#Copyright (c) 2020 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. import math import paddle import paddle.nn as nn from paddle.nn import Conv2D, AvgPool2D, MaxPool2D, BatchNorm, Linear __all__ = ['InceptionV3'] class InceptionV3(nn.Layer): DEFAULT_BLOCK_INDEX = 3 BLOCK_INDEX_BY_DIM = { 64: 0, # First max pooling features 192: 1, # Second max pooling featurs 768: 2, # Pre-aux classifier features 2048: 3 # Final average pooling features } def __init__(self, output_blocks=[DEFAULT_BLOCK_INDEX], class_dim=1000, aux_logits=False, resize_input=True, normalize_input=True): super(InceptionV3, self).__init__() self.resize_input = resize_input self.normalize_input = normalize_input self.output_blocks = sorted(output_blocks) self.last_needed_block = max(output_blocks) self.class_dim = class_dim self.aux_logits = aux_logits assert self.last_needed_block <= 3, 'Last possible output block index is 3' self.blocks = [] self.Conv2d_1a_3x3 = ConvBNLayer(3, 32, 3, stride=2, name='Conv2d_1a_3x3') self.Conv2d_2a_3x3 = ConvBNLayer(32, 32, 3, name='Conv2d_2a_3x3') self.Conv2d_2b_3x3 = ConvBNLayer(32, 64, 3, padding=1, name='Conv2d_2b_3x3') self.maxpool1 = MaxPool2D(pool_size=3, pool_stride=2) block0 = [ self.Conv2d_1a_3x3, self.Conv2d_2a_3x3, self.Conv2d_2b_3x3, self.maxpool1 ] self.blocks.append(nn.Sequential(*block0)) ### block1 if self.last_needed_block >= 1: self.Conv2d_3b_1x1 = ConvBNLayer(64, 80, 1, name='Conv2d_3b_1x1') self.Conv2d_4a_3x3 = ConvBNLayer(80, 192, 3, name='Conv2d_4a_3x3') self.maxpool2 = MaxPool2D(pool_size=3, pool_stride=2) block1 = [self.Conv2d_3b_1x1, self.Conv2d_4a_3x3, self.maxpool2] self.blocks.append(nn.Sequential(*block1)) ### block2 ### Mixed_5b 5c 5d if self.last_needed_block >= 2: self.Mixed_5b = Fid_inceptionA(192, pool_features=32, name='Mixed_5b') self.Mixed_5c = Fid_inceptionA(256, pool_features=64, name='Mixed_5c') self.Mixed_5d = Fid_inceptionA(288, pool_features=64, name='Mixed_5d') ### Mixed_6 self.Mixed_6a = InceptionB(288, name='Mixed_6a') self.Mixed_6b = Fid_inceptionC(768, c7=128, name='Mixed_6b') self.Mixed_6c = Fid_inceptionC(768, c7=160, name='Mixed_6c') self.Mixed_6d = Fid_inceptionC(768, c7=160, name='Mixed_6d') self.Mixed_6e = Fid_inceptionC(768, c7=192, name='Mixed_6e') block2 = [ self.Mixed_5b, self.Mixed_5c, self.Mixed_5d, self.Mixed_6a, self.Mixed_6b, self.Mixed_6c, self.Mixed_6d, self.Mixed_6e ] self.blocks.append(nn.Sequential(*block2)) if self.aux_logits: self.AuxLogits = InceptionAux(768, self.class_dim, name='AuxLogits') ### block3 ### Mixed_7 if self.last_needed_block >= 3: self.Mixed_7a = InceptionD(768, name='Mixed_7a') self.Mixed_7b = Fid_inceptionE_1(1280, name='Mixed_7b') self.Mixed_7c = Fid_inceptionE_2(2048, name='Mixed_7c') self.avgpool = AvgPool2D(global_pooling=True) block3 = [self.Mixed_7a, self.Mixed_7b, self.Mixed_7c, self.avgpool] self.blocks.append(nn.Sequential(*block3)) def forward(self, x): out = [] aux = None if self.resize_input: x = nn.functional.interpolate(x, size=[299, 299], mode='bilinear', align_corners=False, align_mode=0) if self.normalize_input: x = x * 2 - 1 for idx, block in enumerate(self.blocks): x = block(x) if self.aux_logits and (idx == 2): aux = self.AuxLogits(x) if idx in self.output_blocks: out.append(x) if idx == self.last_needed_block: break return out, aux class InceptionA(nn.Layer): def __init__(self, in_channels, pool_features, name=None): super(InceptionA, self).__init__() self.branch1x1 = ConvBNLayer(in_channels, 64, 1, name=name + '.branch1x1') self.branch5x5_1 = ConvBNLayer(in_channels, 48, 1, name=name + '.branch5x5_1') self.branch5x5_2 = ConvBNLayer(48, 64, 5, padding=2, name=name + '.branch5x5_2') self.branch3x3dbl_1 = ConvBNLayer(in_channels, 64, 1, name=name + '.branch3x3dbl_1') self.branch3x3dbl_2 = ConvBNLayer(64, 96, 3, padding=1, name=name + '.branch3x3dbl_2') self.branch3x3dbl_3 = ConvBNLayer(96, 96, 3, padding=1, name=name + '.branch3x3dbl_3') self.branch_pool0 = AvgPool2D(pool_size=3, pool_stride=1, pool_padding=1, exclusive=True) self.branch_pool = ConvBNLayer(in_channels, pool_features, 1, name=name + '.branch_pool') def forward(self, x): branch1x1 = self.branch1x1(x) branch5x5 = self.branch5x5_1(x) branch5x5 = self.branch5x5_2(branch5x5) branch3x3dbl = self.branch3x3dbl_1(x) branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) branch3x3dbl = self.branch3x3dbl_3(branch3x3dbl) branch_pool = self.branch_pool0(x) branch_pool = self.branch_pool(branch_pool) return paddle.concat( [branch1x1, branch5x5, branch3x3dbl, branch_pool], axis=1) class InceptionB(nn.Layer): def __init__(self, in_channels, name=None): super(InceptionB, self).__init__() self.branch3x3 = ConvBNLayer(in_channels, 384, 3, stride=2, name=name + '.branch3x3') self.branch3x3dbl_1 = ConvBNLayer(in_channels, 64, 1, name=name + '.branch3x3dbl_1') self.branch3x3dbl_2 = ConvBNLayer(64, 96, 3, padding=1, name=name + '.branch3x3dbl_2') self.branch3x3dbl_3 = ConvBNLayer(96, 96, 3, stride=2, name=name + '.branch3x3dbl_3') self.branch_pool = MaxPool2D(pool_size=3, pool_stride=2) def forward(self, x): branch3x3 = self.branch3x3(x) branch3x3dbl = self.branch3x3dbl_1(x) branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) branch3x3dbl = self.branch3x3dbl_3(branch3x3dbl) branch_pool = self.branch_pool(x) return paddle.concat([branch3x3, branch3x3dbl, branch_pool], axis=1) class InceptionC(nn.Layer): def __init__(self, in_channels, c7, name=None): super(InceptionC, self).__init__() self.branch1x1 = ConvBNLayer(in_channels, 192, 1, name=name + '.branch1x1') self.branch7x7_1 = ConvBNLayer(in_channels, c7, 1, name=name + '.branch7x7_1') self.branch7x7_2 = ConvBNLayer(c7, c7, (1, 7), padding=(0, 3), name=name + '.branch7x7_2') self.branch7x7_3 = ConvBNLayer(c7, 192, (7, 1), padding=(3, 0), name=name + '.branch7x7_3') self.branch7x7dbl_1 = ConvBNLayer(in_channels, c7, 1, name=name + '.branch7x7dbl_1') self.branch7x7dbl_2 = ConvBNLayer(c7, c7, (7, 1), padding=(3, 0), name=name + '.branch7x7dbl_2') self.branch7x7dbl_3 = ConvBNLayer(c7, c7, (1, 7), padding=(0, 3), name=name + '.branch7x7dbl_3') self.branch7x7dbl_4 = ConvBNLayer(c7, c7, (7, 1), padding=(3, 0), name=name + '.branch7x7dbl_4') self.branch7x7dbl_5 = ConvBNLayer(c7, 192, (1, 7), padding=(0, 3), name=name + '.branch7x7dbl_5') self.branch_pool0 = AvgPool2D(pool_size=3, pool_stride=1, pool_padding=1, exclusive=True) self.branch_pool = ConvBNLayer(in_channels, 192, 1, name=name + '.branch_pool') def forward(self, x): branch1x1 = self.branch1x1(x) branch7x7 = self.branch7x7_1(x) branch7x7 = self.branch7x7_2(branch7x7) branch7x7 = self.branch7x7_3(branch7x7) branch7x7dbl = self.branch7x7dbl_1(x) branch7x7dbl = self.branch7x7dbl_2(branch7x7dbl) branch7x7dbl = self.branch7x7dbl_3(branch7x7dbl) branch7x7dbl = self.branch7x7dbl_4(branch7x7dbl) branch7x7dbl = self.branch7x7dbl_5(branch7x7dbl) branch_pool = self.branch_pool0(x) branch_pool = self.branch_pool(branch_pool) return paddle.concat( [branch1x1, branch7x7, branch7x7dbl, branch_pool], axis=1) class InceptionD(nn.Layer): def __init__(self, in_channels, name=None): super(InceptionD, self).__init__() self.branch3x3_1 = ConvBNLayer(in_channels, 192, 1, name=name + '.branch3x3_1') self.branch3x3_2 = ConvBNLayer(192, 320, 3, stride=2, name=name + '.branch3x3_2') self.branch7x7x3_1 = ConvBNLayer(in_channels, 192, 1, name=name + '.branch7x7x3_1') self.branch7x7x3_2 = ConvBNLayer(192, 192, (1, 7), padding=(0, 3), name=name + '.branch7x7x3_2') self.branch7x7x3_3 = ConvBNLayer(192, 192, (7, 1), padding=(3, 0), name=name + '.branch7x7x3_3') self.branch7x7x3_4 = ConvBNLayer(192, 192, 3, stride=2, name=name + '.branch7x7x3_4') self.branch_pool = MaxPool2D(pool_size=3, pool_stride=2) def forward(self, x): branch3x3 = self.branch3x3_1(x) branch3x3 = self.branch3x3_2(branch3x3) branch7x7x3 = self.branch7x7x3_1(x) branch7x7x3 = self.branch7x7x3_2(branch7x7x3) branch7x7x3 = self.branch7x7x3_3(branch7x7x3) branch7x7x3 = self.branch7x7x3_4(branch7x7x3) branch_pool = self.branch_pool(x) return paddle.concat([branch3x3, branch7x7x3, branch_pool], axis=1) class InceptionE(nn.Layer): def __init__(self, in_channels, name=None): super(InceptionE, self).__init__() self.branch1x1 = ConvBNLayer(in_channels, 320, 1, name=name + '.branch1x1') self.branch3x3_1 = ConvBNLayer(in_channels, 384, 1, name=name + '.branch3x3_1') self.branch3x3_2a = ConvBNLayer(384, 384, (1, 3), padding=(0, 1), name=name + '.branch3x3_2a') self.branch3x3_2b = ConvBNLayer(384, 384, (3, 1), padding=(1, 0), name=name + '.branch3x3_2b') self.branch3x3dbl_1 = ConvBNLayer(in_channels, 448, 1, name=name + '.branch3x3dbl_1') self.branch3x3dbl_2 = ConvBNLayer(448, 384, 3, padding=1, name=name + '.branch3x3dbl_2') self.branch3x3dbl_3a = ConvBNLayer(384, 384, (1, 3), padding=(0, 1), name=name + '.branch3x3dbl_3a') self.branch3x3dbl_3b = ConvBNLayer(384, 384, (3, 1), padding=(1, 0), name=name + '.branch3x3dbl_3b') self.branch_pool0 = AvgPool2D(pool_size=3, pool_stride=1, pool_padding=1, exclusive=True) self.branch_pool = ConvBNLayer(in_channels, 192, 1, name=name + '.branch_pool') def forward(self, x): branch1x1 = self.branch1x1(x) branch3x3_1 = self.branch3x3_1(x) branch3x3_2a = self.branch3x3_2a(branch3x3_1) branch3x3_2b = self.branch3x3_2b(branch3x3_1) branch3x3 = paddle.concat([branch3x3_2a, branch3x3_2b], axis=1) branch3x3dbl = self.branch3x3dbl_1(x) branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) branch3x3dbl_3a = self.branch3x3dbl_3a(branch3x3dbl) branch3x3dbl_3b = self.branch3x3dbl_3b(branch3x3dbl) branch3x3dbl = paddle.concat([branch3x3dbl_3a, branch3x3dbl_3b], axis=1) branch_pool = self.branch_pool0(x) branch_pool = self.branch_pool(branch_pool) return paddle.concat( [branch1x1, branch3x3, branch3x3dbl, branch_pool], axis=1) class InceptionAux(nn.Layer): def __init__(self, in_channels, num_classes, name=None): super(InceptionAux, self).__init__() self.num_classes = num_classes self.pool0 = AvgPool2D(pool_size=5, pool_stride=3) self.conv0 = ConvBNLayer(in_channels, 128, 1, name=name + '.conv0') self.conv1 = ConvBNLayer(128, 768, 5, name=name + '.conv1') self.pool1 = AvgPool2D(global_pooling=True) def forward(self, x): x = self.pool0(x) x = self.conv0(x) x = self.conv1(x) x = self.pool1(x) x = paddle.flatten(x, axis=1) x = paddle.static.nn.fc(x, size=self.num_classes) return x class Fid_inceptionA(nn.Layer): """ FID block in inception v3 """ def __init__(self, in_channels, pool_features, name=None): super(Fid_inceptionA, self).__init__() self.branch1x1 = ConvBNLayer(in_channels, 64, 1, name=name + '.branch1x1') self.branch5x5_1 = ConvBNLayer(in_channels, 48, 1, name=name + '.branch5x5_1') self.branch5x5_2 = ConvBNLayer(48, 64, 5, padding=2, name=name + '.branch5x5_2') self.branch3x3dbl_1 = ConvBNLayer(in_channels, 64, 1, name=name + '.branch3x3dbl_1') self.branch3x3dbl_2 = ConvBNLayer(64, 96, 3, padding=1, name=name + '.branch3x3dbl_2') self.branch3x3dbl_3 = ConvBNLayer(96, 96, 3, padding=1, name=name + '.branch3x3dbl_3') self.branch_pool0 = AvgPool2D(pool_size=3, pool_stride=1, pool_padding=1, exclusive=True) self.branch_pool = ConvBNLayer(in_channels, pool_features, 1, name=name + '.branch_pool') def forward(self, x): branch1x1 = self.branch1x1(x) branch5x5 = self.branch5x5_1(x) branch5x5 = self.branch5x5_2(branch5x5) branch3x3dbl = self.branch3x3dbl_1(x) branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) branch3x3dbl = self.branch3x3dbl_3(branch3x3dbl) branch_pool = self.branch_pool0(x) branch_pool = self.branch_pool(branch_pool) return paddle.concat( [branch1x1, branch5x5, branch3x3dbl, branch_pool], axis=1) class Fid_inceptionC(nn.Layer): """ FID block in inception v3 """ def __init__(self, in_channels, c7, name=None): super(Fid_inceptionC, self).__init__() self.branch1x1 = ConvBNLayer(in_channels, 192, 1, name=name + '.branch1x1') self.branch7x7_1 = ConvBNLayer(in_channels, c7, 1, name=name + '.branch7x7_1') self.branch7x7_2 = ConvBNLayer(c7, c7, (1, 7), padding=(0, 3), name=name + '.branch7x7_2') self.branch7x7_3 = ConvBNLayer(c7, 192, (7, 1), padding=(3, 0), name=name + '.branch7x7_3') self.branch7x7dbl_1 = ConvBNLayer(in_channels, c7, 1, name=name + '.branch7x7dbl_1') self.branch7x7dbl_2 = ConvBNLayer(c7, c7, (7, 1), padding=(3, 0), name=name + '.branch7x7dbl_2') self.branch7x7dbl_3 = ConvBNLayer(c7, c7, (1, 7), padding=(0, 3), name=name + '.branch7x7dbl_3') self.branch7x7dbl_4 = ConvBNLayer(c7, c7, (7, 1), padding=(3, 0), name=name + '.branch7x7dbl_4') self.branch7x7dbl_5 = ConvBNLayer(c7, 192, (1, 7), padding=(0, 3), name=name + '.branch7x7dbl_5') self.branch_pool0 = AvgPool2D(pool_size=3, pool_stride=1, pool_padding=1, exclusive=True) self.branch_pool = ConvBNLayer(in_channels, 192, 1, name=name + '.branch_pool') def forward(self, x): branch1x1 = self.branch1x1(x) branch7x7 = self.branch7x7_1(x) branch7x7 = self.branch7x7_2(branch7x7) branch7x7 = self.branch7x7_3(branch7x7) branch7x7dbl = self.branch7x7dbl_1(x) branch7x7dbl = self.branch7x7dbl_2(branch7x7dbl) branch7x7dbl = self.branch7x7dbl_3(branch7x7dbl) branch7x7dbl = self.branch7x7dbl_4(branch7x7dbl) branch7x7dbl = self.branch7x7dbl_5(branch7x7dbl) branch_pool = self.branch_pool0(x) branch_pool = self.branch_pool(branch_pool) return paddle.concat( [branch1x1, branch7x7, branch7x7dbl, branch_pool], axis=1) class Fid_inceptionE_1(nn.Layer): """ FID block in inception v3 """ def __init__(self, in_channels, name=None): super(Fid_inceptionE_1, self).__init__() self.branch1x1 = ConvBNLayer(in_channels, 320, 1, name=name + '.branch1x1') self.branch3x3_1 = ConvBNLayer(in_channels, 384, 1, name=name + '.branch3x3_1') self.branch3x3_2a = ConvBNLayer(384, 384, (1, 3), padding=(0, 1), name=name + '.branch3x3_2a') self.branch3x3_2b = ConvBNLayer(384, 384, (3, 1), padding=(1, 0), name=name + '.branch3x3_2b') self.branch3x3dbl_1 = ConvBNLayer(in_channels, 448, 1, name=name + '.branch3x3dbl_1') self.branch3x3dbl_2 = ConvBNLayer(448, 384, 3, padding=1, name=name + '.branch3x3dbl_2') self.branch3x3dbl_3a = ConvBNLayer(384, 384, (1, 3), padding=(0, 1), name=name + '.branch3x3dbl_3a') self.branch3x3dbl_3b = ConvBNLayer(384, 384, (3, 1), padding=(1, 0), name=name + '.branch3x3dbl_3b') self.branch_pool0 = AvgPool2D(pool_size=3, pool_stride=1, pool_padding=1, exclusive=True) self.branch_pool = ConvBNLayer(in_channels, 192, 1, name=name + '.branch_pool') def forward(self, x): branch1x1 = self.branch1x1(x) branch3x3_1 = self.branch3x3_1(x) branch3x3_2a = self.branch3x3_2a(branch3x3_1) branch3x3_2b = self.branch3x3_2b(branch3x3_1) branch3x3 = paddle.concat([branch3x3_2a, branch3x3_2b], axis=1) branch3x3dbl = self.branch3x3dbl_1(x) branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) branch3x3dbl_3a = self.branch3x3dbl_3a(branch3x3dbl) branch3x3dbl_3b = self.branch3x3dbl_3b(branch3x3dbl) branch3x3dbl = paddle.concat([branch3x3dbl_3a, branch3x3dbl_3b], axis=1) branch_pool = self.branch_pool0(x) branch_pool = self.branch_pool(branch_pool) return paddle.concat( [branch1x1, branch3x3, branch3x3dbl, branch_pool], axis=1) class Fid_inceptionE_2(nn.Layer): """ FID block in inception v3 """ def __init__(self, in_channels, name=None): super(Fid_inceptionE_2, self).__init__() self.branch1x1 = ConvBNLayer(in_channels, 320, 1, name=name + '.branch1x1') self.branch3x3_1 = ConvBNLayer(in_channels, 384, 1, name=name + '.branch3x3_1') self.branch3x3_2a = ConvBNLayer(384, 384, (1, 3), padding=(0, 1), name=name + '.branch3x3_2a') self.branch3x3_2b = ConvBNLayer(384, 384, (3, 1), padding=(1, 0), name=name + '.branch3x3_2b') self.branch3x3dbl_1 = ConvBNLayer(in_channels, 448, 1, name=name + '.branch3x3dbl_1') self.branch3x3dbl_2 = ConvBNLayer(448, 384, 3, padding=1, name=name + '.branch3x3dbl_2') self.branch3x3dbl_3a = ConvBNLayer(384, 384, (1, 3), padding=(0, 1), name=name + '.branch3x3dbl_3a') self.branch3x3dbl_3b = ConvBNLayer(384, 384, (3, 1), padding=(1, 0), name=name + '.branch3x3dbl_3b') ### same with paper self.branch_pool0 = MaxPool2D(pool_size=3, pool_stride=1, pool_padding=1) self.branch_pool = ConvBNLayer(in_channels, 192, 1, name=name + '.branch_pool') def forward(self, x): branch1x1 = self.branch1x1(x) branch3x3_1 = self.branch3x3_1(x) branch3x3_2a = self.branch3x3_2a(branch3x3_1) branch3x3_2b = self.branch3x3_2b(branch3x3_1) branch3x3 = paddle.concat([branch3x3_2a, branch3x3_2b], axis=1) branch3x3dbl = self.branch3x3dbl_1(x) branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) branch3x3dbl_3a = self.branch3x3dbl_3a(branch3x3dbl) branch3x3dbl_3b = self.branch3x3dbl_3b(branch3x3dbl) branch3x3dbl = paddle.concat([branch3x3dbl_3a, branch3x3dbl_3b], axis=1) branch_pool = self.branch_pool0(x) branch_pool = self.branch_pool(branch_pool) return paddle.concat( [branch1x1, branch3x3, branch3x3dbl, branch_pool], axis=1) class ConvBNLayer(nn.Layer): def __init__(self, in_channels, num_filters, filter_size, stride=1, padding=0, groups=1, act='relu', name=None): super(ConvBNLayer, self).__init__() self.conv = Conv2D(num_channels=in_channels, num_filters=num_filters, filter_size=filter_size, stride=stride, padding=padding, groups=groups, act=None, param_attr=paddle.ParamAttr(name=name + ".conv.weight"), bias_attr=False) self.bn = BatchNorm(num_filters, act=act, epsilon=0.001, param_attr=paddle.ParamAttr(name=name + ".bn.weight"), bias_attr=paddle.ParamAttr(name=name + ".bn.bias"), moving_mean_name=name + '.bn.running_mean', moving_variance_name=name + '.bn.running_var') def forward(self, inputs): y = self.conv(inputs) y = self.bn(y) return y