quant_nn.py 23.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
#   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.

from paddle.fluid.dygraph import layers
from paddle.fluid import core
from paddle.fluid import dygraph_utils
from paddle.fluid import unique_name
from paddle.fluid.param_attr import ParamAttr
from paddle.fluid.framework import _varbase_creator
from paddle.fluid.framework import in_dygraph_mode
from paddle.fluid.initializer import Constant
from paddle.fluid.data_feeder import check_variable_and_dtype
H
huangxu96 已提交
24
from paddle.nn import functional as F
25 26
import logging
from paddle.fluid.log_helper import get_logger
27 28

__all__ = [
29 30 31 32 33 34 35 36 37
    'FakeQuantMovingAverageAbsMax',
    'FakeQuantAbsMax',
    'FakeQuantChannelWiseAbsMax',
    'QuantizedConv2D',
    'QuantizedLinear',
    'QuantizedNoweightLayer',
    'MovingAverageAbsMaxScale',
    'MAOutputScaleLayer',
    'FakeQuantMAOutputScaleLayer',
38 39
]

40 41
_logger = get_logger(
    __name__, logging.INFO, fmt='%(asctime)s-%(levelname)s: %(message)s')
42

43 44

class FakeQuantMovingAverageAbsMax(layers.Layer):
45
    r"""
46
    FakeQuantMovingAverageAbsMax layer does the moving_average_abs_max quant and then dequant.
47 48 49 50 51 52 53 54 55 56 57 58
    Its computational formula is described as below:

    :math:`scale = (moving\_rate*accum+max(abs(x)))/(moving\_rate*state+1)`
    :math:`range = 2^{bit\_length - 1} - 1`
    :math:`Out = round(X / scale * range) * scale / range`
    """

    def __init__(self,
                 name=None,
                 moving_rate=0.9,
                 quant_bits=8,
                 dtype='float32'):
59
        super(FakeQuantMovingAverageAbsMax, self).__init__()
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
        self._moving_rate = moving_rate
        self._quant_bits = quant_bits

        scale_prefix = "{}.scale".format(
            name) if name else 'quant_dequant.scale'
        scale_attr = ParamAttr(
            name=unique_name.generate(scale_prefix),
            initializer=Constant(0.001),
            trainable=False)
        self._scale = self.create_parameter(
            shape=[1], attr=scale_attr, dtype=dtype)
        self._scale.stop_gradient = True

        state_prefix = "{}.state".format(
            name) if name else 'quant_dequant.state'
        state_attr = ParamAttr(
            name=unique_name.generate(state_prefix),
            initializer=Constant(1),
            trainable=False)
        self._state = self.create_parameter(
            shape=[1], attr=state_attr, dtype=dtype)
        self._state.stop_gradient = True

        accum_prefix = "{}.accum".format(
            name) if name else 'quant_dequant.accum'
        accum_attr = ParamAttr(
            name=unique_name.generate(accum_prefix),
            initializer=Constant(1),
            trainable=False)
        self._accum = self.create_parameter(
            shape=[1], attr=accum_attr, dtype=dtype)
        self._accum.stop_gradient = True

    def forward(self, input):
        if in_dygraph_mode():
            attrs = ('moving_rate', self._moving_rate, 'bit_length',
                     self._quant_bits, 'is_test', not self.training)
            quant_out = _varbase_creator(
                type=input.type,
                name="{}.quantized.dequantized".format(input.name),
                shape=input.shape,
                dtype=input.dtype,
                persistable=False)
            state = self._state if self.training else None
            accum = self._accum if self.training else None

            out, _, _, _ = core.ops.fake_quantize_dequantize_moving_average_abs_max(
                input, self._scale, accum, state, quant_out, self._scale, state,
                accum, *attrs)
            return out

        check_variable_and_dtype(input, 'input', ['float32'],
112
                                 "FakeQuantMovingAverageAbsMax")
113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142
        attrs = {
            'moving_rate': self._moving_rate,
            'bit_length': self._quant_bits,
            'is_test': not self.training
        }
        inputs = {"X": [input], "InScale": [self._scale]}
        quant_out = self._helper.create_variable(
            name="{}.quantized.dequantized".format(input.name),
            dtype=input.dtype,
            type=core.VarDesc.VarType.LOD_TENSOR,
            persistable=False,
            stop_gradient=False)
        outputs = {"Out": [quant_out], "OutScale": [self._scale]}

        if self.training:
            inputs['InState'] = [self._state]
            inputs['InAccum'] = [self._accum]
            outputs['OutState'] = [self._state]
            outputs['OutAccum'] = [self._accum]

        self._helper.append_op(
            type="fake_quantize_dequantize_moving_average_abs_max",
            inputs=inputs,
            outputs=outputs,
            attrs=attrs)

        return quant_out


class FakeQuantAbsMax(layers.Layer):
143
    r"""
144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223
    FakeQuantAbsMax layer does the abs_max quant and then dequant.
    Its computational formula is described as below:

    :math:`scale = max(abs(X))`
    :math:`range = 2^{bit\_length - 1} - 1`
    :math:`Out = round(X / scale * range) * scale / range`
    """

    def __init__(self,
                 name=None,
                 quant_bits=8,
                 dtype='float32',
                 quant_on_weight=False):
        super(FakeQuantAbsMax, self).__init__()
        self._quant_bits = quant_bits
        self._name = name
        scale_prefix = "{}.scale".format(
            name) if name else 'quant_dequant.scale'
        self._scale_name = unique_name.generate(scale_prefix)
        if quant_on_weight:
            scale_attr = ParamAttr(
                name=self._scale_name,
                initializer=Constant(0.0),
                trainable=False)
            self._scale = self.create_parameter(
                shape=[1], attr=scale_attr, dtype=self._dtype)
            self._scale.stop_gradient = True
        else:
            self._scale = None

    def forward(self, input):
        if in_dygraph_mode():
            attrs = ('bit_length', self._quant_bits)
            quant_out = _varbase_creator(
                type=input.type,
                name="{}.quantized.dequantized".format(input.name),
                shape=input.shape,
                dtype=input.dtype,
                persistable=False)
            out_scale = self._scale
            if not out_scale:
                out_scale = _varbase_creator(
                    type=core.VarDesc.VarType.LOD_TENSOR,
                    name=self._scale_name,
                    shape=[1],
                    dtype=self._dtype,
                    persistable=False)
                out_scale.stop_gradient = True
            out, _, = core.ops.fake_quantize_dequantize_abs_max(
                input, quant_out, out_scale, *attrs)
            return out

        check_variable_and_dtype(input, 'input', ['float32'], "FakeQuantAbsMax")
        attrs = {'bit_length': self._quant_bits}
        inputs = {"X": [input]}
        quant_out = self._helper.create_variable(
            name="{}.quantized.dequantized".format(input.name),
            dtype=input.dtype,
            type=core.VarDesc.VarType.LOD_TENSOR,
            persistable=False,
            stop_gradient=False)
        out_scale = self._scale
        if not out_scale:
            out_scale = self._helper.create_variable(
                name=self._scale_name,
                dtype=self._dtype,
                type=core.VarDesc.VarType.LOD_TENSOR,
                persistable=False,
                stop_gradient=True)
        outputs = {"Out": [quant_out], "OutScale": [out_scale]}

        self._helper.append_op(
            type="fake_quantize_dequantize_abs_max",
            inputs=inputs,
            outputs=outputs,
            attrs=attrs)

        return quant_out


224
class FakeQuantChannelWiseAbsMax(layers.Layer):
H
huangxu96 已提交
225 226 227 228 229 230 231 232
    def __init__(self,
                 name=None,
                 channel_num=None,
                 quant_bits=8,
                 quant_axis=0,
                 dtype='float32',
                 quant_on_weight=False):
        assert quant_on_weight == True, "Channel_wise only can be used on weight quantization."
233
        super(FakeQuantChannelWiseAbsMax, self).__init__()
H
huangxu96 已提交
234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278
        self._quant_bits = quant_bits
        self._quant_axis = quant_axis
        self._dtype = dtype
        self._name = name
        self._channel_num = channel_num
        scale_prefix = "{}.scale".format(
            name) if name else 'quant_dequant.scale'
        self._scale_name = unique_name.generate(scale_prefix)
        if quant_on_weight:
            scale_attr = ParamAttr(
                name=self._scale_name,
                initializer=Constant(0.0),
                trainable=False)
            self._scale = self.create_parameter(
                shape=[self._channel_num], attr=scale_attr, dtype=self._dtype)
            self._scale.stop_gradient = True
        else:
            self._scale = None

    def forward(self, input):
        if in_dygraph_mode():
            attrs = ('bit_length', self._quant_bits, 'quant_axis',
                     self._quant_axis)
            quant_out = _varbase_creator(
                type=input.type,
                name="{}.quantized.dequantized".format(input.name),
                shape=input.shape,
                dtype=input.dtype,
                persistable=False)

            out_scale = self._scale
            if out_scale is None:
                out_scale = _varbase_creator(
                    type=core.VarDesc.VarType.LOD_TENSOR,
                    name=self._scale_name,
                    shape=[self._channel_num],
                    dtype=self._dtype,
                    persistable=False)
                out_scale.stop_gradient = True

            out, _, = core.ops.fake_channel_wise_quantize_dequantize_abs_max(
                input, quant_out, out_scale, *attrs)
            return out

        check_variable_and_dtype(input, 'input', ['float32'],
279
                                 "FakeQuantChannelWiseAbsMax")
H
huangxu96 已提交
280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306
        attrs = {'bit_length': self._quant_bits, 'quant_axis': self._quant_axis}
        inputs = {"X": [input]}
        quant_out = self._helper.create_variable(
            name="{}.quantized.dequantized".format(input.name),
            dtype=input.dtype,
            type=core.VarDesc.VarType.LOD_TENSOR,
            persistable=False,
            stop_gradient=False)
        out_scale = self._scale
        if not out_scale:
            out_scale = self._helper.create_variable(
                name=self._scale_name,
                dtype=self._dtype,
                type=core.VarDesc.VarType.LOD_TENSOR,
                persistable=False,
                stop_gradient=True)
        outputs = {"Out": [quant_out], "OutScale": [out_scale]}

        self._helper.append_op(
            type="fake_channel_wise_quantize_dequantize_abs_max",
            inputs=inputs,
            outputs=outputs,
            attrs=attrs)

        return quant_out


307 308 309 310 311 312 313 314 315 316 317
def _get_fake_quant_type(quant_type, **kwargs):
    call_args = {
        "name": kwargs.get("name", None),
        "quant_bits": kwargs.get("quant_bits", 8),
        "dtype": kwargs.get("dtype", "float32")
    }

    if quant_type == 'abs_max':
        call_args["quant_on_weight"] = kwargs.get("quant_on_weight", False)
    elif quant_type == 'moving_average_abs_max':
        call_args["moving_rate"] = kwargs.get("moving_rate", 0.9)
H
huangxu96 已提交
318 319 320 321 322 323 324
    elif quant_type == 'channel_wise_abs_max':
        call_args["quant_on_weight"] = kwargs.get("quant_on_weight", False)
        call_args["channel_num"] = kwargs.get("channel_num", None)
        call_args["quant_axis"] = kwargs.get("quant_axis", 0)
        assert call_args["channel_num"] is not None, (
            "You need to input channel_num"
            "when you use channel_wise_abs_max strategy.")
325
    fake_quant_map = {
326
        'abs_max': FakeQuantAbsMax,
327 328
        'moving_average_abs_max': FakeQuantMovingAverageAbsMax,
        'channel_wise_abs_max': FakeQuantChannelWiseAbsMax
329
    }
330 331

    return fake_quant_map[quant_type](**call_args)
332 333 334 335 336 337 338 339 340 341 342 343 344 345


class QuantizedConv2D(layers.Layer):
    """
    The computational logic of QuantizedConv2D is the same with Conv2D.
    The only difference is that its inputs are all fake quantized.
    """

    def __init__(self,
                 layer,
                 weight_bits=8,
                 activation_bits=8,
                 moving_rate=0.9,
                 weight_quantize_type='abs_max',
346 347 348 349 350
                 activation_quantize_type='abs_max',
                 weight_pre_layer=None,
                 act_pre_layer=None,
                 weight_quant_layer=None,
                 act_quant_layer=None):
351 352 353 354 355
        super(QuantizedConv2D, self).__init__()
        # For Conv2D
        self._groups = getattr(layer, '_groups')
        self._stride = getattr(layer, '_stride')
        self._padding = getattr(layer, '_padding')
H
huangxu96 已提交
356 357 358 359
        self._padding_mode = getattr(layer, '_padding_mode')
        if self._padding_mode != 'zeros':
            self._reversed_padding_repeated_twice = getattr(
                layer, '_reversed_padding_repeated_twice')
360
        self._dilation = getattr(layer, '_dilation')
H
huangxu96 已提交
361
        self._data_format = getattr(layer, '_data_format')
362 363
        self.weight = getattr(layer, 'weight')
        self.bias = getattr(layer, 'bias')
H
huangxu96 已提交
364

365
        # For FakeQuant
H
huangxu96 已提交
366
        self._conv2d_quant_axis = 0
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
        if weight_quant_layer is not None:
            self._fake_quant_weight = weight_quant_layer()
        else:
            self._fake_quant_weight = _get_fake_quant_type(
                weight_quantize_type,
                name=self.weight.name,
                moving_rate=moving_rate,
                quant_bits=weight_bits,
                dtype=self._dtype,
                quant_on_weight=True,
                channel_num=self.weight.shape[self._conv2d_quant_axis],
                quant_axis=self._conv2d_quant_axis)
        if act_quant_layer is not None:
            self._fake_quant_input = act_quant_layer()
        else:
            self._fake_quant_input = _get_fake_quant_type(
                activation_quantize_type,
                name=layer.full_name(),
                moving_rate=moving_rate,
                quant_bits=activation_bits,
                dtype=self._dtype,
                quant_on_weight=False)

        self._act_preprocess = act_pre_layer(
        ) if act_pre_layer is not None else None
        self._weight_preprocess = weight_pre_layer(
        ) if weight_pre_layer is not None else None
394 395

    def forward(self, input):
396 397
        if self._act_preprocess is not None:
            input = self._act_preprocess(input)
398
        quant_input = self._fake_quant_input(input)
399 400 401 402 403

        weight = self.weight
        if self._weight_preprocess is not None:
            weight = self._weight_preprocess(self.weight)
        quant_weight = self._fake_quant_weight(weight)
404

H
huangxu96 已提交
405 406 407 408 409 410
        if self._padding_mode != 'zeros':
            quant_input = F.pad(quant_input,
                                self._reversed_padding_repeated_twice,
                                mode=self._padding_mode,
                                data_format=self._data_format)
            self._padding = 0
411

H
huangxu96 已提交
412 413 414 415 416 417 418 419 420
        return F.conv2d(
            quant_input,
            quant_weight,
            bias=self.bias,
            padding=self._padding,
            stride=self._stride,
            dilation=self._dilation,
            groups=self._groups,
            data_format=self._data_format)
421 422 423 424 425 426 427 428 429 430 431 432 433 434


class QuantizedLinear(layers.Layer):
    """
    The computational logic of QuantizedLinear is the same with Linear.
    The only difference is that its inputs are all fake quantized.
    """

    def __init__(self,
                 layer,
                 weight_bits=8,
                 activation_bits=8,
                 moving_rate=0.9,
                 weight_quantize_type='abs_max',
435 436 437 438 439
                 activation_quantize_type='abs_max',
                 weight_pre_layer=None,
                 act_pre_layer=None,
                 weight_quant_layer=None,
                 act_quant_layer=None):
440 441 442 443
        super(QuantizedLinear, self).__init__()
        # For Linear
        self.weight = getattr(layer, 'weight')
        self.bias = getattr(layer, 'bias')
H
huangxu96 已提交
444
        self.name = getattr(layer, 'name')
445
        # For FakeQuant
H
huangxu96 已提交
446
        self._linear_quant_axis = 1
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

        if weight_quant_layer is not None:
            self._fake_quant_weight = weight_quant_layer()
        else:
            self._fake_quant_weight = _get_fake_quant_type(
                weight_quantize_type,
                name=self.weight.name,
                moving_rate=moving_rate,
                quant_bits=weight_bits,
                dtype=self._dtype,
                quant_on_weight=True,
                channel_num=self.weight.shape[self._linear_quant_axis],
                quant_axis=self._linear_quant_axis)

        if act_quant_layer is not None:
            self._fake_quant_input = act_quant_layer()
        else:
            self._fake_quant_input = _get_fake_quant_type(
                activation_quantize_type,
                name=layer.full_name(),
                moving_rate=moving_rate,
                quant_bits=activation_bits,
                dtype=self._dtype,
                quant_on_weight=False)

        self._act_preprocess = act_pre_layer(
        ) if act_pre_layer is not None else None
        self._weight_preprocess = weight_pre_layer(
        ) if weight_pre_layer is not None else None
476 477

    def forward(self, input):
478 479
        if self._act_preprocess is not None:
            input = self._act_preprocess(input)
480
        quant_input = self._fake_quant_input(input)
481 482 483 484 485 486

        weight = self.weight
        if self._weight_preprocess is not None:
            weight = self._weight_preprocess(self.weight)
        quant_weight = self._fake_quant_weight(weight)

H
huangxu96 已提交
487 488 489
        out = F.linear(
            x=quant_input, weight=quant_weight, bias=self.bias, name=self.name)
        return out
490 491


492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511
class QuantizedNoweightLayer(layers.Layer):
    def __init__(self,
                 layer,
                 weight_bits=8,
                 activation_bits=8,
                 moving_rate=0.9,
                 *args,
                 **kwargs):

        super(QuantizedNoweightLayer, self).__init__()
        self._layer = layer
        self._fake_quant_input = _get_fake_quant_type(
            'moving_average_abs_max',
            name=layer.full_name(),
            moving_rate=moving_rate,
            quant_bits=activation_bits,
            dtype=self._dtype,
            quant_on_weight=False)

    def forward(self, input):
512
        return self._layer.forward(self._fake_quant_input(input))
513 514


515
class MovingAverageAbsMaxScale(layers.Layer):
516
    def __init__(self, name=None, moving_rate=0.9, dtype='float32'):
517
        r"""
518 519
        MovingAverageMaxScale layer is used to calculating the output quantization
        scale of Layer. Its computational formula is described as below:
520 521 522 523 524 525 526

        :math:`scale = (moving\_rate*accum+max(abs(x)))/(moving\_rate*state+1)`
        :math:`Out = X`
        """
        super(MovingAverageAbsMaxScale, self).__init__()
        self._moving_rate = moving_rate

527 528 529 530 531 532 533
        scale_prefix = '{}.scale'.format(name) if name else 'outscale.scale'
        scale_name = unique_name.generate(scale_prefix)
        scale_attr = ParamAttr(
            name=scale_name, initializer=Constant(1), trainable=False)
        self._scale = self.create_parameter(
            shape=[1], attr=scale_attr, dtype=dtype)
        self._scale.stop_gradient = True
534

535 536 537 538 539 540 541 542
        state_prefix = "{}.state".format(name) if name else 'outscale.state'
        state_attr = ParamAttr(
            name=unique_name.generate(state_prefix),
            initializer=Constant(1),
            trainable=False)
        self._state = self.create_parameter(
            shape=[1], attr=state_attr, dtype=dtype)
        self._state.stop_gradient = True
543

544 545 546 547 548 549 550 551
        accum_prefix = "{}.accum".format(name) if name else 'outscale.accum'
        accum_attr = ParamAttr(
            name=unique_name.generate(accum_prefix),
            initializer=Constant(1),
            trainable=False)
        self._accum = self.create_parameter(
            shape=[1], attr=accum_attr, dtype=dtype)
        self._accum.stop_gradient = True
552 553 554 555 556 557 558

    def forward(self, input):
        if in_dygraph_mode():
            attrs = ('moving_rate', self._moving_rate, 'is_test',
                     not self.training)
            state = self._state if self.training else None
            accum = self._accum if self.training else None
559 560 561 562 563 564
            quant_out = _varbase_creator(
                type=input.type,
                name="{}.tmp".format(input.name),
                shape=input.shape,
                dtype=input.dtype,
                persistable=False)
565

566 567 568 569
            out, _, _, _ = core.ops.moving_average_abs_max_scale(
                input, accum, state, quant_out, self._scale, state, accum,
                *attrs)
            return out
570 571 572 573 574 575

        check_variable_and_dtype(input, 'input', ['float32', 'float64'],
                                 'MovingAverageAbsMaxScale')

        attrs = {'moving_rate': self._moving_rate, 'is_test': not self.training}
        inputs = {"X": [input]}
576 577 578 579 580 581 582
        quant_out = self._helper.create_variable(
            name="{}.tmp".format(input.name),
            dtype=input.dtype,
            type=core.VarDesc.VarType.LOD_TENSOR,
            persistable=False,
            stop_gradient=False)
        outputs = {"Out": [quant_out], "OutScale": [self._scale]}
583 584 585 586 587 588 589 590 591 592 593 594 595

        if self.training:
            inputs['InState'] = [self._state]
            inputs['InAccum'] = [self._accum]
            outputs['OutState'] = [self._state]
            outputs['OutAccum'] = [self._accum]

        self._helper.append_op(
            type="moving_average_abs_max_scale",
            inputs=inputs,
            outputs=outputs,
            attrs=attrs)

596 597 598
        return quant_out


599 600 601 602 603 604 605
class MAOutputScaleLayer(layers.Layer):
    """
    Calculate the scale (moving average abs max) for the output of the input layer.
    Add MovingAverageMaxScale layer to the behind of the input layer.
    """

    def __init__(self, layer=None, moving_rate=0.9, name=None, dtype='float32'):
606
        r"""
607
        Construct
608
        """
609
        super(MAOutputScaleLayer, self).__init__()
610
        self._layer = layer
611 612 613 614 615 616 617 618 619 620 621 622
        if name is None:
            name = layer.full_name()
        self._ma_output_scale = \
            MovingAverageAbsMaxScale(name, moving_rate, dtype)

    def forward(self, *inputs, **kwargs):
        out = self._layer(*inputs, **kwargs)
        # TODO (jc): support the ops of several outputs
        if (isinstance(out, list) or isinstance(out, tuple)) and len(out) > 1:
            return out
        else:
            return self._ma_output_scale(out)
623

624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651

class FakeQuantMAOutputScaleLayer(layers.Layer):
    def __init__(self,
                 layer,
                 weight_bits=8,
                 activation_bits=8,
                 moving_rate=0.9,
                 name=None,
                 *args,
                 **kwargs):

        super(FakeQuantMAOutputScaleLayer, self).__init__()
        self._layer = layer
        self._fake_quant_output = _get_fake_quant_type(
            'moving_average_abs_max',
            name=layer.full_name() if name is None else name,
            moving_rate=moving_rate,
            quant_bits=activation_bits,
            dtype=self._dtype,
            quant_on_weight=False)

    def forward(self, *inputs, **kwargs):
        out = self._layer(*inputs, **kwargs)
        # TODO (jc): support the ops of several outputs
        if (isinstance(out, list) or isinstance(out, tuple)) and len(out) > 1:
            return out
        else:
            return self._fake_quant_output(out)