base_network.py 12.6 KB
Newer Older
L
lvmengsi 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
#copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
#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 __future__ import division
import paddle.fluid as fluid
import numpy as np
L
lvmengsi 已提交
18
import math
L
lvmengsi 已提交
19
import os
L
lvmengsi 已提交
20
import warnings
L
lvmengsi 已提交
21 22 23 24 25 26

use_cudnn = True
if 'ce_mode' in os.environ:
    use_cudnn = False


L
lvmengsi 已提交
27 28 29 30 31 32 33 34 35 36
def cal_padding(img_size, stride, filter_size, dilation=1):
    """Calculate padding size."""
    valid_filter_size = dilation * (filter_size - 1) + 1
    if img_size % stride == 0:
        out_size = max(filter_size - stride, 0)
    else:
        out_size = max(filter_size - (img_size % stride), 0)
    return out_size // 2, out_size - out_size // 2


L
lvmengsi 已提交
37
def norm_layer(input, norm_type='batch_norm', name=None, is_test=False):
L
lvmengsi 已提交
38 39
    if norm_type == 'batch_norm':
        param_attr = fluid.ParamAttr(
L
lvmengsi 已提交
40
            name=name + '_w', initializer=fluid.initializer.Constant(1.0))
L
lvmengsi 已提交
41 42 43 44 45 46
        bias_attr = fluid.ParamAttr(
            name=name + '_b', initializer=fluid.initializer.Constant(value=0.0))
        return fluid.layers.batch_norm(
            input,
            param_attr=param_attr,
            bias_attr=bias_attr,
L
lvmengsi 已提交
47
            is_test=is_test,
L
lvmengsi 已提交
48 49 50 51 52 53 54 55 56 57 58 59 60 61 62
            moving_mean_name=name + '_mean',
            moving_variance_name=name + '_var')

    elif norm_type == 'instance_norm':
        helper = fluid.layer_helper.LayerHelper("instance_norm", **locals())
        dtype = helper.input_dtype()
        epsilon = 1e-5
        mean = fluid.layers.reduce_mean(input, dim=[2, 3], keep_dim=True)
        var = fluid.layers.reduce_mean(
            fluid.layers.square(input - mean), dim=[2, 3], keep_dim=True)
        if name is not None:
            scale_name = name + "_scale"
            offset_name = name + "_offset"
        scale_param = fluid.ParamAttr(
            name=scale_name,
L
lvmengsi 已提交
63
            initializer=fluid.initializer.Constant(1.0),
L
lvmengsi 已提交
64 65 66 67 68 69 70 71 72 73 74 75 76 77 78
            trainable=True)
        offset_param = fluid.ParamAttr(
            name=offset_name,
            initializer=fluid.initializer.Constant(0.0),
            trainable=True)
        scale = helper.create_parameter(
            attr=scale_param, shape=input.shape[1:2], dtype=dtype)
        offset = helper.create_parameter(
            attr=offset_param, shape=input.shape[1:2], dtype=dtype)

        tmp = fluid.layers.elementwise_mul(x=(input - mean), y=scale, axis=1)
        tmp = tmp / fluid.layers.sqrt(var + epsilon)
        tmp = fluid.layers.elementwise_add(tmp, offset, axis=1)
        return tmp
    else:
L
lvmengsi 已提交
79
        raise NotImplementedError("norm type: [%s] is not support" % norm_type)
L
lvmengsi 已提交
80 81


L
lvmengsi 已提交
82
def initial_type(name,
L
lvmengsi 已提交
83 84 85
                 input,
                 op_type,
                 fan_out,
L
lvmengsi 已提交
86 87 88 89 90
                 init="normal",
                 use_bias=False,
                 filter_size=0,
                 stddev=0.02):
    if init == "kaiming":
L
lvmengsi 已提交
91 92 93 94 95 96 97 98 99
        if op_type == 'conv':
            fan_in = input.shape[1] * filter_size * filter_size
        elif op_type == 'deconv':
            fan_in = fan_out * filter_size * filter_size
        else:
            if len(input.shape) > 2:
                fan_in = input.shape[1] * input.shape[2] * input.shape[3]
            else:
                fan_in = input.shape[1]
L
lvmengsi 已提交
100 101 102
        bound = 1 / math.sqrt(fan_in)
        param_attr = fluid.ParamAttr(
            name=name + "_w",
L
lvmengsi 已提交
103 104
            initializer=fluid.initializer.Uniform(
                low=-bound, high=bound))
L
lvmengsi 已提交
105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124
        if use_bias == True:
            bias_attr = fluid.ParamAttr(
                name=name + '_b',
                initializer=fluid.initializer.Uniform(
                    low=-bound, high=bound))
        else:
            bias_attr = False
    else:
        param_attr = fluid.ParamAttr(
            name=name + "_w",
            initializer=fluid.initializer.NormalInitializer(
                loc=0.0, scale=stddev))
        if use_bias == True:
            bias_attr = fluid.ParamAttr(
                name=name + "_b", initializer=fluid.initializer.Constant(0.0))
        else:
            bias_attr = False
    return param_attr, bias_attr


L
lvmengsi 已提交
125 126 127 128 129 130 131 132 133 134
def conv2d(input,
           num_filters=64,
           filter_size=7,
           stride=1,
           stddev=0.02,
           padding=0,
           name="conv2d",
           norm=None,
           activation_fn=None,
           relufactor=0.0,
L
lvmengsi 已提交
135 136
           use_bias=False,
           padding_type=None,
L
lvmengsi 已提交
137 138
           initial="normal",
           is_test=False):
L
lvmengsi 已提交
139 140 141 142 143 144 145 146

    if padding != 0 and padding_type != None:
        warnings.warn(
            'padding value and padding type are set in the same time, and the final padding width and padding height are computed by padding_type'
        )

    param_attr, bias_attr = initial_type(
        name=name,
L
lvmengsi 已提交
147 148 149
        input=input,
        op_type='conv',
        fan_out=num_filters,
L
lvmengsi 已提交
150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171
        init=initial,
        use_bias=use_bias,
        filter_size=filter_size,
        stddev=stddev)

    need_crop = False
    if padding_type == "SAME":
        top_padding, bottom_padding = cal_padding(input.shape[2], stride,
                                                  filter_size)
        left_padding, right_padding = cal_padding(input.shape[2], stride,
                                                  filter_size)
        height_padding = bottom_padding
        width_padding = right_padding
        if top_padding != bottom_padding or left_padding != right_padding:
            height_padding = top_padding + stride
            width_padding = left_padding + stride
            need_crop = True
        padding = [height_padding, width_padding]
    elif padding_type == "VALID":
        height_padding = 0
        width_padding = 0
        padding = [height_padding, width_padding]
L
lvmengsi 已提交
172
    else:
L
lvmengsi 已提交
173
        padding = padding
L
lvmengsi 已提交
174 175 176 177 178 179 180 181 182 183 184

    conv = fluid.layers.conv2d(
        input,
        num_filters,
        filter_size,
        name=name,
        stride=stride,
        padding=padding,
        use_cudnn=use_cudnn,
        param_attr=param_attr,
        bias_attr=bias_attr)
L
lvmengsi 已提交
185 186 187 188 189
    if need_crop:
        conv = fluid.layers.crop(
            conv,
            shape=(-1, conv.shape[1], conv.shape[2] - 1, conv.shape[3] - 1),
            offsets=(0, 0, 1, 1))
L
lvmengsi 已提交
190
    if norm is not None:
L
lvmengsi 已提交
191 192
        conv = norm_layer(
            input=conv, norm_type=norm, name=name + "_norm", is_test=is_test)
L
lvmengsi 已提交
193 194 195 196 197 198 199
    if activation_fn == 'relu':
        conv = fluid.layers.relu(conv, name=name + '_relu')
    elif activation_fn == 'leaky_relu':
        conv = fluid.layers.leaky_relu(
            conv, alpha=relufactor, name=name + '_leaky_relu')
    elif activation_fn == 'tanh':
        conv = fluid.layers.tanh(conv, name=name + '_tanh')
L
lvmengsi 已提交
200 201
    elif activation_fn == 'sigmoid':
        conv = fluid.layers.sigmoid(conv, name=name + '_sigmoid')
L
lvmengsi 已提交
202 203 204 205 206 207 208 209 210 211 212 213 214 215
    elif activation_fn == None:
        conv = conv
    else:
        raise NotImplementedError("activation: [%s] is not support" %
                                  activation_fn)

    return conv


def deconv2d(input,
             num_filters=64,
             filter_size=7,
             stride=1,
             stddev=0.02,
L
lvmengsi 已提交
216
             padding=0,
L
lvmengsi 已提交
217 218 219 220 221 222
             outpadding=[0, 0, 0, 0],
             name="deconv2d",
             norm=None,
             activation_fn=None,
             relufactor=0.0,
             use_bias=False,
L
lvmengsi 已提交
223 224
             padding_type=None,
             output_size=None,
L
lvmengsi 已提交
225 226
             initial="normal",
             is_test=False):
L
lvmengsi 已提交
227 228 229 230 231 232 233 234

    if padding != 0 and padding_type != None:
        warnings.warn(
            'padding value and padding type are set in the same time, and the final padding width and padding height are computed by padding_type'
        )

    param_attr, bias_attr = initial_type(
        name=name,
L
lvmengsi 已提交
235 236 237
        input=input,
        op_type='deconv',
        fan_out=num_filters,
L
lvmengsi 已提交
238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259
        init=initial,
        use_bias=use_bias,
        filter_size=filter_size,
        stddev=stddev)

    need_crop = False
    if padding_type == "SAME":
        top_padding, bottom_padding = cal_padding(input.shape[2], stride,
                                                  filter_size)
        left_padding, right_padding = cal_padding(input.shape[2], stride,
                                                  filter_size)
        height_padding = bottom_padding
        width_padding = right_padding
        if top_padding != bottom_padding or left_padding != right_padding:
            height_padding = top_padding + stride
            width_padding = left_padding + stride
            need_crop = True
        padding = [height_padding, width_padding]
    elif padding_type == "VALID":
        height_padding = 0
        width_padding = 0
        padding = [height_padding, width_padding]
L
lvmengsi 已提交
260
    else:
L
lvmengsi 已提交
261
        padding = padding
L
lvmengsi 已提交
262 263 264 265 266 267 268 269 270 271 272 273 274

    conv = fluid.layers.conv2d_transpose(
        input,
        num_filters,
        output_size=output_size,
        name=name,
        filter_size=filter_size,
        stride=stride,
        padding=padding,
        use_cudnn=use_cudnn,
        param_attr=param_attr,
        bias_attr=bias_attr)

L
lvmengsi 已提交
275 276 277
    if outpadding != 0 and padding_type == None:
        conv = fluid.layers.pad2d(
            conv, paddings=outpadding, mode='constant', pad_value=0.0)
L
lvmengsi 已提交
278 279

    if norm is not None:
L
lvmengsi 已提交
280 281
        conv = norm_layer(
            input=conv, norm_type=norm, name=name + "_norm", is_test=is_test)
L
lvmengsi 已提交
282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308
    if activation_fn == 'relu':
        conv = fluid.layers.relu(conv, name=name + '_relu')
    elif activation_fn == 'leaky_relu':
        if relufactor == 0.0:
            raise Warning(
                "the activation is leaky_relu, but the relufactor is 0")
        conv = fluid.layers.leaky_relu(
            conv, alpha=relufactor, name=name + '_leaky_relu')
    elif activation_fn == 'tanh':
        conv = fluid.layers.tanh(conv, name=name + '_tanh')
    elif activation_fn == 'sigmoid':
        conv = fluid.layers.sigmoid(conv, name=name + '_sigmoid')
    elif activation_fn == None:
        conv = conv
    else:
        raise NotImplementedError("activation: [%s] is not support" %
                                  activation_fn)

    return conv


def linear(input,
           output_size,
           norm=None,
           stddev=0.02,
           activation_fn=None,
           relufactor=0.2,
L
lvmengsi 已提交
309
           name="linear",
L
lvmengsi 已提交
310 311
           initial="normal",
           is_test=False):
L
lvmengsi 已提交
312 313 314

    param_attr, bias_attr = initial_type(
        name=name,
L
lvmengsi 已提交
315 316 317
        input=input,
        op_type='linear',
        fan_out=output_size,
L
lvmengsi 已提交
318 319 320 321 322
        init=initial,
        use_bias=True,
        filter_size=1,
        stddev=stddev)

L
lvmengsi 已提交
323 324 325 326 327 328 329
    linear = fluid.layers.fc(input,
                             output_size,
                             param_attr=param_attr,
                             bias_attr=bias_attr,
                             name=name)

    if norm is not None:
L
lvmengsi 已提交
330 331
        linear = norm_layer(
            input=linear, norm_type=norm, name=name + '_norm', is_test=is_test)
L
lvmengsi 已提交
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
    if activation_fn == 'relu':
        linear = fluid.layers.relu(linear, name=name + '_relu')
    elif activation_fn == 'leaky_relu':
        if relufactor == 0.0:
            raise Warning(
                "the activation is leaky_relu, but the relufactor is 0")
        linear = fluid.layers.leaky_relu(
            linear, alpha=relufactor, name=name + '_leaky_relu')
    elif activation_fn == 'tanh':
        linear = fluid.layers.tanh(linear, name=name + '_tanh')
    elif activation_fn == 'sigmoid':
        linear = fluid.layers.sigmoid(linear, name=name + '_sigmoid')
    elif activation_fn == None:
        linear = linear
    else:
        raise NotImplementedError("activation: [%s] is not support" %
                                  activation_fn)

    return linear


def conv_cond_concat(x, y):
    ones = fluid.layers.fill_constant_batch_size_like(
        x, [-1, y.shape[1], x.shape[2], x.shape[3]], "float32", 1.0)
    out = fluid.layers.concat([x, ones * y], 1)
    return out


def conv_and_pool(x, num_filters, name, stddev=0.02, act=None):
    param_attr = fluid.ParamAttr(
        name=name + '_w',
        initializer=fluid.initializer.NormalInitializer(
            loc=0.0, scale=stddev))
    bias_attr = fluid.ParamAttr(
        name=name + "_b", initializer=fluid.initializer.Constant(0.0))

    out = fluid.nets.simple_img_conv_pool(
        input=x,
        filter_size=5,
        num_filters=num_filters,
        pool_size=2,
        pool_stride=2,
        param_attr=param_attr,
        bias_attr=bias_attr,
        act=act)
    return out