce_sum(input=square)
F
update  
fengjiayi 已提交
317
        context[self.group_name].append(local_norm_var)
F
fengjiayi 已提交
318

F
update  
fengjiayi 已提交
319
        self.context = context
320

Y
yuyang18 已提交
321
    def _create_operators(self, param, grad):
F
update  
fengjiayi 已提交
322 323 324
        group_scale_name = self.group_name + "_scale"
        if group_scale_name not in self.context:
            group_norm_var = layers.sums(input=self.context[self.group_name])
T
tensor-tang 已提交
325
            group_norm_var = layers.sqrt(x=group_norm_var)
F
update  
fengjiayi 已提交
326 327 328
            clip_var = self.context[self.group_name + "_clip"]
            group_scale_var = layers.elementwise_div(
                x=clip_var,
F
fengjiayi 已提交
329
                y=layers.elementwise_max(
F
update  
fengjiayi 已提交
330
                    x=clip_var, y=group_norm_var))
331
            assert group_scale_var.shape == (1, )
F
update  
fengjiayi 已提交
332
            self.context[group_scale_name] = group_scale_var
F
fengjiayi 已提交
333

F
update  
fengjiayi 已提交
334 335
        new_grad = layers.elementwise_mul(
            x=grad, y=self.context[group_scale_name])
C
chengduo 已提交
336

337
        return param, new_grad
F
fengjiayi 已提交
338 339


J
Jiabin Yang 已提交
340
@_not_support
F
fengjiayi 已提交
341
def set_gradient_clip(clip, param_list=None, program=None):
F
fengjiayi 已提交
342
    """
343 344 345 346
    To specify parameters that require gradient clip.

    Args:
        clip(BaseGradientClipAttr): An instance of some derived class of BaseGradientClipAttr,
347
                for example :ref:`api_fluid_clip_GradientClipByGlobalNorm` ,
348
                which describes the type and detailed attributes of required gradient clip.
349
        param_list(list(Variable), optional): Parameters that require gradient clip.
350
                It can be a list of parameter or a list of parameter's name.
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
                Default None, meaning that all parameters in the program will be included.
        program(Program, optional): The program where parameters are located.
                Default None, meaning that using :ref:`api_fluid_default_main_program` .

    Returns:
        None

    Examples:
        .. code-block:: python
            
            import paddle.fluid as fluid

            def network():
                image = fluid.layers.data(name='image', shape=[28], dtype='float32')
                param_attr1 = fluid.ParamAttr("fc1_param")
                fc1 = fluid.layers.fc(image, size=10, param_attr=param_attr1)
                param_attr2 = fluid.ParamAttr("fc2_param")
                fc2 = fluid.layers.fc(fc1, size=10, param_attr=param_attr2)
                loss = fluid.layers.reduce_mean(fc2)
                return loss


            # network 1: clip all parameter gradient
            with fluid.program_guard(fluid.Program(), fluid.Program()):
                loss = network()
                fluid.clip.set_gradient_clip(
                    fluid.clip.GradientClipByGlobalNorm(clip_norm=2.0))
                sgd = fluid.optimizer.SGD(learning_rate=1e-3)
                sgd.minimize(loss)

            # network 2: clip parameter gradient by name
            with fluid.program_guard(fluid.Program(), fluid.Program()):
                loss = network()
                fluid.clip.set_gradient_clip(
                    fluid.clip.GradientClipByValue(min=-1.0, max=1.0),
                    param_list=["fc1_param", "fc2_param"])
                sgd = fluid.optimizer.SGD(learning_rate=1e-3)
                sgd.minimize(loss)

            # network 3: clip parameter gradient by var
            with fluid.program_guard(fluid.Program(), fluid.Program()):
                loss = network()
                param_var1 = fluid.default_main_program().global_block().var("fc1_param")
                param_var2 = fluid.default_main_program().global_block().var("fc2_param")
                fluid.clip.set_gradient_clip(
                    fluid.clip.GradientClipByValue(min=-1.0, max=1.0),
                    param_list=[param_var1, param_var2])
                sgd = fluid.optimizer.SGD(learning_rate=1e-3)
                sgd.minimize(loss)
F
fengjiayi 已提交
400
    """
F
fengjiayi 已提交
401 402 403 404
    if not isinstance(clip, BaseGradientClipAttr):
        raise TypeError(
            "'clip' should be an instance of BaseGradientClipAttr's derived class"
        )
F
fengjiayi 已提交
405 406 407 408
    if program is None:
        program = framework.default_main_program()
    if param_list is None:
        param_list = program.block(0).all_parameters()
409
    if all(isinstance(elem, six.string_types) for elem in param_list):
F
fengjiayi 已提交
410 411 412 413 414 415 416
        param_list = [program.block(0).var(elem) for elem in param_list]
    if not all(isinstance(elem, framework.Parameter) for elem in param_list):
        raise TypeError(
            "'param_list' should be a list of Parameter or basestring(parameter's name)."
        )

    for param in param_list:
F
fengjiayi 已提交
417
        param.gradient_clip_attr = copy.deepcopy(clip)
F
fengjiayi 已提交
418 419


420
def append_gradient_clip_ops(param_grads):
Y
Yu Yang 已提交
421
    context = dict()
422 423 424
    for p, g in param_grads:
        if g is None:
            continue
X
Xin Pan 已提交
425 426
        with p.block.program._optimized_guard(
            [p, g]), framework.name_scope('append_clip'):
Y
yuyang18 已提交
427 428 429 430 431 432 433
            clip_attr = getattr(p, 'gradient_clip_attr', NullGradientClipAttr())
            if clip_attr is None:
                clip_attr = NullGradientClipAttr()
            if not isinstance(clip_attr, BaseGradientClipAttr):
                raise TypeError(
                    "clip attribute should be an instance of BaseGradientClipAttr"
                )
Y
Yu Yang 已提交
434

Y
yuyang18 已提交
435
            clip_attr._process_context(context=context, param=p, grad=g)
Y
yuyang18 已提交
436 437

    res = []
438 439 440
    for p, g in param_grads:
        if g is None:
            continue
X
Xin Pan 已提交
441 442
        with p.block.program._optimized_guard(
            [p, g]), framework.name_scope('append_graident_clip'):
Y
yuyang18 已提交
443
            res.append(clip_attr._create_operators(param=p, grad=g))
Y
Yu Yang 已提交
444

Y
yuyang18 已提交
445
    return res
Y
Yu Yang 已提交
446 447 448


ClipByValue = GradientClipByValue
F
fengjiayi 已提交
449 450
ClipByNorm = GradientClipByNorm
ClipByGlobalNorm = GradientClipByGlobalNorm