Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 1 年 前同步成功

通知 2298
Star 20931
Fork 5422
P
Paddle

体验新版 GitCode，发现更多精彩内容 >>

未验证 提交 f6219dda 编写于 作者: N Nyakku Shigure 提交者： GitHub
浏览文件
操作

reuse ConvNormActivation in some vision models (#40431) 

* reuse ConvNormActivation in some vision models
上级 34ac7b74
隐藏空白更改
内联 并排
Showing with 372 addition and 382 deletion
python/paddle/vision/models/inceptionv3.py
浏览文件 @ f6219dda
...@@ -19,75 +19,60 @@ from __future__ import print_function ...@@ -19,75 +19,60 @@ from __future__ import print_function
import math import math
import paddle import paddle
import paddle.nn as nn import paddle.nn as nn
from paddle.nn import Conv2D, BatchNorm, Linear, Dropout from paddle.nn import Linear, Dropout
from paddle.nn import AdaptiveAvgPool2D, MaxPool2D, AvgPool2D from paddle.nn import AdaptiveAvgPool2D, MaxPool2D, AvgPool2D
from paddle.nn.initializer import Uniform from paddle.nn.initializer import Uniform
from paddle.fluid.param_attr import ParamAttr from paddle.fluid.param_attr import ParamAttr
from paddle.utils.download import get_weights_path_from_url from paddle.utils.download import get_weights_path_from_url
from ..ops import ConvNormActivation
__all__ = [] __all__ = []
model_urls = { model_urls = {
"inception_v3": "inception_v3":
("https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/legendary_models/InceptionV3_pretrained.pdparams", ("https://paddle-hapi.bj.bcebos.com/models/inception_v3.pdparams",
"e4d0905a818f6bb7946e881777a8a935") "649a4547c3243e8b59c656f41fe330b8")
} }
class ConvBNLayer(nn.Layer):
def __init__(self,
num_channels,
num_filters,
filter_size,
stride=1,
padding=0,
groups=1,
act="relu"):
super().__init__()
self.act = act
self.conv = Conv2D(
in_channels=num_channels,
out_channels=num_filters,
kernel_size=filter_size,
stride=stride,
padding=padding,
groups=groups,
bias_attr=False)
self.bn = BatchNorm(num_filters)
self.relu = nn.ReLU()
def forward(self, x):
x = self.conv(x)
x = self.bn(x)
if self.act:
x = self.relu(x)
return x
class InceptionStem(nn.Layer): class InceptionStem(nn.Layer):
def __init__(self): def __init__(self):
super().__init__() super().__init__()
self.conv_1a_3x3 = ConvBNLayer( self.conv_1a_3x3 = ConvNormActivation(
num_channels=3, num_filters=32, filter_size=3, stride=2, act="relu") in_channels=3,
self.conv_2a_3x3 = ConvBNLayer( out_channels=32,
num_channels=32, kernel_size=3,
num_filters=32, stride=2,
filter_size=3, padding=0,
activation_layer=nn.ReLU)
self.conv_2a_3x3 = ConvNormActivation(
in_channels=32,
out_channels=32,
kernel_size=3,
stride=1, stride=1,
act="relu") padding=0,
self.conv_2b_3x3 = ConvBNLayer( activation_layer=nn.ReLU)
num_channels=32, self.conv_2b_3x3 = ConvNormActivation(
num_filters=64, in_channels=32,
filter_size=3, out_channels=64,
kernel_size=3,
padding=1, padding=1,
act="relu") activation_layer=nn.ReLU)
self.max_pool = MaxPool2D(kernel_size=3, stride=2, padding=0) self.max_pool = MaxPool2D(kernel_size=3, stride=2, padding=0)
self.conv_3b_1x1 = ConvBNLayer( self.conv_3b_1x1 = ConvNormActivation(
num_channels=64, num_filters=80, filter_size=1, act="relu") in_channels=64,
self.conv_4a_3x3 = ConvBNLayer( out_channels=80,
num_channels=80, num_filters=192, filter_size=3, act="relu") kernel_size=1,
padding=0,
activation_layer=nn.ReLU)
self.conv_4a_3x3 = ConvNormActivation(
in_channels=80,
out_channels=192,
kernel_size=3,
padding=0,
activation_layer=nn.ReLU)
def forward(self, x): def forward(self, x):
x = self.conv_1a_3x3(x) x = self.conv_1a_3x3(x)
...@@ -103,47 +88,53 @@ class InceptionStem(nn.Layer): ...@@ -103,47 +88,53 @@ class InceptionStem(nn.Layer):
class InceptionA(nn.Layer): class InceptionA(nn.Layer):
def __init__(self, num_channels, pool_features): def __init__(self, num_channels, pool_features):
super().__init__() super().__init__()
self.branch1x1 = ConvBNLayer( self.branch1x1 = ConvNormActivation(
num_channels=num_channels, in_channels=num_channels,
num_filters=64, out_channels=64,
filter_size=1, kernel_size=1,
act="relu") padding=0,
self.branch5x5_1 = ConvBNLayer( activation_layer=nn.ReLU)
num_channels=num_channels,
num_filters=48, self.branch5x5_1 = ConvNormActivation(
filter_size=1, in_channels=num_channels,
act="relu") out_channels=48,
self.branch5x5_2 = ConvBNLayer( kernel_size=1,
num_channels=48, padding=0,
num_filters=64, activation_layer=nn.ReLU)
filter_size=5, self.branch5x5_2 = ConvNormActivation(
in_channels=48,
out_channels=64,
kernel_size=5,
padding=2, padding=2,
act="relu") activation_layer=nn.ReLU)
self.branch3x3dbl_1 = ConvBNLayer( self.branch3x3dbl_1 = ConvNormActivation(
num_channels=num_channels, in_channels=num_channels,
num_filters=64, out_channels=64,
filter_size=1, kernel_size=1,
act="relu") padding=0,
self.branch3x3dbl_2 = ConvBNLayer( activation_layer=nn.ReLU)
num_channels=64, self.branch3x3dbl_2 = ConvNormActivation(
num_filters=96, in_channels=64,
filter_size=3, out_channels=96,
kernel_size=3,
padding=1, padding=1,
act="relu") activation_layer=nn.ReLU)
self.branch3x3dbl_3 = ConvBNLayer( self.branch3x3dbl_3 = ConvNormActivation(
num_channels=96, in_channels=96,
num_filters=96, out_channels=96,
filter_size=3, kernel_size=3,
padding=1, padding=1,
act="relu") activation_layer=nn.ReLU)
self.branch_pool = AvgPool2D( self.branch_pool = AvgPool2D(
kernel_size=3, stride=1, padding=1, exclusive=False) kernel_size=3, stride=1, padding=1, exclusive=False)
self.branch_pool_conv = ConvBNLayer( self.branch_pool_conv = ConvNormActivation(
num_channels=num_channels, in_channels=num_channels,
num_filters=pool_features, out_channels=pool_features,
filter_size=1, kernel_size=1,
act="relu") padding=0,
activation_layer=nn.ReLU)
def forward(self, x): def forward(self, x):
branch1x1 = self.branch1x1(x) branch1x1 = self.branch1x1(x)
...@@ -164,29 +155,34 @@ class InceptionA(nn.Layer): ...@@ -164,29 +155,34 @@ class InceptionA(nn.Layer):
class InceptionB(nn.Layer): class InceptionB(nn.Layer):
def __init__(self, num_channels): def __init__(self, num_channels):
super().__init__() super().__init__()
self.branch3x3 = ConvBNLayer( self.branch3x3 = ConvNormActivation(
num_channels=num_channels, in_channels=num_channels,
num_filters=384, out_channels=384,
filter_size=3, kernel_size=3,
stride=2, stride=2,
act="relu") padding=0,
self.branch3x3dbl_1 = ConvBNLayer( activation_layer=nn.ReLU)
num_channels=num_channels,
num_filters=64, self.branch3x3dbl_1 = ConvNormActivation(
filter_size=1, in_channels=num_channels,
act="relu") out_channels=64,
self.branch3x3dbl_2 = ConvBNLayer( kernel_size=1,
num_channels=64, padding=0,
num_filters=96, activation_layer=nn.ReLU)
filter_size=3, self.branch3x3dbl_2 = ConvNormActivation(
in_channels=64,
out_channels=96,
kernel_size=3,
padding=1, padding=1,
act="relu") activation_layer=nn.ReLU)
self.branch3x3dbl_3 = ConvBNLayer( self.branch3x3dbl_3 = ConvNormActivation(
num_channels=96, in_channels=96,
num_filters=96, out_channels=96,
filter_size=3, kernel_size=3,
stride=2, stride=2,
act="relu") padding=0,
activation_layer=nn.ReLU)
self.branch_pool = MaxPool2D(kernel_size=3, stride=2) self.branch_pool = MaxPool2D(kernel_size=3, stride=2)
def forward(self, x): def forward(self, x):
...@@ -206,70 +202,74 @@ class InceptionB(nn.Layer): ...@@ -206,70 +202,74 @@ class InceptionB(nn.Layer):
class InceptionC(nn.Layer): class InceptionC(nn.Layer):
def __init__(self, num_channels, channels_7x7): def __init__(self, num_channels, channels_7x7):
super().__init__() super().__init__()
self.branch1x1 = ConvBNLayer( self.branch1x1 = ConvNormActivation(
num_channels=num_channels, in_channels=num_channels,
num_filters=192, out_channels=192,
filter_size=1, kernel_size=1,
act="relu") padding=0,
activation_layer=nn.ReLU)
self.branch7x7_1 = ConvBNLayer(
num_channels=num_channels, self.branch7x7_1 = ConvNormActivation(
num_filters=channels_7x7, in_channels=num_channels,
filter_size=1, out_channels=channels_7x7,
kernel_size=1,
stride=1, stride=1,
act="relu") padding=0,
self.branch7x7_2 = ConvBNLayer( activation_layer=nn.ReLU)
num_channels=channels_7x7, self.branch7x7_2 = ConvNormActivation(
num_filters=channels_7x7, in_channels=channels_7x7,
filter_size=(1, 7), out_channels=channels_7x7,
kernel_size=(1, 7),
stride=1, stride=1,
padding=(0, 3), padding=(0, 3),
act="relu") activation_layer=nn.ReLU)
self.branch7x7_3 = ConvBNLayer( self.branch7x7_3 = ConvNormActivation(
num_channels=channels_7x7, in_channels=channels_7x7,
num_filters=192, out_channels=192,
filter_size=(7, 1), kernel_size=(7, 1),
stride=1, stride=1,
padding=(3, 0), padding=(3, 0),
act="relu") activation_layer=nn.ReLU)
self.branch7x7dbl_1 = ConvBNLayer( self.branch7x7dbl_1 = ConvNormActivation(
num_channels=num_channels, in_channels=num_channels,
num_filters=channels_7x7, out_channels=channels_7x7,
filter_size=1, kernel_size=1,
act="relu") padding=0,
self.branch7x7dbl_2 = ConvBNLayer( activation_layer=nn.ReLU)
num_channels=channels_7x7, self.branch7x7dbl_2 = ConvNormActivation(
num_filters=channels_7x7, in_channels=channels_7x7,
filter_size=(7, 1), out_channels=channels_7x7,
kernel_size=(7, 1),
padding=(3, 0), padding=(3, 0),
act="relu") activation_layer=nn.ReLU)
self.branch7x7dbl_3 = ConvBNLayer( self.branch7x7dbl_3 = ConvNormActivation(
num_channels=channels_7x7, in_channels=channels_7x7,
num_filters=channels_7x7, out_channels=channels_7x7,
filter_size=(1, 7), kernel_size=(1, 7),
padding=(0, 3), padding=(0, 3),
act="relu") activation_layer=nn.ReLU)
self.branch7x7dbl_4 = ConvBNLayer( self.branch7x7dbl_4 = ConvNormActivation(
num_channels=channels_7x7, in_channels=channels_7x7,
num_filters=channels_7x7, out_channels=channels_7x7,
filter_size=(7, 1), kernel_size=(7, 1),
padding=(3, 0), padding=(3, 0),
act="relu") activation_layer=nn.ReLU)
self.branch7x7dbl_5 = ConvBNLayer( self.branch7x7dbl_5 = ConvNormActivation(
num_channels=channels_7x7, in_channels=channels_7x7,
num_filters=192, out_channels=192,
filter_size=(1, 7), kernel_size=(1, 7),
padding=(0, 3), padding=(0, 3),
act="relu") activation_layer=nn.ReLU)
self.branch_pool = AvgPool2D( self.branch_pool = AvgPool2D(
kernel_size=3, stride=1, padding=1, exclusive=False) kernel_size=3, stride=1, padding=1, exclusive=False)
self.branch_pool_conv = ConvBNLayer( self.branch_pool_conv = ConvNormActivation(
num_channels=num_channels, in_channels=num_channels,
num_filters=192, out_channels=192,
filter_size=1, kernel_size=1,
act="relu") padding=0,
activation_layer=nn.ReLU)
def forward(self, x): def forward(self, x):
branch1x1 = self.branch1x1(x) branch1x1 = self.branch1x1(x)
...@@ -296,40 +296,46 @@ class InceptionC(nn.Layer): ...@@ -296,40 +296,46 @@ class InceptionC(nn.Layer):
class InceptionD(nn.Layer): class InceptionD(nn.Layer):
def __init__(self, num_channels): def __init__(self, num_channels):
super().__init__() super().__init__()
self.branch3x3_1 = ConvBNLayer( self.branch3x3_1 = ConvNormActivation(
num_channels=num_channels, in_channels=num_channels,
num_filters=192, out_channels=192,
filter_size=1, kernel_size=1,
act="relu") padding=0,
self.branch3x3_2 = ConvBNLayer( activation_layer=nn.ReLU)
num_channels=192, self.branch3x3_2 = ConvNormActivation(
num_filters=320, in_channels=192,
filter_size=3, out_channels=320,
kernel_size=3,
stride=2, stride=2,
act="relu") padding=0,
self.branch7x7x3_1 = ConvBNLayer( activation_layer=nn.ReLU)
num_channels=num_channels,
num_filters=192, self.branch7x7x3_1 = ConvNormActivation(
filter_size=1, in_channels=num_channels,
act="relu") out_channels=192,
self.branch7x7x3_2 = ConvBNLayer( kernel_size=1,
num_channels=192, padding=0,
num_filters=192, activation_layer=nn.ReLU)
filter_size=(1, 7), self.branch7x7x3_2 = ConvNormActivation(
in_channels=192,
out_channels=192,
kernel_size=(1, 7),
padding=(0, 3), padding=(0, 3),
act="relu") activation_layer=nn.ReLU)
self.branch7x7x3_3 = ConvBNLayer( self.branch7x7x3_3 = ConvNormActivation(
num_channels=192, in_channels=192,
num_filters=192, out_channels=192,
filter_size=(7, 1), kernel_size=(7, 1),
padding=(3, 0), padding=(3, 0),
act="relu") activation_layer=nn.ReLU)
self.branch7x7x3_4 = ConvBNLayer( self.branch7x7x3_4 = ConvNormActivation(
num_channels=192, in_channels=192,
num_filters=192, out_channels=192,
filter_size=3, kernel_size=3,
stride=2, stride=2,
act="relu") padding=0,
activation_layer=nn.ReLU)
self.branch_pool = MaxPool2D(kernel_size=3, stride=2) self.branch_pool = MaxPool2D(kernel_size=3, stride=2)
def forward(self, x): def forward(self, x):
...@@ -350,59 +356,64 @@ class InceptionD(nn.Layer): ...@@ -350,59 +356,64 @@ class InceptionD(nn.Layer):
class InceptionE(nn.Layer): class InceptionE(nn.Layer):
def __init__(self, num_channels): def __init__(self, num_channels):
super().__init__() super().__init__()
self.branch1x1 = ConvBNLayer( self.branch1x1 = ConvNormActivation(
num_channels=num_channels, in_channels=num_channels,
num_filters=320, out_channels=320,
filter_size=1, kernel_size=1,
act="relu") padding=0,
self.branch3x3_1 = ConvBNLayer( activation_layer=nn.ReLU)
num_channels=num_channels, self.branch3x3_1 = ConvNormActivation(
num_filters=384, in_channels=num_channels,
filter_size=1, out_channels=384,
act="relu") kernel_size=1,
self.branch3x3_2a = ConvBNLayer( padding=0,
num_channels=384, activation_layer=nn.ReLU)
num_filters=384, self.branch3x3_2a = ConvNormActivation(
filter_size=(1, 3), in_channels=384,
out_channels=384,
kernel_size=(1, 3),
padding=(0, 1), padding=(0, 1),
act="relu") activation_layer=nn.ReLU)
self.branch3x3_2b = ConvBNLayer( self.branch3x3_2b = ConvNormActivation(
num_channels=384, in_channels=384,
num_filters=384, out_channels=384,
filter_size=(3, 1), kernel_size=(3, 1),
padding=(1, 0), padding=(1, 0),
act="relu") activation_layer=nn.ReLU)
self.branch3x3dbl_1 = ConvBNLayer( self.branch3x3dbl_1 = ConvNormActivation(
num_channels=num_channels, in_channels=num_channels,
num_filters=448, out_channels=448,
filter_size=1, kernel_size=1,
act="relu") padding=0,
self.branch3x3dbl_2 = ConvBNLayer( activation_layer=nn.ReLU)
num_channels=448, self.branch3x3dbl_2 = ConvNormActivation(
num_filters=384, in_channels=448,
filter_size=3, out_channels=384,
kernel_size=3,
padding=1, padding=1,
act="relu") activation_layer=nn.ReLU)
self.branch3x3dbl_3a = ConvBNLayer( self.branch3x3dbl_3a = ConvNormActivation(
num_channels=384, in_channels=384,
num_filters=384, out_channels=384,
filter_size=(1, 3), kernel_size=(1, 3),
padding=(0, 1), padding=(0, 1),
act="relu") activation_layer=nn.ReLU)
self.branch3x3dbl_3b = ConvBNLayer( self.branch3x3dbl_3b = ConvNormActivation(
num_channels=384, in_channels=384,
num_filters=384, out_channels=384,
filter_size=(3, 1), kernel_size=(3, 1),
padding=(1, 0), padding=(1, 0),
act="relu") activation_layer=nn.ReLU)
self.branch_pool = AvgPool2D( self.branch_pool = AvgPool2D(
kernel_size=3, stride=1, padding=1, exclusive=False) kernel_size=3, stride=1, padding=1, exclusive=False)
self.branch_pool_conv = ConvBNLayer( self.branch_pool_conv = ConvNormActivation(
num_channels=num_channels, in_channels=num_channels,
num_filters=192, out_channels=192,
filter_size=1, kernel_size=1,
act="relu") padding=0,
activation_layer=nn.ReLU)
def forward(self, x): def forward(self, x):
branch1x1 = self.branch1x1(x) branch1x1 = self.branch1x1(x)
......
python/paddle/vision/models/mobilenetv1.py
浏览文件 @ f6219dda
...@@ -16,59 +16,31 @@ import paddle ...@@ -16,59 +16,31 @@ import paddle
import paddle.nn as nn import paddle.nn as nn
from paddle.utils.download import get_weights_path_from_url from paddle.utils.download import get_weights_path_from_url
from ..ops import ConvNormActivation
__all__ = [] __all__ = []
model_urls = { model_urls = {
'mobilenetv1_1.0': 'mobilenetv1_1.0':
('https://paddle-hapi.bj.bcebos.com/models/mobilenet_v1_x1.0.pdparams', ('https://paddle-hapi.bj.bcebos.com/models/mobilenetv1_1.0.pdparams',
'42a154c2f26f86e7457d6daded114e8c') '3033ab1975b1670bef51545feb65fc45')
} }
class ConvBNLayer(nn.Layer):
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride,
padding,
num_groups=1):
super(ConvBNLayer, self).__init__()
self._conv = nn.Conv2D(
in_channels,
out_channels,
kernel_size,
stride=stride,
padding=padding,
groups=num_groups,
bias_attr=False)
self._norm_layer = nn.BatchNorm2D(out_channels)
self._act = nn.ReLU()
def forward(self, x):
x = self._conv(x)
x = self._norm_layer(x)
x = self._act(x)
return x
class DepthwiseSeparable(nn.Layer): class DepthwiseSeparable(nn.Layer):
def __init__(self, in_channels, out_channels1, out_channels2, num_groups, def __init__(self, in_channels, out_channels1, out_channels2, num_groups,
stride, scale): stride, scale):
super(DepthwiseSeparable, self).__init__() super(DepthwiseSeparable, self).__init__()
self._depthwise_conv = ConvBNLayer( self._depthwise_conv = ConvNormActivation(
in_channels, in_channels,
int(out_channels1 * scale), int(out_channels1 * scale),
kernel_size=3, kernel_size=3,
stride=stride, stride=stride,
padding=1, padding=1,
num_groups=int(num_groups * scale)) groups=int(num_groups * scale))
self._pointwise_conv = ConvBNLayer( self._pointwise_conv = ConvNormActivation(
int(out_channels1 * scale), int(out_channels1 * scale),
int(out_channels2 * scale), int(out_channels2 * scale),
kernel_size=1, kernel_size=1,
...@@ -94,9 +66,15 @@ class MobileNetV1(nn.Layer): ...@@ -94,9 +66,15 @@ class MobileNetV1(nn.Layer):
Examples: Examples:
.. code-block:: python .. code-block:: python
import paddle
from paddle.vision.models import MobileNetV1 from paddle.vision.models import MobileNetV1
model = MobileNetV1() model = MobileNetV1()
x = paddle.rand([1, 3, 224, 224])
out = model(x)
print(out.shape)
""" """
def __init__(self, scale=1.0, num_classes=1000, with_pool=True): def __init__(self, scale=1.0, num_classes=1000, with_pool=True):
...@@ -106,7 +84,7 @@ class MobileNetV1(nn.Layer): ...@@ -106,7 +84,7 @@ class MobileNetV1(nn.Layer):
self.num_classes = num_classes self.num_classes = num_classes
self.with_pool = with_pool self.with_pool = with_pool
self.conv1 = ConvBNLayer( self.conv1 = ConvNormActivation(
in_channels=3, in_channels=3,
out_channels=int(32 * scale), out_channels=int(32 * scale),
kernel_size=3, kernel_size=3,
...@@ -257,6 +235,7 @@ def mobilenet_v1(pretrained=False, scale=1.0, **kwargs): ...@@ -257,6 +235,7 @@ def mobilenet_v1(pretrained=False, scale=1.0, **kwargs):
Examples: Examples:
.. code-block:: python .. code-block:: python
import paddle
from paddle.vision.models import mobilenet_v1 from paddle.vision.models import mobilenet_v1
# build model # build model
...@@ -266,7 +245,12 @@ def mobilenet_v1(pretrained=False, scale=1.0, **kwargs): ...@@ -266,7 +245,12 @@ def mobilenet_v1(pretrained=False, scale=1.0, **kwargs):
# model = mobilenet_v1(pretrained=True) # model = mobilenet_v1(pretrained=True)
# build mobilenet v1 with scale=0.5 # build mobilenet v1 with scale=0.5
model = mobilenet_v1(scale=0.5) model_scale = mobilenet_v1(scale=0.5)
x = paddle.rand([1, 3, 224, 224])
out = model(x)
print(out.shape)
""" """
model = _mobilenet( model = _mobilenet(
'mobilenetv1_' + str(scale), pretrained, scale=scale, **kwargs) 'mobilenetv1_' + str(scale), pretrained, scale=scale, **kwargs)
......
python/paddle/vision/models/mobilenetv2.py
浏览文件 @ f6219dda
...@@ -17,6 +17,7 @@ import paddle.nn as nn ...@@ -17,6 +17,7 @@ import paddle.nn as nn
from paddle.utils.download import get_weights_path_from_url from paddle.utils.download import get_weights_path_from_url
from .utils import _make_divisible from .utils import _make_divisible
from ..ops import ConvNormActivation
__all__ = [] __all__ = []
...@@ -27,29 +28,6 @@ model_urls = { ...@@ -27,29 +28,6 @@ model_urls = {
} }
class ConvBNReLU(nn.Sequential):
def __init__(self,
in_planes,
out_planes,
kernel_size=3,
stride=1,
groups=1,
norm_layer=nn.BatchNorm2D):
padding = (kernel_size - 1) // 2
super(ConvBNReLU, self).__init__(
nn.Conv2D(
in_planes,
out_planes,
kernel_size,
stride,
padding,
groups=groups,
bias_attr=False),
norm_layer(out_planes),
nn.ReLU6())
class InvertedResidual(nn.Layer): class InvertedResidual(nn.Layer):
def __init__(self, def __init__(self,
inp, inp,
...@@ -67,15 +45,20 @@ class InvertedResidual(nn.Layer): ...@@ -67,15 +45,20 @@ class InvertedResidual(nn.Layer):
layers = [] layers = []
if expand_ratio != 1: if expand_ratio != 1:
layers.append( layers.append(
ConvBNReLU( ConvNormActivation(
inp, hidden_dim, kernel_size=1, norm_layer=norm_layer)) inp,
hidden_dim,
kernel_size=1,
norm_layer=norm_layer,
activation_layer=nn.ReLU6))
layers.extend([ layers.extend([
ConvBNReLU( ConvNormActivation(
hidden_dim, hidden_dim,
hidden_dim, hidden_dim,
stride=stride, stride=stride,
groups=hidden_dim, groups=hidden_dim,
norm_layer=norm_layer), norm_layer=norm_layer,
activation_layer=nn.ReLU6),
nn.Conv2D( nn.Conv2D(
hidden_dim, oup, 1, 1, 0, bias_attr=False), hidden_dim, oup, 1, 1, 0, bias_attr=False),
norm_layer(oup), norm_layer(oup),
...@@ -90,23 +73,30 @@ class InvertedResidual(nn.Layer): ...@@ -90,23 +73,30 @@ class InvertedResidual(nn.Layer):
class MobileNetV2(nn.Layer): class MobileNetV2(nn.Layer):
def __init__(self, scale=1.0, num_classes=1000, with_pool=True): """MobileNetV2 model from
"""MobileNetV2 model from `"MobileNetV2: Inverted Residuals and Linear Bottlenecks" <https://arxiv.org/abs/1801.04381>`_.
`"MobileNetV2: Inverted Residuals and Linear Bottlenecks" <https://arxiv.org/abs/1801.04381>`_.
Args:
scale (float): scale of channels in each layer. Default: 1.0.
num_classes (int): output dim of last fc layer. If num_classes <=0, last fc layer
will not be defined. Default: 1000.
with_pool (bool): use pool before the last fc layer or not. Default: True.
Args: Examples:
scale (float): scale of channels in each layer. Default: 1.0. .. code-block:: python
num_classes (int): output dim of last fc layer. If num_classes <=0, last fc layer
will not be defined. Default: 1000. import paddle
with_pool (bool): use pool before the last fc layer or not. Default: True. from paddle.vision.models import MobileNetV2
Examples: model = MobileNetV2()
.. code-block:: python
from paddle.vision.models import MobileNetV2 x = paddle.rand([1, 3, 224, 224])
out = model(x)
print(out.shape)
"""
model = MobileNetV2() def __init__(self, scale=1.0, num_classes=1000, with_pool=True):
"""
super(MobileNetV2, self).__init__() super(MobileNetV2, self).__init__()
self.num_classes = num_classes self.num_classes = num_classes
self.with_pool = with_pool self.with_pool = with_pool
...@@ -130,8 +120,12 @@ class MobileNetV2(nn.Layer): ...@@ -130,8 +120,12 @@ class MobileNetV2(nn.Layer):
self.last_channel = _make_divisible(last_channel * max(1.0, scale), self.last_channel = _make_divisible(last_channel * max(1.0, scale),
round_nearest) round_nearest)
features = [ features = [
ConvBNReLU( ConvNormActivation(
3, input_channel, stride=2, norm_layer=norm_layer) 3,
input_channel,
stride=2,
norm_layer=norm_layer,
activation_layer=nn.ReLU6)
] ]
for t, c, n, s in inverted_residual_setting: for t, c, n, s in inverted_residual_setting:
...@@ -148,11 +142,12 @@ class MobileNetV2(nn.Layer): ...@@ -148,11 +142,12 @@ class MobileNetV2(nn.Layer):
input_channel = output_channel input_channel = output_channel
features.append( features.append(
ConvBNReLU( ConvNormActivation(
input_channel, input_channel,
self.last_channel, self.last_channel,
kernel_size=1, kernel_size=1,
norm_layer=norm_layer)) norm_layer=norm_layer,
activation_layer=nn.ReLU6))
self.features = nn.Sequential(*features) self.features = nn.Sequential(*features)
...@@ -199,6 +194,7 @@ def mobilenet_v2(pretrained=False, scale=1.0, **kwargs): ...@@ -199,6 +194,7 @@ def mobilenet_v2(pretrained=False, scale=1.0, **kwargs):
Examples: Examples:
.. code-block:: python .. code-block:: python
import paddle
from paddle.vision.models import mobilenet_v2 from paddle.vision.models import mobilenet_v2
# build model # build model
...@@ -209,6 +205,11 @@ def mobilenet_v2(pretrained=False, scale=1.0, **kwargs): ...@@ -209,6 +205,11 @@ def mobilenet_v2(pretrained=False, scale=1.0, **kwargs):
# build mobilenet v2 with scale=0.5 # build mobilenet v2 with scale=0.5
model = mobilenet_v2(scale=0.5) model = mobilenet_v2(scale=0.5)
x = paddle.rand([1, 3, 224, 224])
out = model(x)
print(out.shape)
""" """
model = _mobilenet( model = _mobilenet(
'mobilenetv2_' + str(scale), pretrained, scale=scale, **kwargs) 'mobilenetv2_' + str(scale), pretrained, scale=scale, **kwargs)
......
python/paddle/vision/models/shufflenetv2.py
浏览文件 @ f6219dda
...@@ -18,37 +18,50 @@ from __future__ import print_function ...@@ -18,37 +18,50 @@ from __future__ import print_function
import paddle import paddle
import paddle.nn as nn import paddle.nn as nn
from paddle.fluid.param_attr import ParamAttr from paddle.nn import AdaptiveAvgPool2D, Linear, MaxPool2D
from paddle.nn import AdaptiveAvgPool2D, BatchNorm, Conv2D, Linear, MaxPool2D
from paddle.utils.download import get_weights_path_from_url from paddle.utils.download import get_weights_path_from_url
from ..ops import ConvNormActivation
__all__ = [] __all__ = []
model_urls = { model_urls = {
"shufflenet_v2_x0_25": ( "shufflenet_v2_x0_25": (
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ShuffleNetV2_x0_25_pretrained.pdparams", "https://paddle-hapi.bj.bcebos.com/models/shufflenet_v2_x0_25.pdparams",
"e753404cbd95027759c5f56ecd6c9c4b", ), "1e509b4c140eeb096bb16e214796d03b", ),
"shufflenet_v2_x0_33": ( "shufflenet_v2_x0_33": (
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ShuffleNetV2_x0_33_pretrained.pdparams", "https://paddle-hapi.bj.bcebos.com/models/shufflenet_v2_x0_33.pdparams",
"776e3cf9a4923abdfce789c45b8fe1f2", ), "3d7b3ab0eaa5c0927ff1026d31b729bd", ),
"shufflenet_v2_x0_5": ( "shufflenet_v2_x0_5": (
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ShuffleNetV2_x0_5_pretrained.pdparams", "https://paddle-hapi.bj.bcebos.com/models/shufflenet_v2_x0_5.pdparams",
"e3649cf531566917e2969487d2bc6b60", ), "5e5cee182a7793c4e4c73949b1a71bd4", ),
"shufflenet_v2_x1_0": ( "shufflenet_v2_x1_0": (
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ShuffleNetV2_x1_0_pretrained.pdparams", "https://paddle-hapi.bj.bcebos.com/models/shufflenet_v2_x1_0.pdparams",
"7821c348ea34e58847c43a08a4ac0bdf", ), "122d42478b9e81eb49f8a9ede327b1a4", ),
"shufflenet_v2_x1_5": ( "shufflenet_v2_x1_5": (
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ShuffleNetV2_x1_5_pretrained.pdparams", "https://paddle-hapi.bj.bcebos.com/models/shufflenet_v2_x1_5.pdparams",
"93a07fa557ab2d8803550f39e5b6c391", ), "faced5827380d73531d0ee027c67826d", ),
"shufflenet_v2_x2_0": ( "shufflenet_v2_x2_0": (
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ShuffleNetV2_x2_0_pretrained.pdparams", "https://paddle-hapi.bj.bcebos.com/models/shufflenet_v2_x2_0.pdparams",
"4ab1f622fd0d341e0f84b4e057797563", ), "cd3dddcd8305e7bcd8ad14d1c69a5784", ),
"shufflenet_v2_swish": ( "shufflenet_v2_swish": (
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ShuffleNetV2_swish_pretrained.pdparams", "https://paddle-hapi.bj.bcebos.com/models/shufflenet_v2_swish.pdparams",
"daff38b3df1b3748fccbb13cfdf02519", ), "adde0aa3b023e5b0c94a68be1c394b84", ),
} }
def create_activation_layer(act):
if act == "swish":
return nn.Swish
elif act == "relu":
return nn.ReLU
elif act is None:
return None
else:
raise RuntimeError(
"The activation function is not supported: {}".format(act))
def channel_shuffle(x, groups): def channel_shuffle(x, groups):
batch_size, num_channels, height, width = x.shape[0:4] batch_size, num_channels, height, width = x.shape[0:4]
channels_per_group = num_channels // groups channels_per_group = num_channels // groups
...@@ -65,61 +78,37 @@ def channel_shuffle(x, groups): ...@@ -65,61 +78,37 @@ def channel_shuffle(x, groups):
return x return x
class ConvBNLayer(nn.Layer): class InvertedResidual(nn.Layer):
def __init__(self, def __init__(self,
in_channels, in_channels,
out_channels, out_channels,
kernel_size,
stride, stride,
padding, activation_layer=nn.ReLU):
groups=1,
act=None):
super(ConvBNLayer, self).__init__()
self._conv = Conv2D(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
groups=groups,
weight_attr=ParamAttr(initializer=nn.initializer.KaimingNormal()),
bias_attr=False, )
self._batch_norm = BatchNorm(out_channels, act=act)
def forward(self, inputs):
x = self._conv(inputs)
x = self._batch_norm(x)
return x
class InvertedResidual(nn.Layer):
def __init__(self, in_channels, out_channels, stride, act="relu"):
super(InvertedResidual, self).__init__() super(InvertedResidual, self).__init__()
self._conv_pw = ConvBNLayer( self._conv_pw = ConvNormActivation(
in_channels=in_channels // 2, in_channels=in_channels // 2,
out_channels=out_channels // 2, out_channels=out_channels // 2,
kernel_size=1, kernel_size=1,
stride=1, stride=1,
padding=0, padding=0,
groups=1, groups=1,
act=act) activation_layer=activation_layer)
self._conv_dw = ConvBNLayer( self._conv_dw = ConvNormActivation(
in_channels=out_channels // 2, in_channels=out_channels // 2,
out_channels=out_channels // 2, out_channels=out_channels // 2,
kernel_size=3, kernel_size=3,
stride=stride, stride=stride,
padding=1, padding=1,
groups=out_channels // 2, groups=out_channels // 2,
act=None) activation_layer=None)
self._conv_linear = ConvBNLayer( self._conv_linear = ConvNormActivation(
in_channels=out_channels // 2, in_channels=out_channels // 2,
out_channels=out_channels // 2, out_channels=out_channels // 2,
kernel_size=1, kernel_size=1,
stride=1, stride=1,
padding=0, padding=0,
groups=1, groups=1,
act=act) activation_layer=activation_layer)
def forward(self, inputs): def forward(self, inputs):
x1, x2 = paddle.split( x1, x2 = paddle.split(
...@@ -134,51 +123,55 @@ class InvertedResidual(nn.Layer): ...@@ -134,51 +123,55 @@ class InvertedResidual(nn.Layer):
class InvertedResidualDS(nn.Layer): class InvertedResidualDS(nn.Layer):
def __init__(self, in_channels, out_channels, stride, act="relu"): def __init__(self,
in_channels,
out_channels,
stride,
activation_layer=nn.ReLU):
super(InvertedResidualDS, self).__init__() super(InvertedResidualDS, self).__init__()
# branch1 # branch1
self._conv_dw_1 = ConvBNLayer( self._conv_dw_1 = ConvNormActivation(
in_channels=in_channels, in_channels=in_channels,
out_channels=in_channels, out_channels=in_channels,
kernel_size=3, kernel_size=3,
stride=stride, stride=stride,
padding=1, padding=1,
groups=in_channels, groups=in_channels,
act=None) activation_layer=None)
self._conv_linear_1 = ConvBNLayer( self._conv_linear_1 = ConvNormActivation(
in_channels=in_channels, in_channels=in_channels,
out_channels=out_channels // 2, out_channels=out_channels // 2,
kernel_size=1, kernel_size=1,
stride=1, stride=1,
padding=0, padding=0,
groups=1, groups=1,
act=act) activation_layer=activation_layer)
# branch2 # branch2
self._conv_pw_2 = ConvBNLayer( self._conv_pw_2 = ConvNormActivation(
in_channels=in_channels, in_channels=in_channels,
out_channels=out_channels // 2, out_channels=out_channels // 2,
kernel_size=1, kernel_size=1,
stride=1, stride=1,
padding=0, padding=0,
groups=1, groups=1,
act=act) activation_layer=activation_layer)
self._conv_dw_2 = ConvBNLayer( self._conv_dw_2 = ConvNormActivation(
in_channels=out_channels // 2, in_channels=out_channels // 2,
out_channels=out_channels // 2, out_channels=out_channels // 2,
kernel_size=3, kernel_size=3,
stride=stride, stride=stride,
padding=1, padding=1,
groups=out_channels // 2, groups=out_channels // 2,
act=None) activation_layer=None)
self._conv_linear_2 = ConvBNLayer( self._conv_linear_2 = ConvNormActivation(
in_channels=out_channels // 2, in_channels=out_channels // 2,
out_channels=out_channels // 2, out_channels=out_channels // 2,
kernel_size=1, kernel_size=1,
stride=1, stride=1,
padding=0, padding=0,
groups=1, groups=1,
act=act) activation_layer=activation_layer)
def forward(self, inputs): def forward(self, inputs):
x1 = self._conv_dw_1(inputs) x1 = self._conv_dw_1(inputs)
...@@ -221,6 +214,7 @@ class ShuffleNetV2(nn.Layer): ...@@ -221,6 +214,7 @@ class ShuffleNetV2(nn.Layer):
self.num_classes = num_classes self.num_classes = num_classes
self.with_pool = with_pool self.with_pool = with_pool
stage_repeats = [4, 8, 4] stage_repeats = [4, 8, 4]
activation_layer = create_activation_layer(act)
if scale == 0.25: if scale == 0.25:
stage_out_channels = [-1, 24, 24, 48, 96, 512] stage_out_channels = [-1, 24, 24, 48, 96, 512]
...@@ -238,13 +232,13 @@ class ShuffleNetV2(nn.Layer): ...@@ -238,13 +232,13 @@ class ShuffleNetV2(nn.Layer):
raise NotImplementedError("This scale size:[" + str(scale) + raise NotImplementedError("This scale size:[" + str(scale) +
"] is not implemented!") "] is not implemented!")
# 1. conv1 # 1. conv1
self._conv1 = ConvBNLayer( self._conv1 = ConvNormActivation(
in_channels=3, in_channels=3,
out_channels=stage_out_channels[1], out_channels=stage_out_channels[1],
kernel_size=3, kernel_size=3,
stride=2, stride=2,
padding=1, padding=1,
act=act) activation_layer=activation_layer)
self._max_pool = MaxPool2D(kernel_size=3, stride=2, padding=1) self._max_pool = MaxPool2D(kernel_size=3, stride=2, padding=1)
# 2. bottleneck sequences # 2. bottleneck sequences
...@@ -257,7 +251,7 @@ class ShuffleNetV2(nn.Layer): ...@@ -257,7 +251,7 @@ class ShuffleNetV2(nn.Layer):
in_channels=stage_out_channels[stage_id + 1], in_channels=stage_out_channels[stage_id + 1],
out_channels=stage_out_channels[stage_id + 2], out_channels=stage_out_channels[stage_id + 2],
stride=2, stride=2,
act=act), activation_layer=activation_layer),
name=str(stage_id + 2) + "_" + str(i + 1)) name=str(stage_id + 2) + "_" + str(i + 1))
else: else:
block = self.add_sublayer( block = self.add_sublayer(
...@@ -265,17 +259,17 @@ class ShuffleNetV2(nn.Layer): ...@@ -265,17 +259,17 @@ class ShuffleNetV2(nn.Layer):
in_channels=stage_out_channels[stage_id + 2], in_channels=stage_out_channels[stage_id + 2],
out_channels=stage_out_channels[stage_id + 2], out_channels=stage_out_channels[stage_id + 2],
stride=1, stride=1,
act=act), activation_layer=activation_layer),
name=str(stage_id + 2) + "_" + str(i + 1)) name=str(stage_id + 2) + "_" + str(i + 1))
self._block_list.append(block) self._block_list.append(block)
# 3. last_conv # 3. last_conv
self._last_conv = ConvBNLayer( self._last_conv = ConvNormActivation(
in_channels=stage_out_channels[-2], in_channels=stage_out_channels[-2],
out_channels=stage_out_channels[-1], out_channels=stage_out_channels[-1],
kernel_size=1, kernel_size=1,
stride=1, stride=1,
padding=0, padding=0,
act=act) activation_layer=activation_layer)
# 4. pool # 4. pool
if with_pool: if with_pool:
self._pool2d_avg = AdaptiveAvgPool2D(1) self._pool2d_avg = AdaptiveAvgPool2D(1)
......
python/paddle/vision/ops.py
浏览文件 @ f6219dda
...@@ -1335,13 +1335,13 @@ class ConvNormActivation(Sequential): ...@@ -1335,13 +1335,13 @@ class ConvNormActivation(Sequential):
Args: Args:
in_channels (int): Number of channels in the input image in_channels (int): Number of channels in the input image
out_channels (int): Number of channels produced by the Convolution-Normalzation-Activation block out_channels (int): Number of channels produced by the Convolution-Normalzation-Activation block
kernel_size: (int, optional): Size of the convolving kernel. Default: 3 kernel_size: (int|list|tuple, optional): Size of the convolving kernel. Default: 3
stride (int, optional): Stride of the convolution. Default: 1 stride (int|list|tuple, optional): Stride of the convolution. Default: 1
padding (int, tuple or str, optional): Padding added to all four sides of the input. Default: None, padding (int|str|tuple|list, optional): Padding added to all four sides of the input. Default: None,
in wich case it will calculated as ``padding = (kernel_size - 1) // 2 * dilation`` in wich case it will calculated as ``padding = (kernel_size - 1) // 2 * dilation``
groups (int, optional): Number of blocked connections from input channels to output channels. Default: 1 groups (int, optional): Number of blocked connections from input channels to output channels. Default: 1
norm_layer (Callable[..., paddle.nn.Layer], optional): Norm layer that will be stacked on top of the convolutiuon layer. norm_layer (Callable[..., paddle.nn.Layer], optional): Norm layer that will be stacked on top of the convolutiuon layer.
If ``None`` this layer wont be used. Default: ``paddle.nn.BatchNorm2d`` If ``None`` this layer wont be used. Default: ``paddle.nn.BatchNorm2D``
activation_layer (Callable[..., paddle.nn.Layer], optional): Activation function which will be stacked on top of the normalization activation_layer (Callable[..., paddle.nn.Layer], optional): Activation function which will be stacked on top of the normalization
layer (if not ``None``), otherwise on top of the conv layer. If ``None`` this layer wont be used. Default: ``paddle.nn.ReLU`` layer (if not ``None``), otherwise on top of the conv layer. If ``None`` this layer wont be used. Default: ``paddle.nn.ReLU``
dilation (int): Spacing between kernel elements. Default: 1 dilation (int): Spacing between kernel elements. Default: 1
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或