Skip to content

P
PaddleClas

PaddlePaddle / PaddleClas
大约 2 年 前同步成功

通知 118
Star 4999
Fork 1114
P
PaddleClas
未验证 提交 ea746480 编写于 作者: L littletomatodonkey 提交者： GitHub
浏览文件
操作

Merge pull request #275 from littletomatodonkey/dyg/adp-2.0b 

Fix clas api to paddle2.0.0b0
上级 5e092259 922f33fe
隐藏空白更改
内联 并排
Showing with 1493 addition and 2490 deletion
ppcls/data/reader.py
浏览文件 @ ea746480
...@@ -17,7 +17,7 @@ import imghdr ...@@ -17,7 +17,7 @@ import imghdr
import os import os
import signal import signal
from paddle.fluid.io import multiprocess_reader from paddle.reader import multiprocess_reader
from . import imaug from . import imaug
from .imaug import transform from .imaug import transform
......
ppcls/modeling/architectures/__init__.py
浏览文件 @ ea746480
...@@ -23,7 +23,7 @@ from .se_resnet_vd import SE_ResNet18_vd, SE_ResNet34_vd, SE_ResNet50_vd, SE_Res ...@@ -23,7 +23,7 @@ from .se_resnet_vd import SE_ResNet18_vd, SE_ResNet34_vd, SE_ResNet50_vd, SE_Res
from .se_resnext_vd import SE_ResNeXt50_vd_32x4d, SE_ResNeXt50_vd_32x4d, SENet154_vd from .se_resnext_vd import SE_ResNeXt50_vd_32x4d, SE_ResNeXt50_vd_32x4d, SENet154_vd
from .dpn import DPN68 from .dpn import DPN68
from .densenet import DenseNet121 from .densenet import DenseNet121
from .hrnet import HRNet_W18_C from .hrnet import HRNet_W18_C, HRNet_W30_C, HRNet_W32_C, HRNet_W40_C, HRNet_W44_C, HRNet_W48_C, HRNet_W60_C, HRNet_W64_C, SE_HRNet_W18_C, SE_HRNet_W30_C, SE_HRNet_W32_C, SE_HRNet_W40_C, SE_HRNet_W44_C, SE_HRNet_W48_C, SE_HRNet_W60_C, SE_HRNet_W64_C
from .efficientnet import EfficientNetB0 from .efficientnet import EfficientNetB0
from .resnest import ResNeSt50_fast_1s1x64d, ResNeSt50 from .resnest import ResNeSt50_fast_1s1x64d, ResNeSt50
from .googlenet import GoogLeNet from .googlenet import GoogLeNet
...@@ -31,5 +31,14 @@ from .mobilenet_v1 import MobileNetV1_x0_25, MobileNetV1_x0_5, MobileNetV1_x0_75 ...@@ -31,5 +31,14 @@ from .mobilenet_v1 import MobileNetV1_x0_25, MobileNetV1_x0_5, MobileNetV1_x0_75
from .mobilenet_v2 import MobileNetV2_x0_25, MobileNetV2_x0_5, MobileNetV2_x0_75, MobileNetV2, MobileNetV2_x1_5, MobileNetV2_x2_0 from .mobilenet_v2 import MobileNetV2_x0_25, MobileNetV2_x0_5, MobileNetV2_x0_75, MobileNetV2, MobileNetV2_x1_5, MobileNetV2_x2_0
from .mobilenet_v3 import MobileNetV3_small_x0_35, MobileNetV3_small_x0_5, MobileNetV3_small_x0_75, MobileNetV3_small_x1_0, MobileNetV3_small_x1_25, MobileNetV3_large_x0_35, MobileNetV3_large_x0_5, MobileNetV3_large_x0_75, MobileNetV3_large_x1_0, MobileNetV3_large_x1_25 from .mobilenet_v3 import MobileNetV3_small_x0_35, MobileNetV3_small_x0_5, MobileNetV3_small_x0_75, MobileNetV3_small_x1_0, MobileNetV3_small_x1_25, MobileNetV3_large_x0_35, MobileNetV3_large_x0_5, MobileNetV3_large_x0_75, MobileNetV3_large_x1_0, MobileNetV3_large_x1_25
from .shufflenet_v2 import ShuffleNetV2_x0_25, ShuffleNetV2_x0_33, ShuffleNetV2_x0_5, ShuffleNetV2, ShuffleNetV2_x1_5, ShuffleNetV2_x2_0, ShuffleNetV2_swish from .shufflenet_v2 import ShuffleNetV2_x0_25, ShuffleNetV2_x0_33, ShuffleNetV2_x0_5, ShuffleNetV2, ShuffleNetV2_x1_5, ShuffleNetV2_x2_0, ShuffleNetV2_swish
from .alexnet import AlexNet
from .inception_v4 import InceptionV4
from .xception import Xception41, Xception65, Xception71
from .xception_deeplab import Xception41_deeplab, Xception65_deeplab, Xception71_deeplab
from .resnext101_wsl import ResNeXt101_32x8d_wsl, ResNeXt101_32x16d_wsl, ResNeXt101_32x32d_wsl, ResNeXt101_32x48d_wsl
from .shufflenet_v2 import ShuffleNetV2_x0_25, ShuffleNetV2_x0_33, ShuffleNetV2_x0_5, ShuffleNetV2, ShuffleNetV2_x1_5, ShuffleNetV2_x2_0, ShuffleNetV2_swish
from .squeezenet import SqueezeNet1_0, SqueezeNet1_1
from .vgg import VGG11, VGG13, VGG16, VGG19
from .darknet import DarkNet53
from .distillation_models import ResNet50_vd_distill_MobileNetV3_large_x1_0 from .distillation_models import ResNet50_vd_distill_MobileNetV3_large_x1_0
ppcls/modeling/architectures/alexnet.py
浏览文件 @ ea746480
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear, Dropout, ReLU
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
__all__ = ["AlexNet"] __all__ = ["AlexNet"]
class ConvPoolLayer(fluid.dygraph.Layer):
def __init__(self, class ConvPoolLayer(nn.Layer):
inputc_channels, def __init__(self,
output_channels, input_channels,
filter_size, output_channels,
stride, filter_size,
padding, stride,
stdv, padding,
groups=1, stdv,
act=None, groups=1,
name=None): act=None,
name=None):
super(ConvPoolLayer, self).__init__() super(ConvPoolLayer, self).__init__()
self._conv = Conv2D(num_channels=inputc_channels, self.relu = ReLU() if act == "relu" else None
num_filters=output_channels,
filter_size=filter_size, self._conv = Conv2d(
stride=stride, in_channels=input_channels,
padding=padding, out_channels=output_channels,
groups=groups, kernel_size=filter_size,
param_attr=ParamAttr(name=name + "_weights", stride=stride,
initializer=fluid.initializer.Uniform(-stdv, stdv)), padding=padding,
bias_attr=ParamAttr(name=name + "_offset", groups=groups,
initializer=fluid.initializer.Uniform(-stdv, stdv)), weight_attr=ParamAttr(
act=act) name=name + "_weights", initializer=Uniform(-stdv, stdv)),
self._pool = Pool2D(pool_size=3, bias_attr=ParamAttr(
pool_stride=2, name=name + "_offset", initializer=Uniform(-stdv, stdv)))
pool_padding=0, self._pool = MaxPool2d(kernel_size=3, stride=2, padding=0)
pool_type="max")
def forward(self, inputs): def forward(self, inputs):
x = self._conv(inputs) x = self._conv(inputs)
if self.relu is not None:
x = self.relu(x)
x = self._pool(x) x = self._pool(x)
return x return x
class AlexNetDY(fluid.dygraph.Layer): class AlexNetDY(nn.Layer):
def __init__(self, class_dim=1000): def __init__(self, class_dim=1000):
super(AlexNetDY, self).__init__() super(AlexNetDY, self).__init__()
stdv = 1.0/math.sqrt(3*11*11) stdv = 1.0 / math.sqrt(3 * 11 * 11)
self._conv1 = ConvPoolLayer( self._conv1 = ConvPoolLayer(
3, 64, 11, 4, 2, stdv, act="relu", name="conv1") 3, 64, 11, 4, 2, stdv, act="relu", name="conv1")
stdv = 1.0/math.sqrt(64*5*5) stdv = 1.0 / math.sqrt(64 * 5 * 5)
self._conv2 = ConvPoolLayer( self._conv2 = ConvPoolLayer(
64, 192, 5, 1, 2, stdv, act="relu", name="conv2") 64, 192, 5, 1, 2, stdv, act="relu", name="conv2")
stdv = 1.0/math.sqrt(192*3*3) stdv = 1.0 / math.sqrt(192 * 3 * 3)
self._conv3 = Conv2D(192, 384, 3, stride=1, padding=1, self._conv3 = Conv2d(
param_attr=ParamAttr(name="conv3_weights", initializer=fluid.initializer.Uniform(-stdv, stdv)), 192,
bias_attr=ParamAttr(name="conv3_offset", initializer=fluid.initializer.Uniform(-stdv, stdv)), 384,
act="relu") 3,
stdv = 1.0/math.sqrt(384*3*3) stride=1,
self._conv4 = Conv2D(384, 256, 3, stride=1, padding=1, padding=1,
param_attr=ParamAttr(name="conv4_weights", initializer=fluid.initializer.Uniform(-stdv, stdv)), weight_attr=ParamAttr(
bias_attr=ParamAttr(name="conv4_offset", initializer=fluid.initializer.Uniform(-stdv, stdv)), name="conv3_weights", initializer=Uniform(-stdv, stdv)),
act="relu") bias_attr=ParamAttr(
stdv = 1.0/math.sqrt(256*3*3) name="conv3_offset", initializer=Uniform(-stdv, stdv)))
stdv = 1.0 / math.sqrt(384 * 3 * 3)
self._conv4 = Conv2d(
384,
256,
3,
stride=1,
padding=1,
weight_attr=ParamAttr(
name="conv4_weights", initializer=Uniform(-stdv, stdv)),
bias_attr=ParamAttr(
name="conv4_offset", initializer=Uniform(-stdv, stdv)))
stdv = 1.0 / math.sqrt(256 * 3 * 3)
self._conv5 = ConvPoolLayer( self._conv5 = ConvPoolLayer(
256, 256, 3, 1, 1, stdv, act="relu", name="conv5") 256, 256, 3, 1, 1, stdv, act="relu", name="conv5")
stdv = 1.0/math.sqrt(256*6*6) stdv = 1.0 / math.sqrt(256 * 6 * 6)
self._drop1 = Dropout(p=0.5) self._drop1 = Dropout(p=0.5, mode="downscale_in_infer")
self._fc6 = Linear(input_dim=256*6*6, self._fc6 = Linear(
output_dim=4096, in_features=256 * 6 * 6,
param_attr=ParamAttr(name="fc6_weights", initializer=fluid.initializer.Uniform(-stdv, stdv)), out_features=4096,
bias_attr=ParamAttr(name="fc6_offset", initializer=fluid.initializer.Uniform(-stdv, stdv)), weight_attr=ParamAttr(
act="relu") name="fc6_weights", initializer=Uniform(-stdv, stdv)),
bias_attr=ParamAttr(
self._drop2 = Dropout(p=0.5) name="fc6_offset", initializer=Uniform(-stdv, stdv)))
self._fc7 = Linear(input_dim=4096,
output_dim=4096, self._drop2 = Dropout(p=0.5, mode="downscale_in_infer")
param_attr=ParamAttr(name="fc7_weights", initializer=fluid.initializer.Uniform(-stdv, stdv)), self._fc7 = Linear(
bias_attr=ParamAttr(name="fc7_offset", initializer=fluid.initializer.Uniform(-stdv, stdv)), in_features=4096,
act="relu") out_features=4096,
self._fc8 = Linear(input_dim=4096, weight_attr=ParamAttr(
output_dim=class_dim, name="fc7_weights", initializer=Uniform(-stdv, stdv)),
param_attr=ParamAttr(name="fc8_weights", initializer=fluid.initializer.Uniform(-stdv, stdv)), bias_attr=ParamAttr(
bias_attr=ParamAttr(name="fc8_offset", initializer=fluid.initializer.Uniform(-stdv, stdv))) name="fc7_offset", initializer=Uniform(-stdv, stdv)))
self._fc8 = Linear(
in_features=4096,
out_features=class_dim,
weight_attr=ParamAttr(
name="fc8_weights", initializer=Uniform(-stdv, stdv)),
bias_attr=ParamAttr(
name="fc8_offset", initializer=Uniform(-stdv, stdv)))
def forward(self, inputs): def forward(self, inputs):
x = self._conv1(inputs) x = self._conv1(inputs)
x = self._conv2(x) x = self._conv2(x)
x = self._conv3(x) x = self._conv3(x)
x = F.relu(x)
x = self._conv4(x) x = self._conv4(x)
x = F.relu(x)
x = self._conv5(x) x = self._conv5(x)
x = fluid.layers.flatten(x, axis=0) x = paddle.flatten(x, start_axis=1, stop_axis=-1)
x = self._drop1(x) x = self._drop1(x)
x = self._fc6(x) x = self._fc6(x)
x = F.relu(x)
x = self._drop2(x) x = self._drop2(x)
x = self._fc7(x) x = self._fc7(x)
x = F.relu(x)
x = self._fc8(x) x = self._fc8(x)
return x return x
def AlexNet(**args): def AlexNet(**args):
model = AlexNetDY(**args) model = AlexNetDY(**args)
return model return model
ppcls/modeling/architectures/csp_resnet.py 已删除 100644 → 0
浏览文件 @ 5e092259
# 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.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import math
import paddle.fluid as fluid
from paddle.fluid.param_attr import ParamAttr
__all__ = [
"CSPResNet50_leaky", "CSPResNet50_mish", "CSPResNet101_leaky",
"CSPResNet101_mish"
]
class CSPResNet():
def __init__(self, layers=50, act="leaky_relu"):
self.layers = layers
self.act = act
def net(self, input, class_dim=1000, data_format="NCHW"):
layers = self.layers
supported_layers = [50, 101]
assert layers in supported_layers, \
"supported layers are {} but input layer is {}".format(
supported_layers, layers)
if layers == 50:
depth = [3, 3, 5, 2]
elif layers == 101:
depth = [3, 3, 22, 2]
num_filters = [64, 128, 256, 512]
conv = self.conv_bn_layer(
input=input,
num_filters=64,
filter_size=7,
stride=2,
act=self.act,
name="conv1",
data_format=data_format)
conv = fluid.layers.pool2d(
input=conv,
pool_size=2,
pool_stride=2,
pool_padding=0,
pool_type='max',
data_format=data_format)
for block in range(len(depth)):
conv_name = "res" + str(block + 2) + chr(97)
if block != 0:
conv = self.conv_bn_layer(
input=conv,
num_filters=num_filters[block],
filter_size=3,
stride=2,
act=self.act,
name=conv_name + "_downsample",
data_format=data_format)
# split
left = conv
right = conv
if block == 0:
ch = num_filters[block]
else:
ch = num_filters[block] * 2
right = self.conv_bn_layer(
input=right,
num_filters=ch,
filter_size=1,
act=self.act,
name=conv_name + "_right_first_route",
data_format=data_format)
for i in range(depth[block]):
conv_name = "res" + str(block + 2) + chr(97 + i)
right = self.bottleneck_block(
input=right,
num_filters=num_filters[block],
stride=1,
name=conv_name,
data_format=data_format)
# route
left = self.conv_bn_layer(
input=left,
num_filters=num_filters[block] * 2,
filter_size=1,
act=self.act,
name=conv_name + "_left_route",
data_format=data_format)
right = self.conv_bn_layer(
input=right,
num_filters=num_filters[block] * 2,
filter_size=1,
act=self.act,
name=conv_name + "_right_route",
data_format=data_format)
conv = fluid.layers.concat([left, right], axis=1)
conv = self.conv_bn_layer(
input=conv,
num_filters=num_filters[block] * 2,
filter_size=1,
stride=1,
act=self.act,
name=conv_name + "_merged_transition",
data_format=data_format)
pool = fluid.layers.pool2d(
input=conv,
pool_type='avg',
global_pooling=True,
data_format=data_format)
stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
out = fluid.layers.fc(
input=pool,
size=class_dim,
param_attr=fluid.param_attr.ParamAttr(
name="fc_0.w_0",
initializer=fluid.initializer.Uniform(-stdv, stdv)),
bias_attr=ParamAttr(name="fc_0.b_0"))
return out
def conv_bn_layer(self,
input,
num_filters,
filter_size,
stride=1,
groups=1,
act=None,
name=None,
data_format='NCHW'):
conv = fluid.layers.conv2d(
input=input,
num_filters=num_filters,
filter_size=filter_size,
stride=stride,
padding=(filter_size - 1) // 2,
groups=groups,
act=None,
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False,
name=name + '.conv2d.output.1',
data_format=data_format)
if name == "conv1":
bn_name = "bn_" + name
else:
bn_name = "bn" + name[3:]
bn = fluid.layers.batch_norm(
input=conv,
act=None,
name=bn_name + '.output.1',
param_attr=ParamAttr(name=bn_name + '_scale'),
bias_attr=ParamAttr(bn_name + '_offset'),
moving_mean_name=bn_name + '_mean',
moving_variance_name=bn_name + '_variance',
data_layout=data_format)
if act == "relu":
bn = fluid.layers.relu(bn)
elif act == "leaky_relu":
bn = fluid.layers.leaky_relu(bn)
elif act == "mish":
bn = self._mish(bn)
return bn
def _mish(self, input):
return input * fluid.layers.tanh(self._softplus(input))
def _softplus(self, input):
expf = fluid.layers.exp(fluid.layers.clip(input, -200, 50))
return fluid.layers.log(1 + expf)
def shortcut(self, input, ch_out, stride, is_first, name, data_format):
if data_format == 'NCHW':
ch_in = input.shape[1]
else:
ch_in = input.shape[-1]
if ch_in != ch_out or stride != 1 or is_first is True:
return self.conv_bn_layer(
input, ch_out, 1, stride, name=name, data_format=data_format)
else:
return input
def bottleneck_block(self, input, num_filters, stride, name, data_format):
conv0 = self.conv_bn_layer(
input=input,
num_filters=num_filters,
filter_size=1,
act="leaky_relu",
name=name + "_branch2a",
data_format=data_format)
conv1 = self.conv_bn_layer(
input=conv0,
num_filters=num_filters,
filter_size=3,
stride=stride,
act="leaky_relu",
name=name + "_branch2b",
data_format=data_format)
conv2 = self.conv_bn_layer(
input=conv1,
num_filters=num_filters * 2,
filter_size=1,
act=None,
name=name + "_branch2c",
data_format=data_format)
short = self.shortcut(
input,
num_filters * 2,
stride,
is_first=False,
name=name + "_branch1",
data_format=data_format)
ret = short + conv2
ret = fluid.layers.leaky_relu(ret, alpha=0.1)
return ret
def CSPResNet50_leaky():
model = CSPResNet(layers=50, act="leaky_relu")
return model
def CSPResNet50_mish():
model = CSPResNet(layers=50, act="mish")
return model
def CSPResNet101_leaky():
model = CSPResNet(layers=101, act="leaky_relu")
return model
def CSPResNet101_mish():
model = CSPResNet(layers=101, act="mish")
return model
ppcls/modeling/architectures/darknet.py
浏览文件 @ ea746480
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
__all__ = ["DarkNet53"] __all__ = ["DarkNet53"]
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
output_channels, output_channels,
...@@ -17,14 +20,13 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -17,14 +20,13 @@ class ConvBNLayer(fluid.dygraph.Layer):
name=None): name=None):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self._conv = Conv2D( self._conv = Conv2d(
num_channels=input_channels, in_channels=input_channels,
num_filters=output_channels, out_channels=output_channels,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=padding, padding=padding,
act=None, weight_attr=ParamAttr(name=name + ".conv.weights"),
param_attr=ParamAttr(name=name + ".conv.weights"),
bias_attr=False) bias_attr=False)
bn_name = name + ".bn" bn_name = name + ".bn"
...@@ -42,7 +44,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -42,7 +44,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return x return x
class BasicBlock(fluid.dygraph.Layer): class BasicBlock(nn.Layer):
def __init__(self, input_channels, output_channels, name=None): def __init__(self, input_channels, output_channels, name=None):
super(BasicBlock, self).__init__() super(BasicBlock, self).__init__()
...@@ -54,10 +56,10 @@ class BasicBlock(fluid.dygraph.Layer): ...@@ -54,10 +56,10 @@ class BasicBlock(fluid.dygraph.Layer):
def forward(self, inputs): def forward(self, inputs):
x = self._conv1(inputs) x = self._conv1(inputs)
x = self._conv2(x) x = self._conv2(x)
return fluid.layers.elementwise_add(x=inputs, y=x) return paddle.elementwise_add(x=inputs, y=x)
class DarkNet(fluid.dygraph.Layer): class DarkNet(nn.Layer):
def __init__(self, class_dim=1000): def __init__(self, class_dim=1000):
super(DarkNet, self).__init__() super(DarkNet, self).__init__()
...@@ -102,15 +104,14 @@ class DarkNet(fluid.dygraph.Layer): ...@@ -102,15 +104,14 @@ class DarkNet(fluid.dygraph.Layer):
self._basic_block_43 = BasicBlock(1024, 512, name="stage.4.2") self._basic_block_43 = BasicBlock(1024, 512, name="stage.4.2")
self._basic_block_44 = BasicBlock(1024, 512, name="stage.4.3") self._basic_block_44 = BasicBlock(1024, 512, name="stage.4.3")
self._pool = Pool2D(pool_type="avg", global_pooling=True) self._pool = AdaptiveAvgPool2d(1)
stdv = 1.0 / math.sqrt(1024.0) stdv = 1.0 / math.sqrt(1024.0)
self._out = Linear( self._out = Linear(
input_dim=1024, 1024,
output_dim=class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
name="fc_weights", name="fc_weights", initializer=Uniform(-stdv, stdv)),
initializer=fluid.initializer.Uniform(-stdv, stdv)),
bias_attr=ParamAttr(name="fc_offset")) bias_attr=ParamAttr(name="fc_offset"))
def forward(self, inputs): def forward(self, inputs):
...@@ -150,11 +151,11 @@ class DarkNet(fluid.dygraph.Layer): ...@@ -150,11 +151,11 @@ class DarkNet(fluid.dygraph.Layer):
x = self._basic_block_44(x) x = self._basic_block_44(x)
x = self._pool(x) x = self._pool(x)
x = fluid.layers.squeeze(x, axes=[2, 3]) x = paddle.squeeze(x, axis=[2, 3])
x = self._out(x) x = self._out(x)
return x return x
def DarkNet53(**args): def DarkNet53(**args):
model = DarkNet(**args) model = DarkNet(**args)
return model return model
\ No newline at end of file
ppcls/modeling/architectures/darts_gs.py 已删除 100644 → 0
浏览文件 @ 5e092259
# 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.
#
# Based on:
# --------------------------------------------------------
# DARTS
# Copyright (c) 2018, Hanxiao Liu.
# Licensed under the Apache License, Version 2.0;
# --------------------------------------------------------
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import sys
import numpy as np
import time
import functools
import paddle
import paddle.fluid as fluid
from paddle.fluid.param_attr import ParamAttr
from paddle.fluid.initializer import Xavier
from paddle.fluid.initializer import Normal
from paddle.fluid.initializer import Constant
from collections import namedtuple
Genotype = namedtuple('Genotype', 'normal normal_concat reduce reduce_concat')
arch_dict = {
'DARTS_GS_6M': Genotype(
normal=[('sep_conv_3x3', 1), ('sep_conv_3x3', 0), ('sep_conv_5x5', 1),
('sep_conv_5x5', 0), ('sep_conv_3x3', 2), ('sep_conv_3x3', 1),
('skip_connect', 4), ('sep_conv_3x3', 3)],
normal_concat=range(2, 6),
reduce=[('sep_conv_5x5', 0), ('max_pool_3x3', 1), ('dil_conv_5x5', 2),
('sep_conv_5x5', 0), ('sep_conv_3x3', 1), ('dil_conv_5x5', 3),
('dil_conv_3x3', 1), ('sep_conv_3x3', 2)],
reduce_concat=range(2, 6)),
'DARTS_GS_4M': Genotype(
normal=[('sep_conv_3x3', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 0),
('sep_conv_3x3', 1), ('sep_conv_3x3', 1), ('skip_connect', 0),
('skip_connect', 0), ('dil_conv_3x3', 1)],
normal_concat=range(2, 6),
reduce=[('max_pool_3x3', 0), ('max_pool_3x3', 1), ('max_pool_3x3', 0),
('avg_pool_3x3', 1), ('skip_connect', 3), ('skip_connect', 2),
('sep_conv_3x3', 0), ('sep_conv_5x5', 2)],
reduce_concat=range(2, 6)),
}
__all__ = list(arch_dict.keys())
OPS = {
'none' : lambda input, C, stride, name, affine: Zero(input, stride, name),
'avg_pool_3x3' : lambda input, C, stride, name, affine: fluid.layers.pool2d(input, 3, 'avg', pool_stride=stride, pool_padding=1, name=name),
'max_pool_3x3' : lambda input, C, stride, name, affine: fluid.layers.pool2d(input, 3, 'max', pool_stride=stride, pool_padding=1, name=name),
'skip_connect' : lambda input,C, stride, name, affine: Identity(input, name) if stride == 1 else FactorizedReduce(input, C, name=name, affine=affine),
'sep_conv_3x3' : lambda input,C, stride, name, affine: SepConv(input, C, C, 3, stride, 1, name=name, affine=affine),
'sep_conv_5x5' : lambda input,C, stride, name, affine: SepConv(input, C, C, 5, stride, 2, name=name, affine=affine),
'sep_conv_7x7' : lambda input,C, stride, name, affine: SepConv(input, C, C, 7, stride, 3, name=name, affine=affine),
'dil_conv_3x3' : lambda input,C, stride, name, affine: DilConv(input, C, C, 3, stride, 2, 2, name=name, affine=affine),
'dil_conv_5x5' : lambda input,C, stride, name, affine: DilConv(input, C, C, 5, stride, 4, 2, name=name, affine=affine),
'conv_7x1_1x7' : lambda input,C, stride, name, affine: SevenConv(input, C, name=name, affine=affine)
}
def ReLUConvBN(input,
C_out,
kernel_size,
stride,
padding,
name='',
affine=True):
relu_a = fluid.layers.relu(input)
conv2d_a = fluid.layers.conv2d(
relu_a, C_out, kernel_size, stride, padding, bias_attr=False)
if affine:
reluconvbn_out = fluid.layers.batch_norm(
conv2d_a,
param_attr=ParamAttr(
initializer=Constant(1.), name=name + 'op.2.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=name + 'op.2.bias'),
moving_mean_name=name + 'op.2.running_mean',
moving_variance_name=name + 'op.2.running_var')
else:
reluconvbn_out = fluid.layers.batch_norm(
conv2d_a,
param_attr=ParamAttr(
initializer=Constant(1.),
learning_rate=0.,
name=name + 'op.2.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.),
learning_rate=0.,
name=name + 'op.2.bias'),
moving_mean_name=name + 'op.2.running_mean',
moving_variance_name=name + 'op.2.running_var')
return reluconvbn_out
def DilConv(input,
C_in,
C_out,
kernel_size,
stride,
padding,
dilation,
name='',
affine=True):
relu_a = fluid.layers.relu(input)
conv2d_a = fluid.layers.conv2d(
relu_a,
C_in,
kernel_size,
stride,
padding,
dilation,
groups=C_in,
bias_attr=False,
use_cudnn=False)
conv2d_b = fluid.layers.conv2d(conv2d_a, C_out, 1, bias_attr=False)
if affine:
dilconv_out = fluid.layers.batch_norm(
conv2d_b,
param_attr=ParamAttr(
initializer=Constant(1.), name=name + 'op.3.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=name + 'op.3.bias'),
moving_mean_name=name + 'op.3.running_mean',
moving_variance_name=name + 'op.3.running_var')
else:
dilconv_out = fluid.layers.batch_norm(
conv2d_b,
param_attr=ParamAttr(
initializer=Constant(1.),
learning_rate=0.,
name=name + 'op.3.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.),
learning_rate=0.,
name=name + 'op.3.bias'),
moving_mean_name=name + 'op.3.running_mean',
moving_variance_name=name + 'op.3.running_var')
return dilconv_out
def SepConv(input,
C_in,
C_out,
kernel_size,
stride,
padding,
name='',
affine=True):
relu_a = fluid.layers.relu(input)
conv2d_a = fluid.layers.conv2d(
relu_a,
C_in,
kernel_size,
stride,
padding,
groups=C_in,
bias_attr=False,
use_cudnn=False)
conv2d_b = fluid.layers.conv2d(conv2d_a, C_in, 1, bias_attr=False)
if affine:
bn_a = fluid.layers.batch_norm(
conv2d_b,
param_attr=ParamAttr(
initializer=Constant(1.), name=name + 'op.3.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=name + 'op.3.bias'),
moving_mean_name=name + 'op.3.running_mean',
moving_variance_name=name + 'op.3.running_var')
else:
bn_a = fluid.layers.batch_norm(
conv2d_b,
param_attr=ParamAttr(
initializer=Constant(1.),
learning_rate=0.,
name=name + 'op.3.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.),
learning_rate=0.,
name=name + 'op.3.bias'),
moving_mean_name=name + 'op.3.running_mean',
moving_variance_name=name + 'op.3.running_var')
relu_b = fluid.layers.relu(bn_a)
conv2d_d = fluid.layers.conv2d(
relu_b,
C_in,
kernel_size,
1,
padding,
groups=C_in,
bias_attr=False,
use_cudnn=False)
conv2d_e = fluid.layers.conv2d(conv2d_d, C_out, 1, bias_attr=False)
if affine:
sepconv_out = fluid.layers.batch_norm(
conv2d_e,
param_attr=ParamAttr(
initializer=Constant(1.), name=name + 'op.7.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=name + 'op.7.bias'),
moving_mean_name=name + 'op.7.running_mean',
moving_variance_name=name + 'op.7.running_var')
else:
sepconv_out = fluid.layers.batch_norm(
conv2d_e,
param_attr=ParamAttr(
initializer=Constant(1.),
learning_rate=0.,
name=name + 'op.7.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.),
learning_rate=0.,
name=name + 'op.7.bias'),
moving_mean_name=name + 'op.7.running_mean',
moving_variance_name=name + 'op.7.running_var')
return sepconv_out
def SevenConv(input, C_out, stride, name='', affine=True):
relu_a = fluid.layers.relu(input)
conv2d_a = fluid.layers.conv2d(
relu_a,
C_out, (1, 7), (1, stride), (0, 3),
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=name + 'op.1.weight'),
bias_attr=False)
conv2d_b = fluid.layers.conv2d(
conv2d_a,
C_out, (7, 1), (stride, 1), (3, 0),
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=name + 'op.2.weight'),
bias_attr=False)
if affine:
out = fluid.layers.batch_norm(
conv2d_b,
param_attr=ParamAttr(
initializer=Constant(1.), name=name + 'op.3.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=name + 'op.3.bias'),
moving_mean_name=name + 'op.3.running_mean',
moving_variance_name=name + 'op.3.running_var')
else:
out = fluid.layers.batch_norm(
conv2d_b,
param_attr=ParamAttr(
initializer=Constant(1.),
learning_rate=0.,
name=name + 'op.3.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.),
learning_rate=0.,
name=name + 'op.3.bias'),
moving_mean_name=name + 'op.3.running_mean',
moving_variance_name=name + 'op.3.running_var')
def Identity(input, name=''):
return input
def Zero(input, stride, name=''):
ones = np.ones(input.shape[-2:])
ones[::stride, ::stride] = 0
ones = fluid.layers.assign(ones)
return input * ones
def FactorizedReduce(input, C_out, name='', affine=True):
relu_a = fluid.layers.relu(input)
conv2d_a = fluid.layers.conv2d(
relu_a,
C_out // 2,
1,
2,
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=name + 'conv_1.weight'),
bias_attr=False)
h_end = relu_a.shape[2]
w_end = relu_a.shape[3]
slice_a = fluid.layers.slice(relu_a, [2, 3], [1, 1], [h_end, w_end])
conv2d_b = fluid.layers.conv2d(
slice_a,
C_out // 2,
1,
2,
param_attr=ParamAttr(
initializer=Xavier(
uniform=False, fan_in=0),
name=name + 'conv_2.weight'),
bias_attr=False)
out = fluid.layers.concat([conv2d_a, conv2d_b], axis=1)
if affine:
out = fluid.layers.batch_norm(
out,
param_attr=ParamAttr(
initializer=Constant(1.), name=name + 'bn.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=name + 'bn.bias'),
moving_mean_name=name + 'bn.running_mean',
moving_variance_name=name + 'bn.running_var')
else:
out = fluid.layers.batch_norm(
out,
param_attr=ParamAttr(
initializer=Constant(1.),
learning_rate=0.,
name=name + 'bn.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.),
learning_rate=0.,
name=name + 'bn.bias'),
moving_mean_name=name + 'bn.running_mean',
moving_variance_name=name + 'bn.running_var')
return out
class Cell():
def __init__(self, genotype, C_prev_prev, C_prev, C, reduction,
reduction_prev):
if reduction_prev:
self.preprocess0 = functools.partial(FactorizedReduce, C_out=C)
else:
self.preprocess0 = functools.partial(
ReLUConvBN, C_out=C, kernel_size=1, stride=1, padding=0)
self.preprocess1 = functools.partial(
ReLUConvBN, C_out=C, kernel_size=1, stride=1, padding=0)
if reduction:
op_names, indices = zip(*genotype.reduce)
concat = genotype.reduce_concat
else:
op_names, indices = zip(*genotype.normal)
concat = genotype.normal_concat
print(op_names, indices, concat, reduction)
self._compile(C, op_names, indices, concat, reduction)
def _compile(self, C, op_names, indices, concat, reduction):
assert len(op_names) == len(indices)
self._steps = len(op_names) // 2
self._concat = concat
self.multiplier = len(concat)
self._ops = []
for name, index in zip(op_names, indices):
stride = 2 if reduction and index < 2 else 1
op = functools.partial(OPS[name], C=C, stride=stride, affine=True)
self._ops += [op]
self._indices = indices
def forward(self, s0, s1, drop_prob, is_train, name):
self.training = is_train
preprocess0_name = name + 'preprocess0.'
preprocess1_name = name + 'preprocess1.'
s0 = self.preprocess0(s0, name=preprocess0_name)
s1 = self.preprocess1(s1, name=preprocess1_name)
out = [s0, s1]
for i in range(self._steps):
h1 = out[self._indices[2 * i]]
h2 = out[self._indices[2 * i + 1]]
op1 = self._ops[2 * i]
op2 = self._ops[2 * i + 1]
h3 = op1(h1, name=name + '_ops.' + str(2 * i) + '.')
h4 = op2(h2, name=name + '_ops.' + str(2 * i + 1) + '.')
if self.training and drop_prob > 0.:
if h3 != h1:
h3 = fluid.layers.dropout(
h3,
drop_prob,
dropout_implementation='upscale_in_train')
if h4 != h2:
h4 = fluid.layers.dropout(
h4,
drop_prob,
dropout_implementation='upscale_in_train')
s = h3 + h4
out += [s]
return fluid.layers.concat([out[i] for i in self._concat], axis=1)
def AuxiliaryHeadImageNet(input, num_classes, aux_name='auxiliary_head'):
relu_a = fluid.layers.relu(input)
pool_a = fluid.layers.pool2d(relu_a, 5, 'avg', 2)
conv2d_a = fluid.layers.conv2d(
pool_a, 128, 1, name=aux_name + '.features.2', bias_attr=False)
bn_a_name = aux_name + '.features.3'
bn_a = fluid.layers.batch_norm(
conv2d_a,
act='relu',
name=bn_a_name,
param_attr=ParamAttr(
initializer=Constant(1.), name=bn_a_name + '.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=bn_a_name + '.bias'),
moving_mean_name=bn_a_name + '.running_mean',
moving_variance_name=bn_a_name + '.running_var')
conv2d_b = fluid.layers.conv2d(
bn_a, 768, 2, name=aux_name + '.features.5', bias_attr=False)
bn_b_name = aux_name + '.features.6'
bn_b = fluid.layers.batch_norm(
conv2d_b,
act='relu',
name=bn_b_name,
param_attr=ParamAttr(
initializer=Constant(1.), name=bn_b_name + '.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=bn_b_name + '.bias'),
moving_mean_name=bn_b_name + '.running_mean',
moving_variance_name=bn_b_name + '.running_var')
pool_b = fluid.layers.adaptive_pool2d(bn_b, (1, 1), "avg")
fc_name = aux_name + '.classifier'
fc = fluid.layers.fc(pool_b,
num_classes,
name=fc_name,
param_attr=ParamAttr(
initializer=Normal(scale=1e-3),
name=fc_name + '.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name=fc_name + '.bias'))
return fc
def StemConv0(input, C_out):
conv_a = fluid.layers.conv2d(
input, C_out // 2, 3, stride=2, padding=1, bias_attr=False)
bn_a = fluid.layers.batch_norm(
conv_a,
act='relu',
param_attr=ParamAttr(
initializer=Constant(1.), name='stem0.1.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name='stem0.1.bias'),
moving_mean_name='stem0.1.running_mean',
moving_variance_name='stem0.1.running_var')
conv_b = fluid.layers.conv2d(
bn_a, C_out, 3, stride=2, padding=1, bias_attr=False)
bn_b = fluid.layers.batch_norm(
conv_b,
param_attr=ParamAttr(
initializer=Constant(1.), name='stem0.3.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name='stem0.3.bias'),
moving_mean_name='stem0.3.running_mean',
moving_variance_name='stem0.3.running_var')
return bn_b
def StemConv1(input, C_out):
relu_a = fluid.layers.relu(input)
conv_a = fluid.layers.conv2d(
relu_a, C_out, 3, stride=2, padding=1, bias_attr=False)
bn_a = fluid.layers.batch_norm(
conv_a,
param_attr=ParamAttr(
initializer=Constant(1.), name='stem1.1.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.), name='stem1.1.bias'),
moving_mean_name='stem1.1.running_mean',
moving_variance_name='stem1.1.running_var')
return bn_a
class NetworkImageNet(object):
def __init__(self, arch='DARTS_6M'):
self.class_num = 1000
self.init_channel = 48
self._layers = 14
self._auxiliary = False
self.drop_path_prob = 0
genotype = arch_dict[arch]
C = self.init_channel
layers = self._layers
C_prev_prev, C_prev, C_curr = C, C, C
self.cells = []
reduction_prev = True
for i in range(layers):
if i in [layers // 3, 2 * layers // 3]:
C_curr *= 2
reduction = True
else:
reduction = False
cell = Cell(genotype, C_prev_prev, C_prev, C_curr, reduction,
reduction_prev)
reduction_prev = reduction
self.cells += [cell]
C_prev_prev, C_prev = C_prev, cell.multiplier * C_curr
if i == 2 * layers // 3:
C_to_auxiliary = C_prev
def net(self, input, class_dim=1000, is_train=True):
self.logits_aux = None
num_channel = self.init_channel
s0 = StemConv0(input, num_channel)
s1 = StemConv1(s0, num_channel)
for i, cell in enumerate(self.cells):
name = 'cells.' + str(i) + '.'
s0, s1 = s1, cell.forward(s0, s1, self.drop_path_prob, is_train,
name)
if i == int(2 * self._layers // 3):
if self._auxiliary and is_train:
self.logits_aux = AuxiliaryHeadImageNet(s1, self.class_num)
out = fluid.layers.adaptive_pool2d(s1, (1, 1), "avg")
self.logits = fluid.layers.fc(out,
size=self.class_num,
param_attr=ParamAttr(
initializer=Normal(scale=1e-4),
name='classifier.weight'),
bias_attr=ParamAttr(
initializer=Constant(0.),
name='classifier.bias'))
return self.logits
def DARTS_GS_6M():
return NetworkImageNet(arch='DARTS_GS_6M')
def DARTS_GS_4M():
return NetworkImageNet(arch='DARTS_GS_4M')
ppcls/modeling/architectures/densenet.py
浏览文件 @ ea746480
...@@ -18,9 +18,11 @@ from __future__ import print_function ...@@ -18,9 +18,11 @@ from __future__ import print_function
import numpy as np import numpy as np
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
...@@ -29,7 +31,7 @@ __all__ = [ ...@@ -29,7 +31,7 @@ __all__ = [
] ]
class BNACConvLayer(fluid.dygraph.Layer): class BNACConvLayer(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -49,15 +51,14 @@ class BNACConvLayer(fluid.dygraph.Layer): ...@@ -49,15 +51,14 @@ class BNACConvLayer(fluid.dygraph.Layer):
moving_mean_name=name + '_bn_mean', moving_mean_name=name + '_bn_mean',
moving_variance_name=name + '_bn_variance') moving_variance_name=name + '_bn_variance')
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=pad, padding=pad,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
def forward(self, input): def forward(self, input):
...@@ -66,7 +67,7 @@ class BNACConvLayer(fluid.dygraph.Layer): ...@@ -66,7 +67,7 @@ class BNACConvLayer(fluid.dygraph.Layer):
return y return y
class DenseLayer(fluid.dygraph.Layer): class DenseLayer(nn.Layer):
def __init__(self, num_channels, growth_rate, bn_size, dropout, name=None): def __init__(self, num_channels, growth_rate, bn_size, dropout, name=None):
super(DenseLayer, self).__init__() super(DenseLayer, self).__init__()
self.dropout = dropout self.dropout = dropout
...@@ -88,18 +89,18 @@ class DenseLayer(fluid.dygraph.Layer): ...@@ -88,18 +89,18 @@ class DenseLayer(fluid.dygraph.Layer):
name=name + "_x2") name=name + "_x2")
if dropout: if dropout:
self.dropout_func = Dropout(p=dropout) self.dropout_func = Dropout(p=dropout, mode="downscale_in_infer")
def forward(self, input): def forward(self, input):
conv = self.bn_ac_func1(input) conv = self.bn_ac_func1(input)
conv = self.bn_ac_func2(conv) conv = self.bn_ac_func2(conv)
if self.dropout: if self.dropout:
conv = self.dropout_func(conv) conv = self.dropout_func(conv)
conv = fluid.layers.concat([input, conv], axis=1) conv = paddle.concat([input, conv], axis=1)
return conv return conv
class DenseBlock(fluid.dygraph.Layer): class DenseBlock(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_layers, num_layers,
...@@ -132,7 +133,7 @@ class DenseBlock(fluid.dygraph.Layer): ...@@ -132,7 +133,7 @@ class DenseBlock(fluid.dygraph.Layer):
return conv return conv
class TransitionLayer(fluid.dygraph.Layer): class TransitionLayer(nn.Layer):
def __init__(self, num_channels, num_output_features, name=None): def __init__(self, num_channels, num_output_features, name=None):
super(TransitionLayer, self).__init__() super(TransitionLayer, self).__init__()
...@@ -144,7 +145,7 @@ class TransitionLayer(fluid.dygraph.Layer): ...@@ -144,7 +145,7 @@ class TransitionLayer(fluid.dygraph.Layer):
stride=1, stride=1,
name=name) name=name)
self.pool2d_avg = Pool2D(pool_size=2, pool_stride=2, pool_type='avg') self.pool2d_avg = AvgPool2d(kernel_size=2, stride=2, padding=0)
def forward(self, input): def forward(self, input):
y = self.conv_ac_func(input) y = self.conv_ac_func(input)
...@@ -152,7 +153,7 @@ class TransitionLayer(fluid.dygraph.Layer): ...@@ -152,7 +153,7 @@ class TransitionLayer(fluid.dygraph.Layer):
return y return y
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -164,15 +165,14 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -164,15 +165,14 @@ class ConvBNLayer(fluid.dygraph.Layer):
name=None): name=None):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=pad, padding=pad,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
self._batch_norm = BatchNorm( self._batch_norm = BatchNorm(
num_filters, num_filters,
...@@ -188,7 +188,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -188,7 +188,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return y return y
class DenseNet(fluid.dygraph.Layer): class DenseNet(nn.Layer):
def __init__(self, layers=60, bn_size=4, dropout=0, class_dim=1000): def __init__(self, layers=60, bn_size=4, dropout=0, class_dim=1000):
super(DenseNet, self).__init__() super(DenseNet, self).__init__()
...@@ -214,8 +214,7 @@ class DenseNet(fluid.dygraph.Layer): ...@@ -214,8 +214,7 @@ class DenseNet(fluid.dygraph.Layer):
act='relu', act='relu',
name="conv1") name="conv1")
self.pool2d_max = Pool2D( self.pool2d_max = MaxPool2d(kernel_size=3, stride=2, padding=1)
pool_size=3, pool_stride=2, pool_padding=1, pool_type='max')
self.block_config = block_config self.block_config = block_config
...@@ -257,16 +256,15 @@ class DenseNet(fluid.dygraph.Layer): ...@@ -257,16 +256,15 @@ class DenseNet(fluid.dygraph.Layer):
moving_mean_name='conv5_blk_bn_mean', moving_mean_name='conv5_blk_bn_mean',
moving_variance_name='conv5_blk_bn_variance') moving_variance_name='conv5_blk_bn_variance')
self.pool2d_avg = Pool2D(pool_type='avg', global_pooling=True) self.pool2d_avg = AdaptiveAvgPool2d(1)
stdv = 1.0 / math.sqrt(num_features * 1.0) stdv = 1.0 / math.sqrt(num_features * 1.0)
self.out = Linear( self.out = Linear(
num_features, num_features,
class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv), initializer=Uniform(-stdv, stdv), name="fc_weights"),
name="fc_weights"),
bias_attr=ParamAttr(name="fc_offset")) bias_attr=ParamAttr(name="fc_offset"))
def forward(self, input): def forward(self, input):
...@@ -280,7 +278,7 @@ class DenseNet(fluid.dygraph.Layer): ...@@ -280,7 +278,7 @@ class DenseNet(fluid.dygraph.Layer):
conv = self.batch_norm(conv) conv = self.batch_norm(conv)
y = self.pool2d_avg(conv) y = self.pool2d_avg(conv)
y = fluid.layers.reshape(y, shape=[0, -1]) y = paddle.reshape(y, shape=[0, -1])
y = self.out(y) y = self.out(y)
return y return y
......
ppcls/modeling/architectures/distillation_models.py
浏览文件 @ ea746480
...@@ -19,8 +19,7 @@ from __future__ import print_function ...@@ -19,8 +19,7 @@ from __future__ import print_function
import math import math
import paddle import paddle
import paddle.fluid as fluid import paddle.nn as nn
from paddle.fluid.param_attr import ParamAttr
from .resnet_vd import ResNet50_vd from .resnet_vd import ResNet50_vd
from .mobilenet_v3 import MobileNetV3_large_x1_0 from .mobilenet_v3 import MobileNetV3_large_x1_0
...@@ -32,7 +31,7 @@ __all__ = [ ...@@ -32,7 +31,7 @@ __all__ = [
] ]
class ResNet50_vd_distill_MobileNetV3_large_x1_0(fluid.dygraph.Layer): class ResNet50_vd_distill_MobileNetV3_large_x1_0(nn.Layer):
def __init__(self, class_dim=1000, **args): def __init__(self, class_dim=1000, **args):
super(ResNet50_vd_distill_MobileNetV3_large_x1_0, self).__init__() super(ResNet50_vd_distill_MobileNetV3_large_x1_0, self).__init__()
...@@ -49,7 +48,7 @@ class ResNet50_vd_distill_MobileNetV3_large_x1_0(fluid.dygraph.Layer): ...@@ -49,7 +48,7 @@ class ResNet50_vd_distill_MobileNetV3_large_x1_0(fluid.dygraph.Layer):
return teacher_label, student_label return teacher_label, student_label
class ResNeXt101_32x16d_wsl_distill_ResNet50_vd(fluid.dygraph.Layer): class ResNeXt101_32x16d_wsl_distill_ResNet50_vd(nn.Layer):
def __init__(self, class_dim=1000, **args): def __init__(self, class_dim=1000, **args):
super(ResNet50_vd_distill_MobileNetV3_large_x1_0, self).__init__() super(ResNet50_vd_distill_MobileNetV3_large_x1_0, self).__init__()
......
ppcls/modeling/architectures/dpn.py
浏览文件 @ ea746480
...@@ -19,9 +19,11 @@ from __future__ import print_function ...@@ -19,9 +19,11 @@ from __future__ import print_function
import numpy as np import numpy as np
import sys import sys
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear from paddle.nn import Conv2d, BatchNorm, Linear
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
...@@ -35,7 +37,7 @@ __all__ = [ ...@@ -35,7 +37,7 @@ __all__ = [
] ]
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -47,15 +49,14 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -47,15 +49,14 @@ class ConvBNLayer(fluid.dygraph.Layer):
name=None): name=None):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=pad, padding=pad,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
self._batch_norm = BatchNorm( self._batch_norm = BatchNorm(
num_filters, num_filters,
...@@ -71,7 +72,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -71,7 +72,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return y return y
class BNACConvLayer(fluid.dygraph.Layer): class BNACConvLayer(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -83,7 +84,6 @@ class BNACConvLayer(fluid.dygraph.Layer): ...@@ -83,7 +84,6 @@ class BNACConvLayer(fluid.dygraph.Layer):
name=None): name=None):
super(BNACConvLayer, self).__init__() super(BNACConvLayer, self).__init__()
self.num_channels = num_channels self.num_channels = num_channels
self.name = name
self._batch_norm = BatchNorm( self._batch_norm = BatchNorm(
num_channels, num_channels,
...@@ -93,15 +93,14 @@ class BNACConvLayer(fluid.dygraph.Layer): ...@@ -93,15 +93,14 @@ class BNACConvLayer(fluid.dygraph.Layer):
moving_mean_name=name + '_bn_mean', moving_mean_name=name + '_bn_mean',
moving_variance_name=name + '_bn_variance') moving_variance_name=name + '_bn_variance')
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=pad, padding=pad,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
def forward(self, input): def forward(self, input):
...@@ -110,7 +109,7 @@ class BNACConvLayer(fluid.dygraph.Layer): ...@@ -110,7 +109,7 @@ class BNACConvLayer(fluid.dygraph.Layer):
return y return y
class DualPathFactory(fluid.dygraph.Layer): class DualPathFactory(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_1x1_a, num_1x1_a,
...@@ -183,14 +182,14 @@ class DualPathFactory(fluid.dygraph.Layer): ...@@ -183,14 +182,14 @@ class DualPathFactory(fluid.dygraph.Layer):
def forward(self, input): def forward(self, input):
# PROJ # PROJ
if isinstance(input, list): if isinstance(input, list):
data_in = fluid.layers.concat([input[0], input[1]], axis=1) data_in = paddle.concat([input[0], input[1]], axis=1)
else: else:
data_in = input data_in = input
if self.has_proj: if self.has_proj:
c1x1_w = self.c1x1_w_func(data_in) c1x1_w = self.c1x1_w_func(data_in)
data_o1, data_o2 = fluid.layers.split( data_o1, data_o2 = paddle.split(
c1x1_w, num_or_sections=[self.num_1x1_c, 2 * self.inc], dim=1) c1x1_w, num_or_sections=[self.num_1x1_c, 2 * self.inc], axis=1)
else: else:
data_o1 = input[0] data_o1 = input[0]
data_o2 = input[1] data_o2 = input[1]
...@@ -199,17 +198,17 @@ class DualPathFactory(fluid.dygraph.Layer): ...@@ -199,17 +198,17 @@ class DualPathFactory(fluid.dygraph.Layer):
c3x3_b = self.c3x3_b_func(c1x1_a) c3x3_b = self.c3x3_b_func(c1x1_a)
c1x1_c = self.c1x1_c_func(c3x3_b) c1x1_c = self.c1x1_c_func(c3x3_b)
c1x1_c1, c1x1_c2 = fluid.layers.split( c1x1_c1, c1x1_c2 = paddle.split(
c1x1_c, num_or_sections=[self.num_1x1_c, self.inc], dim=1) c1x1_c, num_or_sections=[self.num_1x1_c, self.inc], axis=1)
# OUTPUTS # OUTPUTS
summ = fluid.layers.elementwise_add(x=data_o1, y=c1x1_c1) summ = paddle.elementwise_add(x=data_o1, y=c1x1_c1)
dense = fluid.layers.concat([data_o2, c1x1_c2], axis=1) dense = paddle.concat([data_o2, c1x1_c2], axis=1)
# tensor, channels # tensor, channels
return [summ, dense] return [summ, dense]
class DPN(fluid.dygraph.Layer): class DPN(nn.Layer):
def __init__(self, layers=60, class_dim=1000): def __init__(self, layers=60, class_dim=1000):
super(DPN, self).__init__() super(DPN, self).__init__()
...@@ -237,8 +236,7 @@ class DPN(fluid.dygraph.Layer): ...@@ -237,8 +236,7 @@ class DPN(fluid.dygraph.Layer):
act='relu', act='relu',
name="conv1") name="conv1")
self.pool2d_max = Pool2D( self.pool2d_max = MaxPool2d(kernel_size=3, stride=2, padding=1)
pool_size=3, pool_stride=2, pool_padding=1, pool_type='max')
num_channel_dpn = init_num_filter num_channel_dpn = init_num_filter
...@@ -303,16 +301,15 @@ class DPN(fluid.dygraph.Layer): ...@@ -303,16 +301,15 @@ class DPN(fluid.dygraph.Layer):
moving_mean_name='final_concat_bn_mean', moving_mean_name='final_concat_bn_mean',
moving_variance_name='final_concat_bn_variance') moving_variance_name='final_concat_bn_variance')
self.pool2d_avg = Pool2D(pool_type='avg', global_pooling=True) self.pool2d_avg = AdaptiveAvgPool2d(1)
stdv = 0.01 stdv = 0.01
self.out = Linear( self.out = Linear(
out_channel, out_channel,
class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv), initializer=Uniform(-stdv, stdv), name="fc_weights"),
name="fc_weights"),
bias_attr=ParamAttr(name="fc_offset")) bias_attr=ParamAttr(name="fc_offset"))
def forward(self, input): def forward(self, input):
...@@ -327,11 +324,11 @@ class DPN(fluid.dygraph.Layer): ...@@ -327,11 +324,11 @@ class DPN(fluid.dygraph.Layer):
convX_x_x = self.dpn_func_list[dpn_idx](convX_x_x) convX_x_x = self.dpn_func_list[dpn_idx](convX_x_x)
dpn_idx += 1 dpn_idx += 1
conv5_x_x = fluid.layers.concat(convX_x_x, axis=1) conv5_x_x = paddle.concat(convX_x_x, axis=1)
conv5_x_x = self.conv5_x_x_bn(conv5_x_x) conv5_x_x = self.conv5_x_x_bn(conv5_x_x)
y = self.pool2d_avg(conv5_x_x) y = self.pool2d_avg(conv5_x_x)
y = fluid.layers.reshape(y, shape=[0, -1]) y = paddle.reshape(y, shape=[0, -1])
y = self.out(y) y = self.out(y)
return y return y
......
ppcls/modeling/architectures/efficientnet.py
浏览文件 @ ea746480
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
import math import math
import collections import collections
import re import re
...@@ -242,15 +244,14 @@ def _drop_connect(inputs, prob, is_test): ...@@ -242,15 +244,14 @@ def _drop_connect(inputs, prob, is_test):
if is_test: if is_test:
return inputs return inputs
keep_prob = 1.0 - prob keep_prob = 1.0 - prob
inputs_shape = fluid.layers.shape(inputs) inputs_shape = paddle.shape(inputs)
random_tensor = keep_prob + fluid.layers.uniform_random( random_tensor = keep_prob + paddle.rand(shape=[inputs_shape[0], 1, 1, 1])
shape=[inputs_shape[0], 1, 1, 1], min=0., max=1.) binary_tensor = paddle.floor(random_tensor)
binary_tensor = fluid.layers.floor(random_tensor)
output = inputs / keep_prob * binary_tensor output = inputs / keep_prob * binary_tensor
return output return output
class Conv2ds(fluid.dygraph.Layer): class Conv2ds(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
output_channels, output_channels,
...@@ -265,6 +266,8 @@ class Conv2ds(fluid.dygraph.Layer): ...@@ -265,6 +266,8 @@ class Conv2ds(fluid.dygraph.Layer):
model_name=None, model_name=None,
cur_stage=None): cur_stage=None):
super(Conv2ds, self).__init__() super(Conv2ds, self).__init__()
assert act in [None, "swish", "sigmoid"]
self.act = act
param_attr, bias_attr = initial_type(name=name, use_bias=use_bias) param_attr, bias_attr = initial_type(name=name, use_bias=use_bias)
...@@ -296,25 +299,31 @@ class Conv2ds(fluid.dygraph.Layer): ...@@ -296,25 +299,31 @@ class Conv2ds(fluid.dygraph.Layer):
else: else:
padding = padding_type padding = padding_type
self._conv = Conv2D( groups = 1 if groups is None else groups
self._conv = Conv2d(
input_channels, input_channels,
output_channels, output_channels,
filter_size, filter_size,
groups=groups, groups=groups,
stride=stride, stride=stride,
act=act, # act=act,
padding=padding, padding=padding,
param_attr=param_attr, weight_attr=param_attr,
bias_attr=bias_attr) bias_attr=bias_attr)
def forward(self, inputs): def forward(self, inputs):
x = self._conv(inputs) x = self._conv(inputs)
if self.act == "swish":
x = F.swish(x)
elif self.act == "sigmoid":
x = F.sigmoid(x)
if self.need_crop: if self.need_crop:
x = x[:, :, 1:, 1:] x = x[:, :, 1:, 1:]
return x return x
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
filter_size, filter_size,
...@@ -369,7 +378,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -369,7 +378,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return self._conv(inputs) return self._conv(inputs)
class ExpandConvNorm(fluid.dygraph.Layer): class ExpandConvNorm(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
block_args, block_args,
...@@ -402,7 +411,7 @@ class ExpandConvNorm(fluid.dygraph.Layer): ...@@ -402,7 +411,7 @@ class ExpandConvNorm(fluid.dygraph.Layer):
return inputs return inputs
class DepthwiseConvNorm(fluid.dygraph.Layer): class DepthwiseConvNorm(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
block_args, block_args,
...@@ -436,7 +445,7 @@ class DepthwiseConvNorm(fluid.dygraph.Layer): ...@@ -436,7 +445,7 @@ class DepthwiseConvNorm(fluid.dygraph.Layer):
return self._conv(inputs) return self._conv(inputs)
class ProjectConvNorm(fluid.dygraph.Layer): class ProjectConvNorm(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
block_args, block_args,
...@@ -464,7 +473,7 @@ class ProjectConvNorm(fluid.dygraph.Layer): ...@@ -464,7 +473,7 @@ class ProjectConvNorm(fluid.dygraph.Layer):
return self._conv(inputs) return self._conv(inputs)
class SEBlock(fluid.dygraph.Layer): class SEBlock(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
num_squeezed_channels, num_squeezed_channels,
...@@ -475,8 +484,7 @@ class SEBlock(fluid.dygraph.Layer): ...@@ -475,8 +484,7 @@ class SEBlock(fluid.dygraph.Layer):
cur_stage=None): cur_stage=None):
super(SEBlock, self).__init__() super(SEBlock, self).__init__()
self._pool = Pool2D( self._pool = AdaptiveAvgPool2d(1)
pool_type="avg", global_pooling=True, use_cudnn=False)
self._conv1 = Conv2ds( self._conv1 = Conv2ds(
input_channels, input_channels,
num_squeezed_channels, num_squeezed_channels,
...@@ -499,10 +507,10 @@ class SEBlock(fluid.dygraph.Layer): ...@@ -499,10 +507,10 @@ class SEBlock(fluid.dygraph.Layer):
x = self._pool(inputs) x = self._pool(inputs)
x = self._conv1(x) x = self._conv1(x)
x = self._conv2(x) x = self._conv2(x)
return fluid.layers.elementwise_mul(inputs, x) return paddle.multiply(inputs, x)
class MbConvBlock(fluid.dygraph.Layer): class MbConvBlock(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
block_args, block_args,
...@@ -565,9 +573,9 @@ class MbConvBlock(fluid.dygraph.Layer): ...@@ -565,9 +573,9 @@ class MbConvBlock(fluid.dygraph.Layer):
x = inputs x = inputs
if self.expand_ratio != 1: if self.expand_ratio != 1:
x = self._ecn(x) x = self._ecn(x)
x = fluid.layers.swish(x) x = F.swish(x)
x = self._dcn(x) x = self._dcn(x)
x = fluid.layers.swish(x) x = F.swish(x)
if self.has_se: if self.has_se:
x = self._se(x) x = self._se(x)
x = self._pcn(x) x = self._pcn(x)
...@@ -576,11 +584,11 @@ class MbConvBlock(fluid.dygraph.Layer): ...@@ -576,11 +584,11 @@ class MbConvBlock(fluid.dygraph.Layer):
self.block_args.input_filters == self.block_args.output_filters: self.block_args.input_filters == self.block_args.output_filters:
if self.drop_connect_rate: if self.drop_connect_rate:
x = _drop_connect(x, self.drop_connect_rate, self.is_test) x = _drop_connect(x, self.drop_connect_rate, self.is_test)
x = fluid.layers.elementwise_add(x, inputs) x = paddle.elementwise_add(x, inputs)
return x return x
class ConvStemNorm(fluid.dygraph.Layer): class ConvStemNorm(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
padding_type, padding_type,
...@@ -608,7 +616,7 @@ class ConvStemNorm(fluid.dygraph.Layer): ...@@ -608,7 +616,7 @@ class ConvStemNorm(fluid.dygraph.Layer):
return self._conv(inputs) return self._conv(inputs)
class ExtractFeatures(fluid.dygraph.Layer): class ExtractFeatures(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
_block_args, _block_args,
...@@ -694,13 +702,13 @@ class ExtractFeatures(fluid.dygraph.Layer): ...@@ -694,13 +702,13 @@ class ExtractFeatures(fluid.dygraph.Layer):
def forward(self, inputs): def forward(self, inputs):
x = self._conv_stem(inputs) x = self._conv_stem(inputs)
x = fluid.layers.swish(x) x = F.swish(x)
for _mc_block in self.conv_seq: for _mc_block in self.conv_seq:
x = _mc_block(x) x = _mc_block(x)
return x return x
class EfficientNet(fluid.dygraph.Layer): class EfficientNet(nn.Layer):
def __init__(self, def __init__(self,
name="b0", name="b0",
is_test=True, is_test=True,
...@@ -753,18 +761,17 @@ class EfficientNet(fluid.dygraph.Layer): ...@@ -753,18 +761,17 @@ class EfficientNet(fluid.dygraph.Layer):
bn_name="_bn1", bn_name="_bn1",
model_name=self.name, model_name=self.name,
cur_stage=7) cur_stage=7)
self._pool = Pool2D(pool_type="avg", global_pooling=True) self._pool = AdaptiveAvgPool2d(1)
if self._global_params.dropout_rate: if self._global_params.dropout_rate:
self._drop = Dropout( self._drop = Dropout(
p=self._global_params.dropout_rate, p=self._global_params.dropout_rate, mode="upscale_in_train")
dropout_implementation="upscale_in_train")
param_attr, bias_attr = init_fc_layer("_fc") param_attr, bias_attr = init_fc_layer("_fc")
self._fc = Linear( self._fc = Linear(
output_channels, output_channels,
class_dim, class_dim,
param_attr=param_attr, weight_attr=param_attr,
bias_attr=bias_attr) bias_attr=bias_attr)
def forward(self, inputs): def forward(self, inputs):
...@@ -773,7 +780,7 @@ class EfficientNet(fluid.dygraph.Layer): ...@@ -773,7 +780,7 @@ class EfficientNet(fluid.dygraph.Layer):
x = self._pool(x) x = self._pool(x)
if self._global_params.dropout_rate: if self._global_params.dropout_rate:
x = self._drop(x) x = self._drop(x)
x = fluid.layers.squeeze(x, axes=[2, 3]) x = paddle.squeeze(x, axis=[2, 3])
x = self._fc(x) x = self._fc(x)
return x return x
......
ppcls/modeling/architectures/googlenet.py
浏览文件 @ ea746480
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
__all__ = ['GoogLeNet'] __all__ = ['GoogLeNet']
...@@ -10,12 +14,11 @@ __all__ = ['GoogLeNet'] ...@@ -10,12 +14,11 @@ __all__ = ['GoogLeNet']
def xavier(channels, filter_size, name): def xavier(channels, filter_size, name):
stdv = (3.0 / (filter_size**2 * channels))**0.5 stdv = (3.0 / (filter_size**2 * channels))**0.5
param_attr = ParamAttr( param_attr = ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv), initializer=Uniform(-stdv, stdv), name=name + "_weights")
name=name + "_weights")
return param_attr return param_attr
class ConvLayer(fluid.dygraph.Layer): class ConvLayer(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -26,15 +29,14 @@ class ConvLayer(fluid.dygraph.Layer): ...@@ -26,15 +29,14 @@ class ConvLayer(fluid.dygraph.Layer):
name=None): name=None):
super(ConvLayer, self).__init__() super(ConvLayer, self).__init__()
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
def forward(self, inputs): def forward(self, inputs):
...@@ -42,7 +44,7 @@ class ConvLayer(fluid.dygraph.Layer): ...@@ -42,7 +44,7 @@ class ConvLayer(fluid.dygraph.Layer):
return y return y
class Inception(fluid.dygraph.Layer): class Inception(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
output_channels, output_channels,
...@@ -71,8 +73,8 @@ class Inception(fluid.dygraph.Layer): ...@@ -71,8 +73,8 @@ class Inception(fluid.dygraph.Layer):
name="inception_" + name + "_5x5_reduce") name="inception_" + name + "_5x5_reduce")
self._conv5 = ConvLayer( self._conv5 = ConvLayer(
filter5R, filter5, 5, name="inception_" + name + "_5x5") filter5R, filter5, 5, name="inception_" + name + "_5x5")
self._pool = Pool2D( self._pool = MaxPool2d(kernel_size=3, stride=1, padding=1)
pool_size=3, pool_type="max", pool_stride=1, pool_padding=1)
self._convprj = ConvLayer( self._convprj = ConvLayer(
input_channels, proj, 1, name="inception_" + name + "_3x3_proj") input_channels, proj, 1, name="inception_" + name + "_3x3_proj")
...@@ -88,16 +90,16 @@ class Inception(fluid.dygraph.Layer): ...@@ -88,16 +90,16 @@ class Inception(fluid.dygraph.Layer):
pool = self._pool(inputs) pool = self._pool(inputs)
convprj = self._convprj(pool) convprj = self._convprj(pool)
cat = fluid.layers.concat([conv1, conv3, conv5, convprj], axis=1) cat = paddle.concat([conv1, conv3, conv5, convprj], axis=1)
cat = fluid.layers.relu(cat) cat = F.relu(cat)
return cat return cat
class GoogleNetDY(fluid.dygraph.Layer): class GoogleNetDY(nn.Layer):
def __init__(self, class_dim=1000): def __init__(self, class_dim=1000):
super(GoogleNetDY, self).__init__() super(GoogleNetDY, self).__init__()
self._conv = ConvLayer(3, 64, 7, 2, name="conv1") self._conv = ConvLayer(3, 64, 7, 2, name="conv1")
self._pool = Pool2D(pool_size=3, pool_type="max", pool_stride=2) self._pool = MaxPool2d(kernel_size=3, stride=2)
self._conv_1 = ConvLayer(64, 64, 1, name="conv2_1x1") self._conv_1 = ConvLayer(64, 64, 1, name="conv2_1x1")
self._conv_2 = ConvLayer(64, 192, 3, name="conv2_3x3") self._conv_2 = ConvLayer(64, 192, 3, name="conv2_3x3")
...@@ -122,42 +124,39 @@ class GoogleNetDY(fluid.dygraph.Layer): ...@@ -122,42 +124,39 @@ class GoogleNetDY(fluid.dygraph.Layer):
self._ince5b = Inception( self._ince5b = Inception(
832, 832, 384, 192, 384, 48, 128, 128, name="ince5b") 832, 832, 384, 192, 384, 48, 128, 128, name="ince5b")
self._pool_5 = Pool2D(pool_size=7, pool_type='avg', pool_stride=7) self._pool_5 = AvgPool2d(kernel_size=7, stride=7)
self._drop = fluid.dygraph.Dropout(p=0.4) self._drop = Dropout(p=0.4, mode="downscale_in_infer")
self._fc_out = Linear( self._fc_out = Linear(
1024, 1024,
class_dim, class_dim,
param_attr=xavier(1024, 1, "out"), weight_attr=xavier(1024, 1, "out"),
bias_attr=ParamAttr(name="out_offset"), bias_attr=ParamAttr(name="out_offset"))
act="softmax") self._pool_o1 = AvgPool2d(kernel_size=5, stride=3)
self._pool_o1 = Pool2D(pool_size=5, pool_stride=3, pool_type="avg")
self._conv_o1 = ConvLayer(512, 128, 1, name="conv_o1") self._conv_o1 = ConvLayer(512, 128, 1, name="conv_o1")
self._fc_o1 = Linear( self._fc_o1 = Linear(
1152, 1152,
1024, 1024,
param_attr=xavier(2048, 1, "fc_o1"), weight_attr=xavier(2048, 1, "fc_o1"),
bias_attr=ParamAttr(name="fc_o1_offset"), bias_attr=ParamAttr(name="fc_o1_offset"))
act="relu") self._drop_o1 = Dropout(p=0.7, mode="downscale_in_infer")
self._drop_o1 = fluid.dygraph.Dropout(p=0.7)
self._out1 = Linear( self._out1 = Linear(
1024, 1024,
class_dim, class_dim,
param_attr=xavier(1024, 1, "out1"), weight_attr=xavier(1024, 1, "out1"),
bias_attr=ParamAttr(name="out1_offset"), bias_attr=ParamAttr(name="out1_offset"))
act="softmax") self._pool_o2 = AvgPool2d(kernel_size=5, stride=3)
self._pool_o2 = Pool2D(pool_size=5, pool_stride=3, pool_type='avg')
self._conv_o2 = ConvLayer(528, 128, 1, name="conv_o2") self._conv_o2 = ConvLayer(528, 128, 1, name="conv_o2")
self._fc_o2 = Linear( self._fc_o2 = Linear(
1152, 1152,
1024, 1024,
param_attr=xavier(2048, 1, "fc_o2"), weight_attr=xavier(2048, 1, "fc_o2"),
bias_attr=ParamAttr(name="fc_o2_offset")) bias_attr=ParamAttr(name="fc_o2_offset"))
self._drop_o2 = fluid.dygraph.Dropout(p=0.7) self._drop_o2 = Dropout(p=0.7, mode="downscale_in_infer")
self._out2 = Linear( self._out2 = Linear(
1024, 1024,
class_dim, class_dim,
param_attr=xavier(1024, 1, "out2"), weight_attr=xavier(1024, 1, "out2"),
bias_attr=ParamAttr(name="out2_offset")) bias_attr=ParamAttr(name="out2_offset"))
def forward(self, inputs): def forward(self, inputs):
...@@ -183,19 +182,22 @@ class GoogleNetDY(fluid.dygraph.Layer): ...@@ -183,19 +182,22 @@ class GoogleNetDY(fluid.dygraph.Layer):
x = self._pool_5(ince5b) x = self._pool_5(ince5b)
x = self._drop(x) x = self._drop(x)
x = fluid.layers.squeeze(x, axes=[2, 3]) x = paddle.squeeze(x, axis=[2, 3])
out = self._fc_out(x) out = self._fc_out(x)
out = F.softmax(out)
x = self._pool_o1(ince4a) x = self._pool_o1(ince4a)
x = self._conv_o1(x) x = self._conv_o1(x)
x = fluid.layers.flatten(x) x = paddle.flatten(x, start_axis=1, stop_axis=-1)
x = self._fc_o1(x) x = self._fc_o1(x)
x = F.relu(x)
x = self._drop_o1(x) x = self._drop_o1(x)
out1 = self._out1(x) out1 = self._out1(x)
out1 = F.softmax(out1)
x = self._pool_o2(ince4d) x = self._pool_o2(ince4d)
x = self._conv_o2(x) x = self._conv_o2(x)
x = fluid.layers.flatten(x) x = paddle.flatten(x, start_axis=1, stop_axis=-1)
x = self._fc_o2(x) x = self._fc_o2(x)
x = self._drop_o2(x) x = self._drop_o2(x)
out2 = self._out2(x) out2 = self._out2(x)
......
ppcls/modeling/architectures/hrnet.py
浏览文件 @ ea746480
...@@ -18,9 +18,12 @@ from __future__ import print_function ...@@ -18,9 +18,12 @@ from __future__ import print_function
import numpy as np import numpy as np
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
...@@ -44,7 +47,7 @@ __all__ = [ ...@@ -44,7 +47,7 @@ __all__ = [
] ]
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -55,15 +58,14 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -55,15 +58,14 @@ class ConvBNLayer(fluid.dygraph.Layer):
name=None): name=None):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
bn_name = name + '_bn' bn_name = name + '_bn'
self._batch_norm = BatchNorm( self._batch_norm = BatchNorm(
...@@ -80,7 +82,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -80,7 +82,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return y return y
class Layer1(fluid.dygraph.Layer): class Layer1(nn.Layer):
def __init__(self, num_channels, has_se=False, name=None): def __init__(self, num_channels, has_se=False, name=None):
super(Layer1, self).__init__() super(Layer1, self).__init__()
...@@ -105,7 +107,7 @@ class Layer1(fluid.dygraph.Layer): ...@@ -105,7 +107,7 @@ class Layer1(fluid.dygraph.Layer):
return conv return conv
class TransitionLayer(fluid.dygraph.Layer): class TransitionLayer(nn.Layer):
def __init__(self, in_channels, out_channels, name=None): def __init__(self, in_channels, out_channels, name=None):
super(TransitionLayer, self).__init__() super(TransitionLayer, self).__init__()
...@@ -148,7 +150,7 @@ class TransitionLayer(fluid.dygraph.Layer): ...@@ -148,7 +150,7 @@ class TransitionLayer(fluid.dygraph.Layer):
return outs return outs
class Branches(fluid.dygraph.Layer): class Branches(nn.Layer):
def __init__(self, def __init__(self,
block_num, block_num,
in_channels, in_channels,
...@@ -183,7 +185,7 @@ class Branches(fluid.dygraph.Layer): ...@@ -183,7 +185,7 @@ class Branches(fluid.dygraph.Layer):
return outs return outs
class BottleneckBlock(fluid.dygraph.Layer): class BottleneckBlock(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -243,11 +245,11 @@ class BottleneckBlock(fluid.dygraph.Layer): ...@@ -243,11 +245,11 @@ class BottleneckBlock(fluid.dygraph.Layer):
if self.has_se: if self.has_se:
conv3 = self.se(conv3) conv3 = self.se(conv3)
y = fluid.layers.elementwise_add(x=conv3, y=residual, act="relu") y = paddle.elementwise_add(x=conv3, y=residual, act="relu")
return y return y
class BasicBlock(fluid.dygraph.Layer): class BasicBlock(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -301,15 +303,15 @@ class BasicBlock(fluid.dygraph.Layer): ...@@ -301,15 +303,15 @@ class BasicBlock(fluid.dygraph.Layer):
if self.has_se: if self.has_se:
conv2 = self.se(conv2) conv2 = self.se(conv2)
y = fluid.layers.elementwise_add(x=conv2, y=residual, act="relu") y = paddle.elementwise_add(x=conv2, y=residual, act="relu")
return y return y
class SELayer(fluid.dygraph.Layer): class SELayer(nn.Layer):
def __init__(self, num_channels, num_filters, reduction_ratio, name=None): def __init__(self, num_channels, num_filters, reduction_ratio, name=None):
super(SELayer, self).__init__() super(SELayer, self).__init__()
self.pool2d_gap = Pool2D(pool_type='avg', global_pooling=True) self.pool2d_gap = AdaptiveAvgPool2d(1)
self._num_channels = num_channels self._num_channels = num_channels
...@@ -320,8 +322,7 @@ class SELayer(fluid.dygraph.Layer): ...@@ -320,8 +322,7 @@ class SELayer(fluid.dygraph.Layer):
med_ch, med_ch,
act="relu", act="relu",
param_attr=ParamAttr( param_attr=ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv), initializer=Uniform(-stdv, stdv), name=name + "_sqz_weights"),
name=name + "_sqz_weights"),
bias_attr=ParamAttr(name=name + '_sqz_offset')) bias_attr=ParamAttr(name=name + '_sqz_offset'))
stdv = 1.0 / math.sqrt(med_ch * 1.0) stdv = 1.0 / math.sqrt(med_ch * 1.0)
...@@ -330,22 +331,21 @@ class SELayer(fluid.dygraph.Layer): ...@@ -330,22 +331,21 @@ class SELayer(fluid.dygraph.Layer):
num_filters, num_filters,
act="sigmoid", act="sigmoid",
param_attr=ParamAttr( param_attr=ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv), initializer=Uniform(-stdv, stdv), name=name + "_exc_weights"),
name=name + "_exc_weights"),
bias_attr=ParamAttr(name=name + '_exc_offset')) bias_attr=ParamAttr(name=name + '_exc_offset'))
def forward(self, input): def forward(self, input):
pool = self.pool2d_gap(input) pool = self.pool2d_gap(input)
pool = fluid.layers.reshape(pool, shape=[-1, self._num_channels]) pool = paddle.reshape(pool, shape=[-1, self._num_channels])
squeeze = self.squeeze(pool) squeeze = self.squeeze(pool)
excitation = self.excitation(squeeze) excitation = self.excitation(squeeze)
excitation = fluid.layers.reshape( excitation = paddle.reshape(
excitation, shape=[-1, self._num_channels, 1, 1]) excitation, shape=[-1, self._num_channels, 1, 1])
out = input * excitation out = input * excitation
return out return out
class Stage(fluid.dygraph.Layer): class Stage(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_modules, num_modules,
...@@ -386,7 +386,7 @@ class Stage(fluid.dygraph.Layer): ...@@ -386,7 +386,7 @@ class Stage(fluid.dygraph.Layer):
return out return out
class HighResolutionModule(fluid.dygraph.Layer): class HighResolutionModule(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -414,7 +414,7 @@ class HighResolutionModule(fluid.dygraph.Layer): ...@@ -414,7 +414,7 @@ class HighResolutionModule(fluid.dygraph.Layer):
return out return out
class FuseLayers(fluid.dygraph.Layer): class FuseLayers(nn.Layer):
def __init__(self, def __init__(self,
in_channels, in_channels,
out_channels, out_channels,
...@@ -482,8 +482,8 @@ class FuseLayers(fluid.dygraph.Layer): ...@@ -482,8 +482,8 @@ class FuseLayers(fluid.dygraph.Layer):
y = self.residual_func_list[residual_func_idx](input[j]) y = self.residual_func_list[residual_func_idx](input[j])
residual_func_idx += 1 residual_func_idx += 1
y = fluid.layers.resize_nearest(input=y, scale=2**(j - i)) y = F.resize_nearest(input=y, scale=2**(j - i))
residual = fluid.layers.elementwise_add( residual = paddle.elementwise_add(
x=residual, y=y, act=None) x=residual, y=y, act=None)
elif j < i: elif j < i:
y = input[j] y = input[j]
...@@ -491,16 +491,16 @@ class FuseLayers(fluid.dygraph.Layer): ...@@ -491,16 +491,16 @@ class FuseLayers(fluid.dygraph.Layer):
y = self.residual_func_list[residual_func_idx](y) y = self.residual_func_list[residual_func_idx](y)
residual_func_idx += 1 residual_func_idx += 1
residual = fluid.layers.elementwise_add( residual = paddle.elementwise_add(
x=residual, y=y, act=None) x=residual, y=y, act=None)
residual = fluid.layers.relu(residual) residual = F.relu(residual)
outs.append(residual) outs.append(residual)
return outs return outs
class LastClsOut(fluid.dygraph.Layer): class LastClsOut(nn.Layer):
def __init__(self, def __init__(self,
num_channel_list, num_channel_list,
has_se, has_se,
...@@ -528,7 +528,7 @@ class LastClsOut(fluid.dygraph.Layer): ...@@ -528,7 +528,7 @@ class LastClsOut(fluid.dygraph.Layer):
return outs return outs
class HRNet(fluid.dygraph.Layer): class HRNet(nn.Layer):
def __init__(self, width=18, has_se=False, class_dim=1000): def __init__(self, width=18, has_se=False, class_dim=1000):
super(HRNet, self).__init__() super(HRNet, self).__init__()
...@@ -623,16 +623,15 @@ class HRNet(fluid.dygraph.Layer): ...@@ -623,16 +623,15 @@ class HRNet(fluid.dygraph.Layer):
stride=1, stride=1,
name="cls_head_last_conv") name="cls_head_last_conv")
self.pool2d_avg = Pool2D(pool_type='avg', global_pooling=True) self.pool2d_avg = AdaptiveAvgPool2d(1)
stdv = 1.0 / math.sqrt(2048 * 1.0) stdv = 1.0 / math.sqrt(2048 * 1.0)
self.out = Linear( self.out = Linear(
2048, 2048,
class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv), initializer=Uniform(-stdv, stdv), name="fc_weights"),
name="fc_weights"),
bias_attr=ParamAttr(name="fc_offset")) bias_attr=ParamAttr(name="fc_offset"))
def forward(self, input): def forward(self, input):
...@@ -658,7 +657,7 @@ class HRNet(fluid.dygraph.Layer): ...@@ -658,7 +657,7 @@ class HRNet(fluid.dygraph.Layer):
y = self.conv_last(y) y = self.conv_last(y)
y = self.pool2d_avg(y) y = self.pool2d_avg(y)
y = fluid.layers.reshape(y, shape=[0, -1]) y = paddle.reshape(y, shape=[0, -1])
y = self.out(y) y = self.out(y)
return y return y
......
ppcls/modeling/architectures/inception_v4.py
浏览文件 @ ea746480
# 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 paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
__all__ = ["InceptionV4"] __all__ = ["InceptionV4"]
class ConvBNLayer(fluid.dygraph.Layer):
class ConvBNLayer(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -18,15 +36,14 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -18,15 +36,14 @@ class ConvBNLayer(fluid.dygraph.Layer):
name=None): name=None):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=padding, padding=padding,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
bn_name = name + "_bn" bn_name = name + "_bn"
self._batch_norm = BatchNorm( self._batch_norm = BatchNorm(
...@@ -43,7 +60,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -43,7 +60,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return y return y
class InceptionStem(fluid.dygraph.Layer): class InceptionStem(nn.Layer):
def __init__(self): def __init__(self):
super(InceptionStem, self).__init__() super(InceptionStem, self).__init__()
self._conv_1 = ConvBNLayer( self._conv_1 = ConvBNLayer(
...@@ -51,7 +68,7 @@ class InceptionStem(fluid.dygraph.Layer): ...@@ -51,7 +68,7 @@ class InceptionStem(fluid.dygraph.Layer):
self._conv_2 = ConvBNLayer(32, 32, 3, act="relu", name="conv2_3x3_s1") self._conv_2 = ConvBNLayer(32, 32, 3, act="relu", name="conv2_3x3_s1")
self._conv_3 = ConvBNLayer( self._conv_3 = ConvBNLayer(
32, 64, 3, padding=1, act="relu", name="conv3_3x3_s1") 32, 64, 3, padding=1, act="relu", name="conv3_3x3_s1")
self._pool = Pool2D(pool_size=3, pool_type="max", pool_stride=2) self._pool = MaxPool2d(kernel_size=3, stride=2, padding=0)
self._conv2 = ConvBNLayer( self._conv2 = ConvBNLayer(
64, 96, 3, stride=2, act="relu", name="inception_stem1_3x3_s2") 64, 96, 3, stride=2, act="relu", name="inception_stem1_3x3_s2")
self._conv1_1 = ConvBNLayer( self._conv1_1 = ConvBNLayer(
...@@ -84,7 +101,7 @@ class InceptionStem(fluid.dygraph.Layer): ...@@ -84,7 +101,7 @@ class InceptionStem(fluid.dygraph.Layer):
pool1 = self._pool(conv) pool1 = self._pool(conv)
conv2 = self._conv2(conv) conv2 = self._conv2(conv)
concat = fluid.layers.concat([pool1, conv2], axis=1) concat = paddle.concat([pool1, conv2], axis=1)
conv1 = self._conv1_1(concat) conv1 = self._conv1_1(concat)
conv1 = self._conv1_2(conv1) conv1 = self._conv1_2(conv1)
...@@ -94,19 +111,19 @@ class InceptionStem(fluid.dygraph.Layer): ...@@ -94,19 +111,19 @@ class InceptionStem(fluid.dygraph.Layer):
conv2 = self._conv2_3(conv2) conv2 = self._conv2_3(conv2)
conv2 = self._conv2_4(conv2) conv2 = self._conv2_4(conv2)
concat = fluid.layers.concat([conv1, conv2], axis=1) concat = paddle.concat([conv1, conv2], axis=1)
conv1 = self._conv3(concat) conv1 = self._conv3(concat)
pool1 = self._pool(concat) pool1 = self._pool(concat)
concat = fluid.layers.concat([conv1, pool1], axis=1) concat = paddle.concat([conv1, pool1], axis=1)
return concat return concat
class InceptionA(fluid.dygraph.Layer): class InceptionA(nn.Layer):
def __init__(self, name): def __init__(self, name):
super(InceptionA, self).__init__() super(InceptionA, self).__init__()
self._pool = Pool2D(pool_size=3, pool_type="avg", pool_padding=1) self._pool = AvgPool2d(kernel_size=3, stride=1, padding=1)
self._conv1 = ConvBNLayer( self._conv1 = ConvBNLayer(
384, 96, 1, act="relu", name="inception_a" + name + "_1x1") 384, 96, 1, act="relu", name="inception_a" + name + "_1x1")
self._conv2 = ConvBNLayer( self._conv2 = ConvBNLayer(
...@@ -154,14 +171,14 @@ class InceptionA(fluid.dygraph.Layer): ...@@ -154,14 +171,14 @@ class InceptionA(fluid.dygraph.Layer):
conv4 = self._conv4_2(conv4) conv4 = self._conv4_2(conv4)
conv4 = self._conv4_3(conv4) conv4 = self._conv4_3(conv4)
concat = fluid.layers.concat([conv1, conv2, conv3, conv4], axis=1) concat = paddle.concat([conv1, conv2, conv3, conv4], axis=1)
return concat return concat
class ReductionA(fluid.dygraph.Layer): class ReductionA(nn.Layer):
def __init__(self): def __init__(self):
super(ReductionA, self).__init__() super(ReductionA, self).__init__()
self._pool = Pool2D(pool_size=3, pool_type="max", pool_stride=2) self._pool = MaxPool2d(kernel_size=3, stride=2, padding=0)
self._conv2 = ConvBNLayer( self._conv2 = ConvBNLayer(
384, 384, 3, stride=2, act="relu", name="reduction_a_3x3") 384, 384, 3, stride=2, act="relu", name="reduction_a_3x3")
self._conv3_1 = ConvBNLayer( self._conv3_1 = ConvBNLayer(
...@@ -177,14 +194,14 @@ class ReductionA(fluid.dygraph.Layer): ...@@ -177,14 +194,14 @@ class ReductionA(fluid.dygraph.Layer):
conv3 = self._conv3_1(inputs) conv3 = self._conv3_1(inputs)
conv3 = self._conv3_2(conv3) conv3 = self._conv3_2(conv3)
conv3 = self._conv3_3(conv3) conv3 = self._conv3_3(conv3)
concat = fluid.layers.concat([pool1, conv2, conv3], axis=1) concat = paddle.concat([pool1, conv2, conv3], axis=1)
return concat return concat
class InceptionB(fluid.dygraph.Layer): class InceptionB(nn.Layer):
def __init__(self, name=None): def __init__(self, name=None):
super(InceptionB, self).__init__() super(InceptionB, self).__init__()
self._pool = Pool2D(pool_size=3, pool_type="avg", pool_padding=1) self._pool = AvgPool2d(kernel_size=3, stride=1, padding=1)
self._conv1 = ConvBNLayer( self._conv1 = ConvBNLayer(
1024, 128, 1, act="relu", name="inception_b" + name + "_1x1") 1024, 128, 1, act="relu", name="inception_b" + name + "_1x1")
self._conv2 = ConvBNLayer( self._conv2 = ConvBNLayer(
...@@ -254,14 +271,14 @@ class InceptionB(fluid.dygraph.Layer): ...@@ -254,14 +271,14 @@ class InceptionB(fluid.dygraph.Layer):
conv4 = self._conv4_4(conv4) conv4 = self._conv4_4(conv4)
conv4 = self._conv4_5(conv4) conv4 = self._conv4_5(conv4)
concat = fluid.layers.concat([conv1, conv2, conv3, conv4], axis=1) concat = paddle.concat([conv1, conv2, conv3, conv4], axis=1)
return concat return concat
class ReductionB(fluid.dygraph.Layer): class ReductionB(nn.Layer):
def __init__(self): def __init__(self):
super(ReductionB, self).__init__() super(ReductionB, self).__init__()
self._pool = Pool2D(pool_size=3, pool_type="max", pool_stride=2) self._pool = MaxPool2d(kernel_size=3, stride=2, padding=0)
self._conv2_1 = ConvBNLayer( self._conv2_1 = ConvBNLayer(
1024, 192, 1, act="relu", name="reduction_b_3x3_reduce") 1024, 192, 1, act="relu", name="reduction_b_3x3_reduce")
self._conv2_2 = ConvBNLayer( self._conv2_2 = ConvBNLayer(
...@@ -294,15 +311,15 @@ class ReductionB(fluid.dygraph.Layer): ...@@ -294,15 +311,15 @@ class ReductionB(fluid.dygraph.Layer):
conv3 = self._conv3_3(conv3) conv3 = self._conv3_3(conv3)
conv3 = self._conv3_4(conv3) conv3 = self._conv3_4(conv3)
concat = fluid.layers.concat([pool1, conv2, conv3], axis=1) concat = paddle.concat([pool1, conv2, conv3], axis=1)
return concat return concat
class InceptionC(fluid.dygraph.Layer): class InceptionC(nn.Layer):
def __init__(self, name=None): def __init__(self, name=None):
super(InceptionC, self).__init__() super(InceptionC, self).__init__()
self._pool = Pool2D(pool_size=3, pool_type="avg", pool_padding=1) self._pool = AvgPool2d(kernel_size=3, stride=1, padding=1)
self._conv1 = ConvBNLayer( self._conv1 = ConvBNLayer(
1536, 256, 1, act="relu", name="inception_c" + name + "_1x1") 1536, 256, 1, act="relu", name="inception_c" + name + "_1x1")
self._conv2 = ConvBNLayer( self._conv2 = ConvBNLayer(
...@@ -364,13 +381,13 @@ class InceptionC(fluid.dygraph.Layer): ...@@ -364,13 +381,13 @@ class InceptionC(fluid.dygraph.Layer):
conv4_1 = self._conv4_1(conv4) conv4_1 = self._conv4_1(conv4)
conv4_2 = self._conv4_2(conv4) conv4_2 = self._conv4_2(conv4)
concat = fluid.layers.concat( concat = paddle.concat(
[conv1, conv2, conv3_1, conv3_2, conv4_1, conv4_2], axis=1) [conv1, conv2, conv3_1, conv3_2, conv4_1, conv4_2], axis=1)
return concat return concat
class InceptionV4DY(fluid.dygraph.Layer): class InceptionV4DY(nn.Layer):
def __init__(self, class_dim=1000): def __init__(self, class_dim=1000):
super(InceptionV4DY, self).__init__() super(InceptionV4DY, self).__init__()
self._inception_stem = InceptionStem() self._inception_stem = InceptionStem()
...@@ -394,15 +411,14 @@ class InceptionV4DY(fluid.dygraph.Layer): ...@@ -394,15 +411,14 @@ class InceptionV4DY(fluid.dygraph.Layer):
self._inceptionC_2 = InceptionC(name="2") self._inceptionC_2 = InceptionC(name="2")
self._inceptionC_3 = InceptionC(name="3") self._inceptionC_3 = InceptionC(name="3")
self.avg_pool = Pool2D(pool_type='avg', global_pooling=True) self.avg_pool = AdaptiveAvgPool2d(1)
self._drop = Dropout(p=0.2) self._drop = Dropout(p=0.2, mode="downscale_in_infer")
stdv = 1.0 / math.sqrt(1536 * 1.0) stdv = 1.0 / math.sqrt(1536 * 1.0)
self.out = Linear( self.out = Linear(
1536, 1536,
class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv), initializer=Uniform(-stdv, stdv), name="final_fc_weights"),
name="final_fc_weights"),
bias_attr=ParamAttr(name="final_fc_offset")) bias_attr=ParamAttr(name="final_fc_offset"))
def forward(self, inputs): def forward(self, inputs):
...@@ -428,7 +444,7 @@ class InceptionV4DY(fluid.dygraph.Layer): ...@@ -428,7 +444,7 @@ class InceptionV4DY(fluid.dygraph.Layer):
x = self._inceptionC_3(x) x = self._inceptionC_3(x)
x = self.avg_pool(x) x = self.avg_pool(x)
x = fluid.layers.squeeze(x, axes=[2, 3]) x = paddle.squeeze(x, axis=[2, 3])
x = self._drop(x) x = self._drop(x)
x = self.out(x) x = self.out(x)
return x return x
...@@ -436,4 +452,4 @@ class InceptionV4DY(fluid.dygraph.Layer): ...@@ -436,4 +452,4 @@ class InceptionV4DY(fluid.dygraph.Layer):
def InceptionV4(**args): def InceptionV4(**args):
model = InceptionV4DY(**args) model = InceptionV4DY(**args)
return model return model
\ No newline at end of file
ppcls/modeling/architectures/layers.py 已删除 100644 → 0
浏览文件 @ 5e092259
#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.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import math
import warnings
import paddle.fluid as fluid
def initial_type(name,
input,
op_type,
fan_out,
init="google",
use_bias=False,
filter_size=0,
stddev=0.02):
if init == "kaiming":
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]
bound = 1 / math.sqrt(fan_in)
param_attr = fluid.ParamAttr(
name=name + "_weights",
initializer=fluid.initializer.Uniform(
low=-bound, high=bound))
if use_bias == True:
bias_attr = fluid.ParamAttr(
name=name + '_offset',
initializer=fluid.initializer.Uniform(
low=-bound, high=bound))
else:
bias_attr = False
elif init == 'google':
n = filter_size * filter_size * fan_out
param_attr = fluid.ParamAttr(
name=name + "_weights",
initializer=fluid.initializer.NormalInitializer(
loc=0.0, scale=math.sqrt(2.0 / n)))
if use_bias == True:
bias_attr = fluid.ParamAttr(
name=name + "_offset",
initializer=fluid.initializer.Constant(0.0))
else:
bias_attr = False
else:
param_attr = fluid.ParamAttr(
name=name + "_weights",
initializer=fluid.initializer.NormalInitializer(
loc=0.0, scale=stddev))
if use_bias == True:
bias_attr = fluid.ParamAttr(
name=name + "_offset",
initializer=fluid.initializer.Constant(0.0))
else:
bias_attr = False
return param_attr, bias_attr
def cal_padding(img_size, stride, filter_size, dilation=1):
"""Calculate padding size."""
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
def init_batch_norm_layer(name="batch_norm"):
param_attr = fluid.ParamAttr(
name=name + '_scale', initializer=fluid.initializer.Constant(1.0))
bias_attr = fluid.ParamAttr(
name=name + '_offset',
initializer=fluid.initializer.Constant(value=0.0))
return param_attr, bias_attr
def init_fc_layer(fout, name='fc'):
n = fout # fan-out
init_range = 1.0 / math.sqrt(n)
param_attr = fluid.ParamAttr(
name=name + '_weights',
initializer=fluid.initializer.UniformInitializer(
low=-init_range, high=init_range))
bias_attr = fluid.ParamAttr(
name=name + '_offset',
initializer=fluid.initializer.Constant(value=0.0))
return param_attr, bias_attr
def norm_layer(input, norm_type='batch_norm', name=None):
if norm_type == 'batch_norm':
param_attr = fluid.ParamAttr(
name=name + '_weights',
initializer=fluid.initializer.Constant(1.0))
bias_attr = fluid.ParamAttr(
name=name + '_offset',
initializer=fluid.initializer.Constant(value=0.0))
return fluid.layers.batch_norm(
input,
param_attr=param_attr,
bias_attr=bias_attr,
moving_mean_name=name + '_mean',
moving_variance_name=name + '_variance')
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,
initializer=fluid.initializer.Constant(1.0),
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:
raise NotImplementedError("norm tyoe: [%s] is not support" % norm_type)
def conv2d(input,
num_filters=64,
filter_size=7,
stride=1,
stddev=0.02,
padding=0,
groups=None,
name="conv2d",
norm=None,
act=None,
relufactor=0.0,
use_bias=False,
padding_type=None,
initial="normal",
use_cudnn=True):
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,
input=input,
op_type='conv',
fan_out=num_filters,
init=initial,
use_bias=use_bias,
filter_size=filter_size,
stddev=stddev)
def get_padding(filter_size, stride=1, dilation=1):
padding = ((stride - 1) + dilation * (filter_size - 1)) // 2
return padding
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]
elif padding_type == "DYNAMIC":
padding = get_padding(filter_size, stride)
else:
padding = padding
conv = fluid.layers.conv2d(
input,
num_filters,
filter_size,
groups=groups,
name=name,
stride=stride,
padding=padding,
use_cudnn=use_cudnn,
param_attr=param_attr,
bias_attr=bias_attr)
if need_crop:
conv = conv[:, :, 1:, 1:]
if norm is not None:
conv = norm_layer(input=conv, norm_type=norm, name=name + "_norm")
if act == 'relu':
conv = fluid.layers.relu(conv, name=name + '_relu')
elif act == 'leaky_relu':
conv = fluid.layers.leaky_relu(
conv, alpha=relufactor, name=name + '_leaky_relu')
elif act == 'tanh':
conv = fluid.layers.tanh(conv, name=name + '_tanh')
elif act == 'sigmoid':
conv = fluid.layers.sigmoid(conv, name=name + '_sigmoid')
elif act == 'swish':
conv = fluid.layers.swish(conv, name=name + '_swish')
elif act == None:
conv = conv
else:
raise NotImplementedError("activation: [%s] is not support" % act)
return conv
ppcls/modeling/architectures/mobilenet_v1.py
浏览文件 @ ea746480
...@@ -18,10 +18,12 @@ from __future__ import print_function ...@@ -18,10 +18,12 @@ from __future__ import print_function
import numpy as np import numpy as np
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout import paddle.nn.functional as F
from paddle.fluid.initializer import MSRA from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import MSRA
import math import math
__all__ = [ __all__ = [
...@@ -29,7 +31,7 @@ __all__ = [ ...@@ -29,7 +31,7 @@ __all__ = [
] ]
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
filter_size, filter_size,
...@@ -39,20 +41,17 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -39,20 +41,17 @@ class ConvBNLayer(fluid.dygraph.Layer):
channels=None, channels=None,
num_groups=1, num_groups=1,
act='relu', act='relu',
use_cudnn=True,
name=None): name=None):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=padding, padding=padding,
groups=num_groups, groups=num_groups,
act=None, weight_attr=ParamAttr(
use_cudnn=use_cudnn,
param_attr=ParamAttr(
initializer=MSRA(), name=name + "_weights"), initializer=MSRA(), name=name + "_weights"),
bias_attr=False) bias_attr=False)
...@@ -70,7 +69,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -70,7 +69,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return y return y
class DepthwiseSeparable(fluid.dygraph.Layer): class DepthwiseSeparable(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters1, num_filters1,
...@@ -88,7 +87,6 @@ class DepthwiseSeparable(fluid.dygraph.Layer): ...@@ -88,7 +87,6 @@ class DepthwiseSeparable(fluid.dygraph.Layer):
stride=stride, stride=stride,
padding=1, padding=1,
num_groups=int(num_groups * scale), num_groups=int(num_groups * scale),
use_cudnn=False,
name=name + "_dw") name=name + "_dw")
self._pointwise_conv = ConvBNLayer( self._pointwise_conv = ConvBNLayer(
...@@ -105,7 +103,7 @@ class DepthwiseSeparable(fluid.dygraph.Layer): ...@@ -105,7 +103,7 @@ class DepthwiseSeparable(fluid.dygraph.Layer):
return y return y
class MobileNet(fluid.dygraph.Layer): class MobileNet(nn.Layer):
def __init__(self, scale=1.0, class_dim=1000): def __init__(self, scale=1.0, class_dim=1000):
super(MobileNet, self).__init__() super(MobileNet, self).__init__()
self.scale = scale self.scale = scale
...@@ -229,12 +227,12 @@ class MobileNet(fluid.dygraph.Layer): ...@@ -229,12 +227,12 @@ class MobileNet(fluid.dygraph.Layer):
name="conv6")) name="conv6"))
self.block_list.append(conv6) self.block_list.append(conv6)
self.pool2d_avg = Pool2D(pool_type='avg', global_pooling=True) self.pool2d_avg = AdaptiveAvgPool2d(1)
self.out = Linear( self.out = Linear(
int(1024 * scale), int(1024 * scale),
class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
initializer=MSRA(), name="fc7_weights"), initializer=MSRA(), name="fc7_weights"),
bias_attr=ParamAttr(name="fc7_offset")) bias_attr=ParamAttr(name="fc7_offset"))
...@@ -243,7 +241,7 @@ class MobileNet(fluid.dygraph.Layer): ...@@ -243,7 +241,7 @@ class MobileNet(fluid.dygraph.Layer):
for block in self.block_list: for block in self.block_list:
y = block(y) y = block(y)
y = self.pool2d_avg(y) y = self.pool2d_avg(y)
y = fluid.layers.reshape(y, shape=[-1, int(1024 * self.scale)]) y = paddle.reshape(y, shape=[-1, int(1024 * self.scale)])
y = self.out(y) y = self.out(y)
return y return y
......
ppcls/modeling/architectures/mobilenet_v2.py
浏览文件 @ ea746480
...@@ -18,9 +18,11 @@ from __future__ import print_function ...@@ -18,9 +18,11 @@ from __future__ import print_function
import numpy as np import numpy as np
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
import math import math
...@@ -30,7 +32,7 @@ __all__ = [ ...@@ -30,7 +32,7 @@ __all__ = [
] ]
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
filter_size, filter_size,
...@@ -43,16 +45,14 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -43,16 +45,14 @@ class ConvBNLayer(fluid.dygraph.Layer):
use_cudnn=True): use_cudnn=True):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=padding, padding=padding,
groups=num_groups, groups=num_groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
use_cudnn=use_cudnn,
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
self._batch_norm = BatchNorm( self._batch_norm = BatchNorm(
...@@ -66,11 +66,11 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -66,11 +66,11 @@ class ConvBNLayer(fluid.dygraph.Layer):
y = self._conv(inputs) y = self._conv(inputs)
y = self._batch_norm(y) y = self._batch_norm(y)
if if_act: if if_act:
y = fluid.layers.relu6(y) y = F.relu6(y)
return y return y
class InvertedResidualUnit(fluid.dygraph.Layer): class InvertedResidualUnit(nn.Layer):
def __init__(self, num_channels, num_in_filter, num_filters, stride, def __init__(self, num_channels, num_in_filter, num_filters, stride,
filter_size, padding, expansion_factor, name): filter_size, padding, expansion_factor, name):
super(InvertedResidualUnit, self).__init__() super(InvertedResidualUnit, self).__init__()
...@@ -108,11 +108,11 @@ class InvertedResidualUnit(fluid.dygraph.Layer): ...@@ -108,11 +108,11 @@ class InvertedResidualUnit(fluid.dygraph.Layer):
y = self._bottleneck_conv(y, if_act=True) y = self._bottleneck_conv(y, if_act=True)
y = self._linear_conv(y, if_act=False) y = self._linear_conv(y, if_act=False)
if ifshortcut: if ifshortcut:
y = fluid.layers.elementwise_add(inputs, y) y = paddle.elementwise_add(inputs, y)
return y return y
class InvresiBlocks(fluid.dygraph.Layer): class InvresiBlocks(nn.Layer):
def __init__(self, in_c, t, c, n, s, name): def __init__(self, in_c, t, c, n, s, name):
super(InvresiBlocks, self).__init__() super(InvresiBlocks, self).__init__()
...@@ -148,7 +148,7 @@ class InvresiBlocks(fluid.dygraph.Layer): ...@@ -148,7 +148,7 @@ class InvresiBlocks(fluid.dygraph.Layer):
return y return y
class MobileNet(fluid.dygraph.Layer): class MobileNet(nn.Layer):
def __init__(self, class_dim=1000, scale=1.0): def __init__(self, class_dim=1000, scale=1.0):
super(MobileNet, self).__init__() super(MobileNet, self).__init__()
self.scale = scale self.scale = scale
...@@ -199,12 +199,12 @@ class MobileNet(fluid.dygraph.Layer): ...@@ -199,12 +199,12 @@ class MobileNet(fluid.dygraph.Layer):
padding=0, padding=0,
name="conv9") name="conv9")
self.pool2d_avg = Pool2D(pool_type="avg", global_pooling=True) self.pool2d_avg = AdaptiveAvgPool2d(1)
self.out = Linear( self.out = Linear(
self.out_c, self.out_c,
class_dim, class_dim,
param_attr=ParamAttr(name="fc10_weights"), weight_attr=ParamAttr(name="fc10_weights"),
bias_attr=ParamAttr(name="fc10_offset")) bias_attr=ParamAttr(name="fc10_offset"))
def forward(self, inputs): def forward(self, inputs):
...@@ -213,7 +213,7 @@ class MobileNet(fluid.dygraph.Layer): ...@@ -213,7 +213,7 @@ class MobileNet(fluid.dygraph.Layer):
y = block(y) y = block(y)
y = self.conv9(y, if_act=True) y = self.conv9(y, if_act=True)
y = self.pool2d_avg(y) y = self.pool2d_avg(y)
y = fluid.layers.reshape(y, shape=[-1, self.out_c]) y = paddle.reshape(y, shape=[-1, self.out_c])
y = self.out(y) y = self.out(y)
return y return y
......
ppcls/modeling/architectures/mobilenet_v3.py
浏览文件 @ ea746480
...@@ -18,9 +18,12 @@ from __future__ import print_function ...@@ -18,9 +18,12 @@ from __future__ import print_function
import numpy as np import numpy as np
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.regularizer import L2Decay
import math import math
...@@ -42,8 +45,12 @@ def make_divisible(v, divisor=8, min_value=None): ...@@ -42,8 +45,12 @@ def make_divisible(v, divisor=8, min_value=None):
return new_v return new_v
class MobileNetV3(fluid.dygraph.Layer): class MobileNetV3(nn.Layer):
def __init__(self, scale=1.0, model_name="small", class_dim=1000): def __init__(self,
scale=1.0,
model_name="small",
dropout_prob=0.2,
class_dim=1000):
super(MobileNetV3, self).__init__() super(MobileNetV3, self).__init__()
inplanes = 16 inplanes = 16
...@@ -130,41 +137,42 @@ class MobileNetV3(fluid.dygraph.Layer): ...@@ -130,41 +137,42 @@ class MobileNetV3(fluid.dygraph.Layer):
act="hard_swish", act="hard_swish",
name="conv_last") name="conv_last")
self.pool = Pool2D( self.pool = AdaptiveAvgPool2d(1)
pool_type="avg", global_pooling=True, use_cudnn=False)
self.last_conv = Conv2D( self.last_conv = Conv2d(
num_channels=make_divisible(scale * self.cls_ch_squeeze), in_channels=make_divisible(scale * self.cls_ch_squeeze),
num_filters=self.cls_ch_expand, out_channels=self.cls_ch_expand,
filter_size=1, kernel_size=1,
stride=1, stride=1,
padding=0, padding=0,
act=None, weight_attr=ParamAttr(name="last_1x1_conv_weights"),
param_attr=ParamAttr(name="last_1x1_conv_weights"),
bias_attr=False) bias_attr=False)
self.dropout = Dropout(p=dropout_prob, mode="downscale_in_infer")
self.out = Linear( self.out = Linear(
input_dim=self.cls_ch_expand, self.cls_ch_expand,
output_dim=class_dim, class_dim,
param_attr=ParamAttr("fc_weights"), weight_attr=ParamAttr("fc_weights"),
bias_attr=ParamAttr(name="fc_offset")) bias_attr=ParamAttr(name="fc_offset"))
def forward(self, inputs, label=None, dropout_prob=0.2): def forward(self, inputs, label=None):
x = self.conv1(inputs) x = self.conv1(inputs)
for block in self.block_list: for block in self.block_list:
x = block(x) x = block(x)
x = self.last_second_conv(x) x = self.last_second_conv(x)
x = self.pool(x) x = self.pool(x)
x = self.last_conv(x) x = self.last_conv(x)
x = fluid.layers.hard_swish(x) x = F.hard_swish(x)
x = fluid.layers.dropout(x=x, dropout_prob=dropout_prob) x = self.dropout(x)
x = fluid.layers.reshape(x, shape=[x.shape[0], x.shape[1]]) x = paddle.reshape(x, shape=[x.shape[0], x.shape[1]])
x = self.out(x) x = self.out(x)
return x return x
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__(self, def __init__(self,
in_c, in_c,
out_c, out_c,
...@@ -179,28 +187,22 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -179,28 +187,22 @@ class ConvBNLayer(fluid.dygraph.Layer):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self.if_act = if_act self.if_act = if_act
self.act = act self.act = act
self.conv = fluid.dygraph.Conv2D( self.conv = Conv2d(
num_channels=in_c, in_channels=in_c,
num_filters=out_c, out_channels=out_c,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=padding, padding=padding,
groups=num_groups, groups=num_groups,
param_attr=ParamAttr(name=name + "_weights"), weight_attr=ParamAttr(name=name + "_weights"),
bias_attr=False, bias_attr=False)
use_cudnn=use_cudnn, self.bn = BatchNorm(
act=None)
self.bn = fluid.dygraph.BatchNorm(
num_channels=out_c, num_channels=out_c,
act=None, act=None,
param_attr=ParamAttr( param_attr=ParamAttr(
name=name + "_bn_scale", name=name + "_bn_scale", regularizer=L2Decay(0.0)),
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0)),
bias_attr=ParamAttr( bias_attr=ParamAttr(
name=name + "_bn_offset", name=name + "_bn_offset", regularizer=L2Decay(0.0)),
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0)),
moving_mean_name=name + "_bn_mean", moving_mean_name=name + "_bn_mean",
moving_variance_name=name + "_bn_variance") moving_variance_name=name + "_bn_variance")
...@@ -209,16 +211,16 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -209,16 +211,16 @@ class ConvBNLayer(fluid.dygraph.Layer):
x = self.bn(x) x = self.bn(x)
if self.if_act: if self.if_act:
if self.act == "relu": if self.act == "relu":
x = fluid.layers.relu(x) x = F.relu(x)
elif self.act == "hard_swish": elif self.act == "hard_swish":
x = fluid.layers.hard_swish(x) x = F.hard_swish(x)
else: else:
print("The activation function is selected incorrectly.") print("The activation function is selected incorrectly.")
exit() exit()
return x return x
class ResidualUnit(fluid.dygraph.Layer): class ResidualUnit(nn.Layer):
def __init__(self, def __init__(self,
in_c, in_c,
mid_c, mid_c,
...@@ -270,40 +272,38 @@ class ResidualUnit(fluid.dygraph.Layer): ...@@ -270,40 +272,38 @@ class ResidualUnit(fluid.dygraph.Layer):
x = self.mid_se(x) x = self.mid_se(x)
x = self.linear_conv(x) x = self.linear_conv(x)
if self.if_shortcut: if self.if_shortcut:
x = fluid.layers.elementwise_add(inputs, x) x = paddle.elementwise_add(inputs, x)
return x return x
class SEModule(fluid.dygraph.Layer): class SEModule(nn.Layer):
def __init__(self, channel, reduction=4, name=""): def __init__(self, channel, reduction=4, name=""):
super(SEModule, self).__init__() super(SEModule, self).__init__()
self.avg_pool = fluid.dygraph.Pool2D( self.avg_pool = AdaptiveAvgPool2d(1)
pool_type="avg", global_pooling=True, use_cudnn=False) self.conv1 = Conv2d(
self.conv1 = fluid.dygraph.Conv2D( in_channels=channel,
num_channels=channel, out_channels=channel // reduction,
num_filters=channel // reduction, kernel_size=1,
filter_size=1,
stride=1, stride=1,
padding=0, padding=0,
act="relu", weight_attr=ParamAttr(name=name + "_1_weights"),
param_attr=ParamAttr(name=name + "_1_weights"),
bias_attr=ParamAttr(name=name + "_1_offset")) bias_attr=ParamAttr(name=name + "_1_offset"))
self.conv2 = fluid.dygraph.Conv2D( self.conv2 = Conv2d(
num_channels=channel // reduction, in_channels=channel // reduction,
num_filters=channel, out_channels=channel,
filter_size=1, kernel_size=1,
stride=1, stride=1,
padding=0, padding=0,
act=None, weight_attr=ParamAttr(name + "_2_weights"),
param_attr=ParamAttr(name + "_2_weights"),
bias_attr=ParamAttr(name=name + "_2_offset")) bias_attr=ParamAttr(name=name + "_2_offset"))
def forward(self, inputs): def forward(self, inputs):
outputs = self.avg_pool(inputs) outputs = self.avg_pool(inputs)
outputs = self.conv1(outputs) outputs = self.conv1(outputs)
outputs = F.relu(outputs)
outputs = self.conv2(outputs) outputs = self.conv2(outputs)
outputs = fluid.layers.hard_sigmoid(outputs) outputs = F.hard_sigmoid(outputs)
return fluid.layers.elementwise_mul(x=inputs, y=outputs, axis=0) return paddle.multiply(x=inputs, y=outputs, axis=0)
def MobileNetV3_small_x0_35(**args): def MobileNetV3_small_x0_35(**args):
......
ppcls/modeling/architectures/model_libs.py 已删除 100644 → 0
浏览文件 @ 5e092259
#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.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import paddle
import paddle.fluid as fluid
import contextlib
bn_regularizer = fluid.regularizer.L2DecayRegularizer(regularization_coeff=0.0)
name_scope = ""
@contextlib.contextmanager
def scope(name):
global name_scope
bk = name_scope
name_scope = name_scope + name + '/'
yield
name_scope = bk
def max_pool(input, kernel, stride, padding):
data = fluid.layers.pool2d(
input,
pool_size=kernel,
pool_type='max',
pool_stride=stride,
pool_padding=padding)
return data
def group_norm(input, G, eps=1e-5, param_attr=None, bias_attr=None):
N, C, H, W = input.shape
if C % G != 0:
# print "group can not divide channle:", C, G
for d in range(10):
for t in [d, -d]:
if G + t <= 0: continue
if C % (G + t) == 0:
G = G + t
break
if C % G == 0:
# print "use group size:", G
break
assert C % G == 0
x = fluid.layers.group_norm(
input,
groups=G,
param_attr=param_attr,
bias_attr=bias_attr,
name=name_scope + 'group_norm')
return x
def bn(*args, **kargs):
with scope('BatchNorm'):
return fluid.layers.batch_norm(
*args,
epsilon=1e-3,
momentum=0.99,
param_attr=fluid.ParamAttr(
name=name_scope + 'gamma', regularizer=bn_regularizer),
bias_attr=fluid.ParamAttr(
name=name_scope + 'beta', regularizer=bn_regularizer),
moving_mean_name=name_scope + 'moving_mean',
moving_variance_name=name_scope + 'moving_variance',
**kargs)
def bn_relu(data):
return fluid.layers.relu(bn(data))
def relu(data):
return fluid.layers.relu(data)
def conv(*args, **kargs):
kargs['param_attr'] = name_scope + 'weights'
if 'bias_attr' in kargs and kargs['bias_attr']:
kargs['bias_attr'] = fluid.ParamAttr(
name=name_scope + 'biases',
regularizer=None,
initializer=fluid.initializer.ConstantInitializer(value=0.0))
else:
kargs['bias_attr'] = False
return fluid.layers.conv2d(*args, **kargs)
def deconv(*args, **kargs):
kargs['param_attr'] = name_scope + 'weights'
if 'bias_attr' in kargs and kargs['bias_attr']:
kargs['bias_attr'] = name_scope + 'biases'
else:
kargs['bias_attr'] = False
return fluid.layers.conv2d_transpose(*args, **kargs)
def seperate_conv(input, channel, stride, filter, dilation=1, act=None):
param_attr = fluid.ParamAttr(
name=name_scope + 'weights',
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0),
initializer=fluid.initializer.TruncatedNormal(
loc=0.0, scale=0.33))
with scope('depthwise'):
input = conv(
input,
input.shape[1],
filter,
stride,
groups=input.shape[1],
padding=(filter // 2) * dilation,
dilation=dilation,
use_cudnn=False,
param_attr=param_attr)
input = bn(input)
if act: input = act(input)
param_attr = fluid.ParamAttr(
name=name_scope + 'weights',
regularizer=None,
initializer=fluid.initializer.TruncatedNormal(
loc=0.0, scale=0.06))
with scope('pointwise'):
input = conv(
input, channel, 1, 1, groups=1, padding=0, param_attr=param_attr)
input = bn(input)
if act: input = act(input)
return input
ppcls/modeling/architectures/res2net.py
浏览文件 @ ea746480
...@@ -18,9 +18,12 @@ from __future__ import print_function ...@@ -18,9 +18,12 @@ from __future__ import print_function
import numpy as np import numpy as np
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
...@@ -31,7 +34,7 @@ __all__ = [ ...@@ -31,7 +34,7 @@ __all__ = [
] ]
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__( def __init__(
self, self,
num_channels, num_channels,
...@@ -43,15 +46,14 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -43,15 +46,14 @@ class ConvBNLayer(fluid.dygraph.Layer):
name=None, ): name=None, ):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
if name == "conv1": if name == "conv1":
bn_name = "bn_" + name bn_name = "bn_" + name
...@@ -71,7 +73,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -71,7 +73,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return y return y
class BottleneckBlock(fluid.dygraph.Layer): class BottleneckBlock(nn.Layer):
def __init__(self, def __init__(self,
num_channels1, num_channels1,
num_channels2, num_channels2,
...@@ -102,8 +104,7 @@ class BottleneckBlock(fluid.dygraph.Layer): ...@@ -102,8 +104,7 @@ class BottleneckBlock(fluid.dygraph.Layer):
act='relu', act='relu',
name=name + '_branch2b_' + str(s + 1))) name=name + '_branch2b_' + str(s + 1)))
self.conv1_list.append(conv1) self.conv1_list.append(conv1)
self.pool2d_avg = Pool2D( self.pool2d_avg = AvgPool2d(kernel_size=3, stride=stride, padding=1)
pool_size=3, pool_stride=stride, pool_padding=1, pool_type='avg')
self.conv2 = ConvBNLayer( self.conv2 = ConvBNLayer(
num_channels=num_filters, num_channels=num_filters,
...@@ -124,7 +125,7 @@ class BottleneckBlock(fluid.dygraph.Layer): ...@@ -124,7 +125,7 @@ class BottleneckBlock(fluid.dygraph.Layer):
def forward(self, inputs): def forward(self, inputs):
y = self.conv0(inputs) y = self.conv0(inputs)
xs = fluid.layers.split(y, self.scales, 1) xs = paddle.split(y, self.scales, 1)
ys = [] ys = []
for s, conv1 in enumerate(self.conv1_list): for s, conv1 in enumerate(self.conv1_list):
if s == 0 or self.stride == 2: if s == 0 or self.stride == 2:
...@@ -135,18 +136,18 @@ class BottleneckBlock(fluid.dygraph.Layer): ...@@ -135,18 +136,18 @@ class BottleneckBlock(fluid.dygraph.Layer):
ys.append(xs[-1]) ys.append(xs[-1])
else: else:
ys.append(self.pool2d_avg(xs[-1])) ys.append(self.pool2d_avg(xs[-1]))
conv1 = fluid.layers.concat(ys, axis=1) conv1 = paddle.concat(ys, axis=1)
conv2 = self.conv2(conv1) conv2 = self.conv2(conv1)
if self.shortcut: if self.shortcut:
short = inputs short = inputs
else: else:
short = self.short(inputs) short = self.short(inputs)
y = fluid.layers.elementwise_add(x=short, y=conv2, act='relu') y = paddle.elementwise_add(x=short, y=conv2, act='relu')
return y return y
class Res2Net(fluid.dygraph.Layer): class Res2Net(nn.Layer):
def __init__(self, layers=50, scales=4, width=26, class_dim=1000): def __init__(self, layers=50, scales=4, width=26, class_dim=1000):
super(Res2Net, self).__init__() super(Res2Net, self).__init__()
...@@ -178,8 +179,7 @@ class Res2Net(fluid.dygraph.Layer): ...@@ -178,8 +179,7 @@ class Res2Net(fluid.dygraph.Layer):
stride=2, stride=2,
act='relu', act='relu',
name="conv1") name="conv1")
self.pool2d_max = Pool2D( self.pool2d_max = MaxPool2d(kernel_size=3, stride=2, padding=1)
pool_size=3, pool_stride=2, pool_padding=1, pool_type='max')
self.block_list = [] self.block_list = []
for block in range(len(depth)): for block in range(len(depth)):
...@@ -207,8 +207,7 @@ class Res2Net(fluid.dygraph.Layer): ...@@ -207,8 +207,7 @@ class Res2Net(fluid.dygraph.Layer):
self.block_list.append(bottleneck_block) self.block_list.append(bottleneck_block)
shortcut = True shortcut = True
self.pool2d_avg = Pool2D( self.pool2d_avg = AdaptiveAvgPool2d(1)
pool_size=7, pool_type='avg', global_pooling=True)
self.pool2d_avg_channels = num_channels[-1] * 2 self.pool2d_avg_channels = num_channels[-1] * 2
...@@ -217,9 +216,8 @@ class Res2Net(fluid.dygraph.Layer): ...@@ -217,9 +216,8 @@ class Res2Net(fluid.dygraph.Layer):
self.out = Linear( self.out = Linear(
self.pool2d_avg_channels, self.pool2d_avg_channels,
class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv), initializer=Uniform(-stdv, stdv), name="fc_weights"),
name="fc_weights"),
bias_attr=ParamAttr(name="fc_offset")) bias_attr=ParamAttr(name="fc_offset"))
def forward(self, inputs): def forward(self, inputs):
...@@ -228,7 +226,7 @@ class Res2Net(fluid.dygraph.Layer): ...@@ -228,7 +226,7 @@ class Res2Net(fluid.dygraph.Layer):
for block in self.block_list: for block in self.block_list:
y = block(y) y = block(y)
y = self.pool2d_avg(y) y = self.pool2d_avg(y)
y = fluid.layers.reshape(y, shape=[-1, self.pool2d_avg_channels]) y = paddle.reshape(y, shape=[-1, self.pool2d_avg_channels])
y = self.out(y) y = self.out(y)
return y return y
......
ppcls/modeling/architectures/res2net_vd.py
浏览文件 @ ea746480
...@@ -18,9 +18,12 @@ from __future__ import print_function ...@@ -18,9 +18,12 @@ from __future__ import print_function
import numpy as np import numpy as np
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
...@@ -31,7 +34,7 @@ __all__ = [ ...@@ -31,7 +34,7 @@ __all__ = [
] ]
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__( def __init__(
self, self,
num_channels, num_channels,
...@@ -45,21 +48,17 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -45,21 +48,17 @@ class ConvBNLayer(fluid.dygraph.Layer):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self.is_vd_mode = is_vd_mode self.is_vd_mode = is_vd_mode
self._pool2d_avg = Pool2D( self._pool2d_avg = AvgPool2d(
pool_size=2, kernel_size=2, stride=2, padding=0, ceil_mode=True)
pool_stride=2, self._conv = Conv2d(
pool_padding=0, in_channels=num_channels,
pool_type='avg', out_channels=num_filters,
ceil_mode=True) kernel_size=filter_size,
self._conv = Conv2D(
num_channels=num_channels,
num_filters=num_filters,
filter_size=filter_size,
stride=stride, stride=stride,
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=groups, groups=groups,
act=None, act=None,
param_attr=ParamAttr(name=name + "_weights"), weight_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
if name == "conv1": if name == "conv1":
bn_name = "bn_" + name bn_name = "bn_" + name
...@@ -81,7 +80,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -81,7 +80,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return y return y
class BottleneckBlock(fluid.dygraph.Layer): class BottleneckBlock(nn.Layer):
def __init__(self, def __init__(self,
num_channels1, num_channels1,
num_channels2, num_channels2,
...@@ -112,8 +111,8 @@ class BottleneckBlock(fluid.dygraph.Layer): ...@@ -112,8 +111,8 @@ class BottleneckBlock(fluid.dygraph.Layer):
act='relu', act='relu',
name=name + '_branch2b_' + str(s + 1))) name=name + '_branch2b_' + str(s + 1)))
self.conv1_list.append(conv1) self.conv1_list.append(conv1)
self.pool2d_avg = Pool2D( self.pool2d_avg = AvgPool2d(
pool_size=3, pool_stride=stride, pool_padding=1, pool_type='avg') kernel_size=3, stride=stride, padding=1, ceil_mode=True)
self.conv2 = ConvBNLayer( self.conv2 = ConvBNLayer(
num_channels=num_filters, num_channels=num_filters,
...@@ -135,7 +134,7 @@ class BottleneckBlock(fluid.dygraph.Layer): ...@@ -135,7 +134,7 @@ class BottleneckBlock(fluid.dygraph.Layer):
def forward(self, inputs): def forward(self, inputs):
y = self.conv0(inputs) y = self.conv0(inputs)
xs = fluid.layers.split(y, self.scales, 1) xs = paddle.split(y, self.scales, 1)
ys = [] ys = []
for s, conv1 in enumerate(self.conv1_list): for s, conv1 in enumerate(self.conv1_list):
if s == 0 or self.stride == 2: if s == 0 or self.stride == 2:
...@@ -146,18 +145,18 @@ class BottleneckBlock(fluid.dygraph.Layer): ...@@ -146,18 +145,18 @@ class BottleneckBlock(fluid.dygraph.Layer):
ys.append(xs[-1]) ys.append(xs[-1])
else: else:
ys.append(self.pool2d_avg(xs[-1])) ys.append(self.pool2d_avg(xs[-1]))
conv1 = fluid.layers.concat(ys, axis=1) conv1 = paddle.concat(ys, axis=1)
conv2 = self.conv2(conv1) conv2 = self.conv2(conv1)
if self.shortcut: if self.shortcut:
short = inputs short = inputs
else: else:
short = self.short(inputs) short = self.short(inputs)
y = fluid.layers.elementwise_add(x=short, y=conv2, act='relu') y = paddle.elementwise_add(x=short, y=conv2, act='relu')
return y return y
class Res2Net_vd(fluid.dygraph.Layer): class Res2Net_vd(nn.Layer):
def __init__(self, layers=50, scales=4, width=26, class_dim=1000): def __init__(self, layers=50, scales=4, width=26, class_dim=1000):
super(Res2Net_vd, self).__init__() super(Res2Net_vd, self).__init__()
...@@ -203,8 +202,7 @@ class Res2Net_vd(fluid.dygraph.Layer): ...@@ -203,8 +202,7 @@ class Res2Net_vd(fluid.dygraph.Layer):
stride=1, stride=1,
act='relu', act='relu',
name="conv1_3") name="conv1_3")
self.pool2d_max = Pool2D( self.pool2d_max = MaxPool2d(kernel_size=3, stride=2, padding=1)
pool_size=3, pool_stride=2, pool_padding=1, pool_type='max')
self.block_list = [] self.block_list = []
for block in range(len(depth)): for block in range(len(depth)):
...@@ -232,8 +230,7 @@ class Res2Net_vd(fluid.dygraph.Layer): ...@@ -232,8 +230,7 @@ class Res2Net_vd(fluid.dygraph.Layer):
self.block_list.append(bottleneck_block) self.block_list.append(bottleneck_block)
shortcut = True shortcut = True
self.pool2d_avg = Pool2D( self.pool2d_avg = AdaptiveAvgPool2d(1)
pool_size=7, pool_type='avg', global_pooling=True)
self.pool2d_avg_channels = num_channels[-1] * 2 self.pool2d_avg_channels = num_channels[-1] * 2
...@@ -242,9 +239,8 @@ class Res2Net_vd(fluid.dygraph.Layer): ...@@ -242,9 +239,8 @@ class Res2Net_vd(fluid.dygraph.Layer):
self.out = Linear( self.out = Linear(
self.pool2d_avg_channels, self.pool2d_avg_channels,
class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv), initializer=Uniform(-stdv, stdv), name="fc_weights"),
name="fc_weights"),
bias_attr=ParamAttr(name="fc_offset")) bias_attr=ParamAttr(name="fc_offset"))
def forward(self, inputs): def forward(self, inputs):
...@@ -255,7 +251,7 @@ class Res2Net_vd(fluid.dygraph.Layer): ...@@ -255,7 +251,7 @@ class Res2Net_vd(fluid.dygraph.Layer):
for block in self.block_list: for block in self.block_list:
y = block(y) y = block(y)
y = self.pool2d_avg(y) y = self.pool2d_avg(y)
y = fluid.layers.reshape(y, shape=[-1, self.pool2d_avg_channels]) y = paddle.reshape(y, shape=[-1, self.pool2d_avg_channels])
y = self.out(y) y = self.out(y)
return y return y
......
ppcls/modeling/architectures/resnest.py
浏览文件 @ ea746480
...@@ -20,11 +20,11 @@ import numpy as np ...@@ -20,11 +20,11 @@ import numpy as np
import paddle import paddle
import math import math
import paddle.nn as nn import paddle.nn as nn
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr from paddle.nn.initializer import MSRA
from paddle.fluid.regularizer import L2DecayRegularizer from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.fluid.initializer import MSRA, ConstantInitializer from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout from paddle.regularizer import L2Decay
__all__ = ["ResNeSt50_fast_1s1x64d", "ResNeSt50"] __all__ = ["ResNeSt50_fast_1s1x64d", "ResNeSt50"]
...@@ -43,26 +43,23 @@ class ConvBNLayer(nn.Layer): ...@@ -43,26 +43,23 @@ class ConvBNLayer(nn.Layer):
bn_decay = 0.0 bn_decay = 0.0
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
dilation=dilation, dilation=dilation,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weight"),
param_attr=ParamAttr(name=name + "_weight"),
bias_attr=False) bias_attr=False)
self._batch_norm = BatchNorm( self._batch_norm = BatchNorm(
num_filters, num_filters,
act=act, act=act,
param_attr=ParamAttr( param_attr=ParamAttr(
name=name + "_scale", name=name + "_scale", regularizer=L2Decay(bn_decay)),
regularizer=L2DecayRegularizer(regularization_coeff=bn_decay)),
bias_attr=ParamAttr( bias_attr=ParamAttr(
name + "_offset", name + "_offset", regularizer=L2Decay(bn_decay)),
regularizer=L2DecayRegularizer(regularization_coeff=bn_decay)),
moving_mean_name=name + "_mean", moving_mean_name=name + "_mean",
moving_variance_name=name + "_variance") moving_variance_name=name + "_variance")
...@@ -124,7 +121,7 @@ class SplatConv(nn.Layer): ...@@ -124,7 +121,7 @@ class SplatConv(nn.Layer):
act="relu", act="relu",
name=name + "_splat1") name=name + "_splat1")
self.avg_pool2d = Pool2D(pool_type='avg', global_pooling=True) self.avg_pool2d = AdaptiveAvgPool2d(1)
inter_channels = int(max(in_channels * radix // reduction_factor, 32)) inter_channels = int(max(in_channels * radix // reduction_factor, 32))
...@@ -139,15 +136,14 @@ class SplatConv(nn.Layer): ...@@ -139,15 +136,14 @@ class SplatConv(nn.Layer):
name=name + "_splat2") name=name + "_splat2")
# to calc atten # to calc atten
self.conv3 = Conv2D( self.conv3 = Conv2d(
num_channels=inter_channels, in_channels=inter_channels,
num_filters=channels * radix, out_channels=channels * radix,
filter_size=1, kernel_size=1,
stride=1, stride=1,
padding=0, padding=0,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(
param_attr=ParamAttr(
name=name + "_splat_weights", initializer=MSRA()), name=name + "_splat_weights", initializer=MSRA()),
bias_attr=False) bias_attr=False)
...@@ -221,11 +217,8 @@ class BottleneckBlock(nn.Layer): ...@@ -221,11 +217,8 @@ class BottleneckBlock(nn.Layer):
name=name + "_conv1") name=name + "_conv1")
if avd and avd_first and (stride > 1 or is_first): if avd and avd_first and (stride > 1 or is_first):
self.avg_pool2d_1 = Pool2D( self.avg_pool2d_1 = AvgPool2d(
pool_size=3, kernel_size=3, stride=stride, padding=1)
pool_stride=stride,
pool_padding=1,
pool_type="avg")
if radix >= 1: if radix >= 1:
self.conv2 = SplatConv( self.conv2 = SplatConv(
...@@ -252,11 +245,8 @@ class BottleneckBlock(nn.Layer): ...@@ -252,11 +245,8 @@ class BottleneckBlock(nn.Layer):
name=name + "_conv2") name=name + "_conv2")
if avd and avd_first == False and (stride > 1 or is_first): if avd and avd_first == False and (stride > 1 or is_first):
self.avg_pool2d_2 = Pool2D( self.avg_pool2d_2 = AvgPool2d(
pool_size=3, kernel_size=3, stride=stride, padding=1)
pool_stride=stride,
pool_padding=1,
pool_type="avg")
self.conv3 = ConvBNLayer( self.conv3 = ConvBNLayer(
num_channels=group_width, num_channels=group_width,
...@@ -270,39 +260,31 @@ class BottleneckBlock(nn.Layer): ...@@ -270,39 +260,31 @@ class BottleneckBlock(nn.Layer):
if stride != 1 or self.inplanes != self.planes * 4: if stride != 1 or self.inplanes != self.planes * 4:
if avg_down: if avg_down:
if dilation == 1: if dilation == 1:
self.avg_pool2d_3 = Pool2D( self.avg_pool2d_3 = AvgPool2d(
pool_size=stride, kernel_size=stride, stride=stride, padding=0)
pool_stride=stride,
pool_type="avg",
ceil_mode=True)
else: else:
self.avg_pool2d_3 = Pool2D( self.avg_pool2d_3 = AvgPool2d(
pool_size=1, kernel_size=1, stride=1, padding=0, ceil_mode=True)
pool_stride=1,
pool_type="avg", self.conv4 = Conv2d(
ceil_mode=True) in_channels=self.inplanes,
out_channels=planes * 4,
self.conv4 = Conv2D( kernel_size=1,
num_channels=self.inplanes,
num_filters=planes * 4,
filter_size=1,
stride=1, stride=1,
padding=0, padding=0,
groups=1, groups=1,
act=None, weight_attr=ParamAttr(
param_attr=ParamAttr(
name=name + "_weights", initializer=MSRA()), name=name + "_weights", initializer=MSRA()),
bias_attr=False) bias_attr=False)
else: else:
self.conv4 = Conv2D( self.conv4 = Conv2d(
num_channels=self.inplanes, in_channels=self.inplanes,
num_filters=planes * 4, out_channels=planes * 4,
filter_size=1, kernel_size=1,
stride=stride, stride=stride,
padding=0, padding=0,
groups=1, groups=1,
act=None, weight_attr=ParamAttr(
param_attr=ParamAttr(
name=name + "_shortcut_weights", initializer=MSRA()), name=name + "_shortcut_weights", initializer=MSRA()),
bias_attr=False) bias_attr=False)
...@@ -312,12 +294,10 @@ class BottleneckBlock(nn.Layer): ...@@ -312,12 +294,10 @@ class BottleneckBlock(nn.Layer):
act=None, act=None,
param_attr=ParamAttr( param_attr=ParamAttr(
name=name + "_shortcut_scale", name=name + "_shortcut_scale",
regularizer=L2DecayRegularizer( regularizer=L2Decay(regularization_coeff=bn_decay)),
regularization_coeff=bn_decay)),
bias_attr=ParamAttr( bias_attr=ParamAttr(
name + "_shortcut_offset", name + "_shortcut_offset",
regularizer=L2DecayRegularizer( regularizer=L2Decay(regularization_coeff=bn_decay)),
regularization_coeff=bn_decay)),
moving_mean_name=name + "_shortcut_mean", moving_mean_name=name + "_shortcut_mean",
moving_variance_name=name + "_shortcut_variance") moving_variance_name=name + "_shortcut_variance")
...@@ -515,8 +495,7 @@ class ResNeSt(nn.Layer): ...@@ -515,8 +495,7 @@ class ResNeSt(nn.Layer):
act="relu", act="relu",
name="conv1") name="conv1")
self.max_pool2d = Pool2D( self.max_pool2d = MaxPool2d(kernel_size=3, stride=2, padding=1)
pool_size=3, pool_stride=2, pool_padding=1, pool_type="max")
self.layer1 = ResNeStLayer( self.layer1 = ResNeStLayer(
inplanes=self.stem_width * 2 inplanes=self.stem_width * 2
...@@ -645,7 +624,7 @@ class ResNeSt(nn.Layer): ...@@ -645,7 +624,7 @@ class ResNeSt(nn.Layer):
stride=2, stride=2,
name="layer4") name="layer4")
self.pool2d_avg = Pool2D(pool_type='avg', global_pooling=True) self.pool2d_avg = AdaptiveAvgPool2d(1)
self.out_channels = 2048 self.out_channels = 2048
...@@ -654,7 +633,7 @@ class ResNeSt(nn.Layer): ...@@ -654,7 +633,7 @@ class ResNeSt(nn.Layer):
self.out = Linear( self.out = Linear(
self.out_channels, self.out_channels,
class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
initializer=nn.initializer.Uniform(-stdv, stdv), initializer=nn.initializer.Uniform(-stdv, stdv),
name="fc_weights"), name="fc_weights"),
bias_attr=ParamAttr(name="fc_offset")) bias_attr=ParamAttr(name="fc_offset"))
......
ppcls/modeling/architectures/resnet.py
浏览文件 @ ea746480
...@@ -18,16 +18,18 @@ from __future__ import print_function ...@@ -18,16 +18,18 @@ from __future__ import print_function
import numpy as np import numpy as np
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
__all__ = ["ResNet18", "ResNet34", "ResNet50", "ResNet101", "ResNet152"] __all__ = ["ResNet18", "ResNet34", "ResNet50", "ResNet101", "ResNet152"]
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -38,15 +40,14 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -38,15 +40,14 @@ class ConvBNLayer(fluid.dygraph.Layer):
name=None): name=None):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
if name == "conv1": if name == "conv1":
bn_name = "bn_" + name bn_name = "bn_" + name
...@@ -66,7 +67,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -66,7 +67,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return y return y
class BottleneckBlock(fluid.dygraph.Layer): class BottleneckBlock(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -117,11 +118,11 @@ class BottleneckBlock(fluid.dygraph.Layer): ...@@ -117,11 +118,11 @@ class BottleneckBlock(fluid.dygraph.Layer):
else: else:
short = self.short(inputs) short = self.short(inputs)
y = fluid.layers.elementwise_add(x=short, y=conv2, act="relu") y = paddle.elementwise_add(x=short, y=conv2, act="relu")
return y return y
class BasicBlock(fluid.dygraph.Layer): class BasicBlock(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -162,11 +163,11 @@ class BasicBlock(fluid.dygraph.Layer): ...@@ -162,11 +163,11 @@ class BasicBlock(fluid.dygraph.Layer):
short = inputs short = inputs
else: else:
short = self.short(inputs) short = self.short(inputs)
y = fluid.layers.elementwise_add(x=short, y=conv1, act="relu") y = paddle.elementwise_add(x=short, y=conv1, act="relu")
return y return y
class ResNet(fluid.dygraph.Layer): class ResNet(nn.Layer):
def __init__(self, layers=50, class_dim=1000): def __init__(self, layers=50, class_dim=1000):
super(ResNet, self).__init__() super(ResNet, self).__init__()
...@@ -195,8 +196,7 @@ class ResNet(fluid.dygraph.Layer): ...@@ -195,8 +196,7 @@ class ResNet(fluid.dygraph.Layer):
stride=2, stride=2,
act="relu", act="relu",
name="conv1") name="conv1")
self.pool2d_max = Pool2D( self.pool2d_max = MaxPool2d(kernel_size=3, stride=2, padding=1)
pool_size=3, pool_stride=2, pool_padding=1, pool_type="max")
self.block_list = [] self.block_list = []
if layers >= 50: if layers >= 50:
...@@ -238,8 +238,7 @@ class ResNet(fluid.dygraph.Layer): ...@@ -238,8 +238,7 @@ class ResNet(fluid.dygraph.Layer):
self.block_list.append(basic_block) self.block_list.append(basic_block)
shortcut = True shortcut = True
self.pool2d_avg = Pool2D( self.pool2d_avg = AdaptiveAvgPool2d(1)
pool_size=7, pool_type='avg', global_pooling=True)
self.pool2d_avg_channels = num_channels[-1] * 2 self.pool2d_avg_channels = num_channels[-1] * 2
...@@ -248,9 +247,8 @@ class ResNet(fluid.dygraph.Layer): ...@@ -248,9 +247,8 @@ class ResNet(fluid.dygraph.Layer):
self.out = Linear( self.out = Linear(
self.pool2d_avg_channels, self.pool2d_avg_channels,
class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv), initializer=Uniform(-stdv, stdv), name="fc_0.w_0"),
name="fc_0.w_0"),
bias_attr=ParamAttr(name="fc_0.b_0")) bias_attr=ParamAttr(name="fc_0.b_0"))
def forward(self, inputs): def forward(self, inputs):
...@@ -259,7 +257,7 @@ class ResNet(fluid.dygraph.Layer): ...@@ -259,7 +257,7 @@ class ResNet(fluid.dygraph.Layer):
for block in self.block_list: for block in self.block_list:
y = block(y) y = block(y)
y = self.pool2d_avg(y) y = self.pool2d_avg(y)
y = fluid.layers.reshape(y, shape=[-1, self.pool2d_avg_channels]) y = paddle.reshape(y, shape=[-1, self.pool2d_avg_channels])
y = self.out(y) y = self.out(y)
return y return y
......
ppcls/modeling/architectures/resnet_vc.py
浏览文件 @ ea746480
...@@ -18,9 +18,11 @@ from __future__ import print_function ...@@ -18,9 +18,11 @@ from __future__ import print_function
import numpy as np import numpy as np
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
...@@ -29,7 +31,7 @@ __all__ = [ ...@@ -29,7 +31,7 @@ __all__ = [
] ]
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -40,15 +42,14 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -40,15 +42,14 @@ class ConvBNLayer(fluid.dygraph.Layer):
name=None): name=None):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
if name == "conv1": if name == "conv1":
bn_name = "bn_" + name bn_name = "bn_" + name
...@@ -68,7 +69,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -68,7 +69,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return y return y
class BottleneckBlock(fluid.dygraph.Layer): class BottleneckBlock(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -119,11 +120,11 @@ class BottleneckBlock(fluid.dygraph.Layer): ...@@ -119,11 +120,11 @@ class BottleneckBlock(fluid.dygraph.Layer):
else: else:
short = self.short(inputs) short = self.short(inputs)
y = fluid.layers.elementwise_add(x=short, y=conv2, act='relu') y = paddle.elementwise_add(x=short, y=conv2, act='relu')
return y return y
class BasicBlock(fluid.dygraph.Layer): class BasicBlock(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -164,11 +165,11 @@ class BasicBlock(fluid.dygraph.Layer): ...@@ -164,11 +165,11 @@ class BasicBlock(fluid.dygraph.Layer):
short = inputs short = inputs
else: else:
short = self.short(inputs) short = self.short(inputs)
y = fluid.layers.elementwise_add(x=short, y=conv1, act='relu') y = paddle.elementwise_add(x=short, y=conv1, act='relu')
return y return y
class ResNet_vc(fluid.dygraph.Layer): class ResNet_vc(nn.Layer):
def __init__(self, layers=50, class_dim=1000): def __init__(self, layers=50, class_dim=1000):
super(ResNet_vc, self).__init__() super(ResNet_vc, self).__init__()
...@@ -212,8 +213,7 @@ class ResNet_vc(fluid.dygraph.Layer): ...@@ -212,8 +213,7 @@ class ResNet_vc(fluid.dygraph.Layer):
act='relu', act='relu',
name="conv1_3") name="conv1_3")
self.pool2d_max = Pool2D( self.pool2d_max = MaxPool2d(kernel_size=3, stride=2, padding=1)
pool_size=3, pool_stride=2, pool_padding=1, pool_type='max')
self.block_list = [] self.block_list = []
if layers >= 50: if layers >= 50:
...@@ -255,8 +255,7 @@ class ResNet_vc(fluid.dygraph.Layer): ...@@ -255,8 +255,7 @@ class ResNet_vc(fluid.dygraph.Layer):
self.block_list.append(basic_block) self.block_list.append(basic_block)
shortcut = True shortcut = True
self.pool2d_avg = Pool2D( self.pool2d_avg = AdaptiveAvgPool2d(1)
pool_size=7, pool_type='avg', global_pooling=True)
self.pool2d_avg_channels = num_channels[-1] * 2 self.pool2d_avg_channels = num_channels[-1] * 2
...@@ -265,9 +264,8 @@ class ResNet_vc(fluid.dygraph.Layer): ...@@ -265,9 +264,8 @@ class ResNet_vc(fluid.dygraph.Layer):
self.out = Linear( self.out = Linear(
self.pool2d_avg_channels, self.pool2d_avg_channels,
class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv), initializer=Uniform(-stdv, stdv), name="fc_0.w_0"),
name="fc_0.w_0"),
bias_attr=ParamAttr(name="fc_0.b_0")) bias_attr=ParamAttr(name="fc_0.b_0"))
def forward(self, inputs): def forward(self, inputs):
...@@ -278,7 +276,7 @@ class ResNet_vc(fluid.dygraph.Layer): ...@@ -278,7 +276,7 @@ class ResNet_vc(fluid.dygraph.Layer):
for block in self.block_list: for block in self.block_list:
y = block(y) y = block(y)
y = self.pool2d_avg(y) y = self.pool2d_avg(y)
y = fluid.layers.reshape(y, shape=[-1, self.pool2d_avg_channels]) y = paddle.reshape(y, shape=[-1, self.pool2d_avg_channels])
y = self.out(y) y = self.out(y)
return y return y
......
ppcls/modeling/architectures/resnet_vd.py
浏览文件 @ ea746480
...@@ -18,9 +18,11 @@ from __future__ import print_function ...@@ -18,9 +18,11 @@ from __future__ import print_function
import numpy as np import numpy as np
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
...@@ -29,7 +31,7 @@ __all__ = [ ...@@ -29,7 +31,7 @@ __all__ = [
] ]
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__( def __init__(
self, self,
num_channels, num_channels,
...@@ -43,21 +45,16 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -43,21 +45,16 @@ class ConvBNLayer(fluid.dygraph.Layer):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self.is_vd_mode = is_vd_mode self.is_vd_mode = is_vd_mode
self._pool2d_avg = Pool2D( self._pool2d_avg = AvgPool2d(
pool_size=2, kernel_size=2, stride=2, padding=0, ceil_mode=True)
pool_stride=2, self._conv = Conv2d(
pool_padding=0, in_channels=num_channels,
pool_type='avg', out_channels=num_filters,
ceil_mode=True) kernel_size=filter_size,
self._conv = Conv2D(
num_channels=num_channels,
num_filters=num_filters,
filter_size=filter_size,
stride=stride, stride=stride,
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
if name == "conv1": if name == "conv1":
bn_name = "bn_" + name bn_name = "bn_" + name
...@@ -79,7 +76,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -79,7 +76,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return y return y
class BottleneckBlock(fluid.dygraph.Layer): class BottleneckBlock(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -129,11 +126,11 @@ class BottleneckBlock(fluid.dygraph.Layer): ...@@ -129,11 +126,11 @@ class BottleneckBlock(fluid.dygraph.Layer):
short = inputs short = inputs
else: else:
short = self.short(inputs) short = self.short(inputs)
y = fluid.layers.elementwise_add(x=short, y=conv2, act='relu') y = paddle.elementwise_add(x=short, y=conv2, act='relu')
return y return y
class BasicBlock(fluid.dygraph.Layer): class BasicBlock(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -176,11 +173,11 @@ class BasicBlock(fluid.dygraph.Layer): ...@@ -176,11 +173,11 @@ class BasicBlock(fluid.dygraph.Layer):
short = inputs short = inputs
else: else:
short = self.short(inputs) short = self.short(inputs)
y = fluid.layers.elementwise_add(x=short, y=conv1, act='relu') y = paddle.elementwise_add(x=short, y=conv1, act='relu')
return y return y
class ResNet_vd(fluid.dygraph.Layer): class ResNet_vd(nn.Layer):
def __init__(self, layers=50, class_dim=1000): def __init__(self, layers=50, class_dim=1000):
super(ResNet_vd, self).__init__() super(ResNet_vd, self).__init__()
...@@ -225,8 +222,7 @@ class ResNet_vd(fluid.dygraph.Layer): ...@@ -225,8 +222,7 @@ class ResNet_vd(fluid.dygraph.Layer):
stride=1, stride=1,
act='relu', act='relu',
name="conv1_3") name="conv1_3")
self.pool2d_max = Pool2D( self.pool2d_max = MaxPool2d(kernel_size=3, stride=2, padding=1)
pool_size=3, pool_stride=2, pool_padding=1, pool_type='max')
self.block_list = [] self.block_list = []
if layers >= 50: if layers >= 50:
...@@ -270,8 +266,7 @@ class ResNet_vd(fluid.dygraph.Layer): ...@@ -270,8 +266,7 @@ class ResNet_vd(fluid.dygraph.Layer):
self.block_list.append(basic_block) self.block_list.append(basic_block)
shortcut = True shortcut = True
self.pool2d_avg = Pool2D( self.pool2d_avg = AdaptiveAvgPool2d(1)
pool_size=7, pool_type='avg', global_pooling=True)
self.pool2d_avg_channels = num_channels[-1] * 2 self.pool2d_avg_channels = num_channels[-1] * 2
...@@ -280,9 +275,8 @@ class ResNet_vd(fluid.dygraph.Layer): ...@@ -280,9 +275,8 @@ class ResNet_vd(fluid.dygraph.Layer):
self.out = Linear( self.out = Linear(
self.pool2d_avg_channels, self.pool2d_avg_channels,
class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv), initializer=Uniform(-stdv, stdv), name="fc_0.w_0"),
name="fc_0.w_0"),
bias_attr=ParamAttr(name="fc_0.b_0")) bias_attr=ParamAttr(name="fc_0.b_0"))
def forward(self, inputs): def forward(self, inputs):
...@@ -293,7 +287,7 @@ class ResNet_vd(fluid.dygraph.Layer): ...@@ -293,7 +287,7 @@ class ResNet_vd(fluid.dygraph.Layer):
for block in self.block_list: for block in self.block_list:
y = block(y) y = block(y)
y = self.pool2d_avg(y) y = self.pool2d_avg(y)
y = fluid.layers.reshape(y, shape=[-1, self.pool2d_avg_channels]) y = paddle.reshape(y, shape=[-1, self.pool2d_avg_channels])
y = self.out(y) y = self.out(y)
return y return y
......
ppcls/modeling/architectures/resnext.py
浏览文件 @ ea746480
...@@ -18,9 +18,11 @@ from __future__ import print_function ...@@ -18,9 +18,11 @@ from __future__ import print_function
import numpy as np import numpy as np
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
...@@ -30,7 +32,7 @@ __all__ = [ ...@@ -30,7 +32,7 @@ __all__ = [
] ]
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -41,15 +43,14 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -41,15 +43,14 @@ class ConvBNLayer(fluid.dygraph.Layer):
name=None): name=None):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
if name == "conv1": if name == "conv1":
bn_name = "bn_" + name bn_name = "bn_" + name
...@@ -69,7 +70,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -69,7 +70,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return y return y
class BottleneckBlock(fluid.dygraph.Layer): class BottleneckBlock(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -121,11 +122,11 @@ class BottleneckBlock(fluid.dygraph.Layer): ...@@ -121,11 +122,11 @@ class BottleneckBlock(fluid.dygraph.Layer):
else: else:
short = self.short(inputs) short = self.short(inputs)
y = fluid.layers.elementwise_add(x=short, y=conv2, act='relu') y = paddle.elementwise_add(x=short, y=conv2, act='relu')
return y return y
class ResNeXt(fluid.dygraph.Layer): class ResNeXt(nn.Layer):
def __init__(self, layers=50, class_dim=1000, cardinality=32): def __init__(self, layers=50, class_dim=1000, cardinality=32):
super(ResNeXt, self).__init__() super(ResNeXt, self).__init__()
...@@ -156,8 +157,7 @@ class ResNeXt(fluid.dygraph.Layer): ...@@ -156,8 +157,7 @@ class ResNeXt(fluid.dygraph.Layer):
stride=2, stride=2,
act='relu', act='relu',
name="res_conv1") name="res_conv1")
self.pool2d_max = Pool2D( self.pool2d_max = MaxPool2d(kernel_size=3, stride=2, padding=1)
pool_size=3, pool_stride=2, pool_padding=1, pool_type='max')
self.block_list = [] self.block_list = []
for block in range(len(depth)): for block in range(len(depth)):
...@@ -183,8 +183,7 @@ class ResNeXt(fluid.dygraph.Layer): ...@@ -183,8 +183,7 @@ class ResNeXt(fluid.dygraph.Layer):
self.block_list.append(bottleneck_block) self.block_list.append(bottleneck_block)
shortcut = True shortcut = True
self.pool2d_avg = Pool2D( self.pool2d_avg = AdaptiveAvgPool2d(1)
pool_size=7, pool_type='avg', global_pooling=True)
self.pool2d_avg_channels = num_channels[-1] * 2 self.pool2d_avg_channels = num_channels[-1] * 2
...@@ -193,9 +192,8 @@ class ResNeXt(fluid.dygraph.Layer): ...@@ -193,9 +192,8 @@ class ResNeXt(fluid.dygraph.Layer):
self.out = Linear( self.out = Linear(
self.pool2d_avg_channels, self.pool2d_avg_channels,
class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv), initializer=Uniform(-stdv, stdv), name="fc_weights"),
name="fc_weights"),
bias_attr=ParamAttr(name="fc_offset")) bias_attr=ParamAttr(name="fc_offset"))
def forward(self, inputs): def forward(self, inputs):
...@@ -204,7 +202,7 @@ class ResNeXt(fluid.dygraph.Layer): ...@@ -204,7 +202,7 @@ class ResNeXt(fluid.dygraph.Layer):
for block in self.block_list: for block in self.block_list:
y = block(y) y = block(y)
y = self.pool2d_avg(y) y = self.pool2d_avg(y)
y = fluid.layers.reshape(y, shape=[-1, self.pool2d_avg_channels]) y = paddle.reshape(y, shape=[-1, self.pool2d_avg_channels])
y = self.out(y) y = self.out(y)
return y return y
......
ppcls/modeling/architectures/resnext101_wsl.py
浏览文件 @ ea746480
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
__all__ = ["ResNeXt101_32x8d_wsl", from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
"ResNeXt101_wsl_32x16d_wsl", from paddle.nn.initializer import Uniform
"ResNeXt101_wsl_32x32d_wsl",
"ResNeXt101_wsl_32x48d_wsl"] __all__ = [
"ResNeXt101_32x8d_wsl", "ResNeXt101_32x16d_wsl", "ResNeXt101_32x32d_wsl",
class ConvBNLayer(fluid.dygraph.Layer): "ResNeXt101_32x48d_wsl"
def __init__(self, ]
input_channels,
output_channels,
filter_size, class ConvBNLayer(nn.Layer):
stride=1, def __init__(self,
groups=1, input_channels,
act=None, output_channels,
name=None): filter_size,
stride=1,
groups=1,
act=None,
name=None):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
if "downsample" in name: if "downsample" in name:
conv_name = name + ".0" conv_name = name + ".0"
else: else:
conv_name = name conv_name = name
self._conv = Conv2D(num_channels=input_channels, self._conv = Conv2d(
num_filters=output_channels, in_channels=input_channels,
filter_size=filter_size, out_channels=output_channels,
stride=stride, kernel_size=filter_size,
padding=(filter_size-1)//2, stride=stride,
groups=groups, padding=(filter_size - 1) // 2,
act=None, groups=groups,
param_attr=ParamAttr(name=conv_name + ".weight"), weight_attr=ParamAttr(name=conv_name + ".weight"),
bias_attr=False) bias_attr=False)
if "downsample" in name: if "downsample" in name:
bn_name = name[:9] + "downsample.1" bn_name = name[:9] + "downsample.1"
else: else:
if "conv1" == name: if "conv1" == name:
bn_name = "bn" + name[-1] bn_name = "bn" + name[-1]
else: else:
bn_name = (name[:10] if name[7:9].isdigit() else name[:9]) + "bn" + name[-1] bn_name = (name[:10] if name[7:9].isdigit() else name[:9]
self._bn = BatchNorm(num_channels=output_channels, ) + "bn" + name[-1]
act=act, self._bn = BatchNorm(
param_attr=ParamAttr(name=bn_name + ".weight"), num_channels=output_channels,
bias_attr=ParamAttr(name=bn_name + ".bias"), act=act,
moving_mean_name=bn_name + ".running_mean", param_attr=ParamAttr(name=bn_name + ".weight"),
moving_variance_name=bn_name + ".running_var") bias_attr=ParamAttr(name=bn_name + ".bias"),
moving_mean_name=bn_name + ".running_mean",
moving_variance_name=bn_name + ".running_var")
def forward(self, inputs): def forward(self, inputs):
x = self._conv(inputs) x = self._conv(inputs)
x = self._bn(x) x = self._bn(x)
return x return x
class ShortCut(fluid.dygraph.Layer):
class ShortCut(nn.Layer):
def __init__(self, input_channels, output_channels, stride, name=None): def __init__(self, input_channels, output_channels, stride, name=None):
super(ShortCut, self).__init__() super(ShortCut, self).__init__()
self.input_channels = input_channels self.input_channels = input_channels
self.output_channels = output_channels self.output_channels = output_channels
self.stride = stride self.stride = stride
if input_channels!=output_channels or stride!=1: if input_channels != output_channels or stride != 1:
self._conv = ConvBNLayer( self._conv = ConvBNLayer(
input_channels, output_channels, filter_size=1, stride=stride, name=name) input_channels,
output_channels,
filter_size=1,
stride=stride,
name=name)
def forward(self, inputs): def forward(self, inputs):
if self.input_channels!= self.output_channels or self.stride!=1: if self.input_channels != self.output_channels or self.stride != 1:
return self._conv(inputs) return self._conv(inputs)
return inputs return inputs
class BottleneckBlock(fluid.dygraph.Layer):
def __init__(self, input_channels, output_channels, stride, cardinality, width, name): class BottleneckBlock(nn.Layer):
def __init__(self, input_channels, output_channels, stride, cardinality,
width, name):
super(BottleneckBlock, self).__init__() super(BottleneckBlock, self).__init__()
self._conv0 = ConvBNLayer( self._conv0 = ConvBNLayer(
input_channels, output_channels, filter_size=1, act="relu", name=name + ".conv1") input_channels,
output_channels,
filter_size=1,
act="relu",
name=name + ".conv1")
self._conv1 = ConvBNLayer( self._conv1 = ConvBNLayer(
output_channels, output_channels, filter_size=3, act="relu", stride=stride, groups=cardinality, name=name + ".conv2") output_channels,
output_channels,
filter_size=3,
act="relu",
stride=stride,
groups=cardinality,
name=name + ".conv2")
self._conv2 = ConvBNLayer( self._conv2 = ConvBNLayer(
output_channels, output_channels//(width//8), filter_size=1, act=None, name=name + ".conv3") output_channels,
output_channels // (width // 8),
filter_size=1,
act=None,
name=name + ".conv3")
self._short = ShortCut( self._short = ShortCut(
input_channels, output_channels//(width//8), stride=stride, name=name + ".downsample") input_channels,
output_channels // (width // 8),
stride=stride,
name=name + ".downsample")
def forward(self, inputs): def forward(self, inputs):
x = self._conv0(inputs) x = self._conv0(inputs)
x = self._conv1(x) x = self._conv1(x)
x = self._conv2(x) x = self._conv2(x)
y = self._short(inputs) y = self._short(inputs)
return fluid.layers.elementwise_add(x, y, act="relu") return paddle.elementwise_add(x, y, act="relu")
class ResNeXt101WSL(fluid.dygraph.Layer): class ResNeXt101WSL(nn.Layer):
def __init__(self, layers=101, cardinality=32, width=48, class_dim=1000): def __init__(self, layers=101, cardinality=32, width=48, class_dim=1000):
super(ResNeXt101WSL, self).__init__() super(ResNeXt101WSL, self).__init__()
...@@ -95,93 +126,257 @@ class ResNeXt101WSL(fluid.dygraph.Layer): ...@@ -95,93 +126,257 @@ class ResNeXt101WSL(fluid.dygraph.Layer):
self.layers = layers self.layers = layers
self.cardinality = cardinality self.cardinality = cardinality
self.width = width self.width = width
self.scale = width//8 self.scale = width // 8
self.depth = [3, 4, 23, 3] self.depth = [3, 4, 23, 3]
self.base_width = cardinality * width self.base_width = cardinality * width
num_filters = [self.base_width*i for i in [1,2,4,8]] #[256, 512, 1024, 2048] num_filters = [self.base_width * i
for i in [1, 2, 4, 8]] # [256, 512, 1024, 2048]
self._conv_stem = ConvBNLayer( self._conv_stem = ConvBNLayer(
3, 64, 7, stride=2, act="relu", name="conv1") 3, 64, 7, stride=2, act="relu", name="conv1")
self._pool = Pool2D(pool_size=3, self._pool = MaxPool2d(kernel_size=3, stride=2, padding=1)
pool_stride=2,
pool_padding=1,
pool_type="max")
self._conv1_0 = BottleneckBlock( self._conv1_0 = BottleneckBlock(
64, num_filters[0], stride=1, cardinality=self.cardinality, width=self.width, name="layer1.0") 64,
num_filters[0],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer1.0")
self._conv1_1 = BottleneckBlock( self._conv1_1 = BottleneckBlock(
num_filters[0]//(width//8), num_filters[0], stride=1, cardinality=self.cardinality, width=self.width, name="layer1.1") num_filters[0] // (width // 8),
num_filters[0],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer1.1")
self._conv1_2 = BottleneckBlock( self._conv1_2 = BottleneckBlock(
num_filters[0]//(width//8), num_filters[0], stride=1, cardinality=self.cardinality, width=self.width, name="layer1.2") num_filters[0] // (width // 8),
num_filters[0],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer1.2")
self._conv2_0 = BottleneckBlock( self._conv2_0 = BottleneckBlock(
num_filters[0]//(width//8), num_filters[1], stride=2, cardinality=self.cardinality, width=self.width, name="layer2.0") num_filters[0] // (width // 8),
num_filters[1],
stride=2,
cardinality=self.cardinality,
width=self.width,
name="layer2.0")
self._conv2_1 = BottleneckBlock( self._conv2_1 = BottleneckBlock(
num_filters[1]//(width//8), num_filters[1], stride=1, cardinality=self.cardinality, width=self.width, name="layer2.1") num_filters[1] // (width // 8),
num_filters[1],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer2.1")
self._conv2_2 = BottleneckBlock( self._conv2_2 = BottleneckBlock(
num_filters[1]//(width//8), num_filters[1], stride=1, cardinality=self.cardinality, width=self.width, name="layer2.2") num_filters[1] // (width // 8),
num_filters[1],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer2.2")
self._conv2_3 = BottleneckBlock( self._conv2_3 = BottleneckBlock(
num_filters[1]//(width//8), num_filters[1], stride=1, cardinality=self.cardinality, width=self.width, name="layer2.3") num_filters[1] // (width // 8),
num_filters[1],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer2.3")
self._conv3_0 = BottleneckBlock( self._conv3_0 = BottleneckBlock(
num_filters[1]//(width//8), num_filters[2], stride=2, cardinality=self.cardinality, width=self.width, name="layer3.0") num_filters[1] // (width // 8),
num_filters[2],
stride=2,
cardinality=self.cardinality,
width=self.width,
name="layer3.0")
self._conv3_1 = BottleneckBlock( self._conv3_1 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.1") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.1")
self._conv3_2 = BottleneckBlock( self._conv3_2 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.2") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.2")
self._conv3_3 = BottleneckBlock( self._conv3_3 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.3") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.3")
self._conv3_4 = BottleneckBlock( self._conv3_4 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.4") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.4")
self._conv3_5 = BottleneckBlock( self._conv3_5 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.5") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.5")
self._conv3_6 = BottleneckBlock( self._conv3_6 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.6") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.6")
self._conv3_7 = BottleneckBlock( self._conv3_7 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.7") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.7")
self._conv3_8 = BottleneckBlock( self._conv3_8 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.8") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.8")
self._conv3_9 = BottleneckBlock( self._conv3_9 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.9") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.9")
self._conv3_10 = BottleneckBlock( self._conv3_10 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.10") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.10")
self._conv3_11 = BottleneckBlock( self._conv3_11 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.11") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.11")
self._conv3_12 = BottleneckBlock( self._conv3_12 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.12") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.12")
self._conv3_13 = BottleneckBlock( self._conv3_13 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.13") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.13")
self._conv3_14 = BottleneckBlock( self._conv3_14 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.14") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.14")
self._conv3_15 = BottleneckBlock( self._conv3_15 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.15") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.15")
self._conv3_16 = BottleneckBlock( self._conv3_16 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.16") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.16")
self._conv3_17 = BottleneckBlock( self._conv3_17 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.17") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.17")
self._conv3_18 = BottleneckBlock( self._conv3_18 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.18") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.18")
self._conv3_19 = BottleneckBlock( self._conv3_19 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.19") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.19")
self._conv3_20 = BottleneckBlock( self._conv3_20 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.20") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.20")
self._conv3_21 = BottleneckBlock( self._conv3_21 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.21") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.21")
self._conv3_22 = BottleneckBlock( self._conv3_22 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[2], stride=1, cardinality=self.cardinality, width=self.width, name="layer3.22") num_filters[2] // (width // 8),
num_filters[2],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer3.22")
self._conv4_0 = BottleneckBlock( self._conv4_0 = BottleneckBlock(
num_filters[2]//(width//8), num_filters[3], stride=2, cardinality=self.cardinality, width=self.width, name="layer4.0") num_filters[2] // (width // 8),
num_filters[3],
stride=2,
cardinality=self.cardinality,
width=self.width,
name="layer4.0")
self._conv4_1 = BottleneckBlock( self._conv4_1 = BottleneckBlock(
num_filters[3]//(width//8), num_filters[3], stride=1, cardinality=self.cardinality, width=self.width, name="layer4.1") num_filters[3] // (width // 8),
num_filters[3],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer4.1")
self._conv4_2 = BottleneckBlock( self._conv4_2 = BottleneckBlock(
num_filters[3]//(width//8), num_filters[3], stride=1, cardinality=self.cardinality, width=self.width, name="layer4.2") num_filters[3] // (width // 8),
num_filters[3],
stride=1,
cardinality=self.cardinality,
width=self.width,
name="layer4.2")
self._avg_pool = Pool2D(pool_type="avg", global_pooling=True) self._avg_pool = AdaptiveAvgPool2d(1)
self._out = Linear(input_dim=num_filters[3]//(width//8), self._out = Linear(
output_dim=class_dim, num_filters[3] // (width // 8),
param_attr=ParamAttr(name="fc.weight"), class_dim,
bias_attr=ParamAttr(name="fc.bias")) weight_attr=ParamAttr(name="fc.weight"),
bias_attr=ParamAttr(name="fc.bias"))
def forward(self, inputs): def forward(self, inputs):
x = self._conv_stem(inputs) x = self._conv_stem(inputs)
...@@ -225,22 +420,26 @@ class ResNeXt101WSL(fluid.dygraph.Layer): ...@@ -225,22 +420,26 @@ class ResNeXt101WSL(fluid.dygraph.Layer):
x = self._conv4_2(x) x = self._conv4_2(x)
x = self._avg_pool(x) x = self._avg_pool(x)
x = fluid.layers.squeeze(x, axes=[2, 3]) x = paddle.squeeze(x, axis=[2, 3])
x = self._out(x) x = self._out(x)
return x return x
def ResNeXt101_32x8d_wsl(**args): def ResNeXt101_32x8d_wsl(**args):
model = ResNeXt101WSL(cardinality=32, width=8, **args) model = ResNeXt101WSL(cardinality=32, width=8, **args)
return model return model
def ResNeXt101_32x16d_wsl(**args): def ResNeXt101_32x16d_wsl(**args):
model = ResNeXt101WSL(cardinality=32, width=16, **args) model = ResNeXt101WSL(cardinality=32, width=16, **args)
return model return model
def ResNeXt101_32x32d_wsl(**args): def ResNeXt101_32x32d_wsl(**args):
model = ResNeXt101WSL(cardinality=32, width=32, **args) model = ResNeXt101WSL(cardinality=32, width=32, **args)
return model return model
def ResNeXt101_32x48d_wsl(**args): def ResNeXt101_32x48d_wsl(**args):
model = ResNeXt101WSL(cardinality=32, width=48, **args) model = ResNeXt101WSL(cardinality=32, width=48, **args)
return model return model
\ No newline at end of file
ppcls/modeling/architectures/resnext_vd.py
浏览文件 @ ea746480
...@@ -18,9 +18,11 @@ from __future__ import print_function ...@@ -18,9 +18,11 @@ from __future__ import print_function
import numpy as np import numpy as np
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
...@@ -30,7 +32,7 @@ __all__ = [ ...@@ -30,7 +32,7 @@ __all__ = [
] ]
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__( def __init__(
self, self,
num_channels, num_channels,
...@@ -44,21 +46,16 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -44,21 +46,16 @@ class ConvBNLayer(fluid.dygraph.Layer):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self.is_vd_mode = is_vd_mode self.is_vd_mode = is_vd_mode
self._pool2d_avg = Pool2D( self._pool2d_avg = AvgPool2d(
pool_size=2, kernel_size=2, stride=2, padding=0, ceil_mode=True)
pool_stride=2, self._conv = Conv2d(
pool_padding=0, in_channels=num_channels,
pool_type='avg', out_channels=num_filters,
ceil_mode=True) kernel_size=filter_size,
self._conv = Conv2D(
num_channels=num_channels,
num_filters=num_filters,
filter_size=filter_size,
stride=stride, stride=stride,
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
if name == "conv1": if name == "conv1":
bn_name = "bn_" + name bn_name = "bn_" + name
...@@ -80,7 +77,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -80,7 +77,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return y return y
class BottleneckBlock(fluid.dygraph.Layer): class BottleneckBlock(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -134,11 +131,11 @@ class BottleneckBlock(fluid.dygraph.Layer): ...@@ -134,11 +131,11 @@ class BottleneckBlock(fluid.dygraph.Layer):
else: else:
short = self.short(inputs) short = self.short(inputs)
y = fluid.layers.elementwise_add(x=short, y=conv2, act='relu') y = paddle.elementwise_add(x=short, y=conv2, act='relu')
return y return y
class ResNeXt(fluid.dygraph.Layer): class ResNeXt(nn.Layer):
def __init__(self, layers=50, class_dim=1000, cardinality=32): def __init__(self, layers=50, class_dim=1000, cardinality=32):
super(ResNeXt, self).__init__() super(ResNeXt, self).__init__()
...@@ -184,8 +181,7 @@ class ResNeXt(fluid.dygraph.Layer): ...@@ -184,8 +181,7 @@ class ResNeXt(fluid.dygraph.Layer):
act='relu', act='relu',
name="conv1_3") name="conv1_3")
self.pool2d_max = Pool2D( self.pool2d_max = MaxPool2d(kernel_size=3, stride=2, padding=1)
pool_size=3, pool_stride=2, pool_padding=1, pool_type='max')
self.block_list = [] self.block_list = []
for block in range(len(depth)): for block in range(len(depth)):
...@@ -212,8 +208,7 @@ class ResNeXt(fluid.dygraph.Layer): ...@@ -212,8 +208,7 @@ class ResNeXt(fluid.dygraph.Layer):
self.block_list.append(bottleneck_block) self.block_list.append(bottleneck_block)
shortcut = True shortcut = True
self.pool2d_avg = Pool2D( self.pool2d_avg = AdaptiveAvgPool2d(1)
pool_size=7, pool_type='avg', global_pooling=True)
self.pool2d_avg_channels = num_channels[-1] * 2 self.pool2d_avg_channels = num_channels[-1] * 2
...@@ -222,9 +217,8 @@ class ResNeXt(fluid.dygraph.Layer): ...@@ -222,9 +217,8 @@ class ResNeXt(fluid.dygraph.Layer):
self.out = Linear( self.out = Linear(
self.pool2d_avg_channels, self.pool2d_avg_channels,
class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv), initializer=Uniform(-stdv, stdv), name="fc_weights"),
name="fc_weights"),
bias_attr=ParamAttr(name="fc_offset")) bias_attr=ParamAttr(name="fc_offset"))
def forward(self, inputs): def forward(self, inputs):
...@@ -235,7 +229,7 @@ class ResNeXt(fluid.dygraph.Layer): ...@@ -235,7 +229,7 @@ class ResNeXt(fluid.dygraph.Layer):
for block in self.block_list: for block in self.block_list:
y = block(y) y = block(y)
y = self.pool2d_avg(y) y = self.pool2d_avg(y)
y = fluid.layers.reshape(y, shape=[-1, self.pool2d_avg_channels]) y = paddle.reshape(y, shape=[-1, self.pool2d_avg_channels])
y = self.out(y) y = self.out(y)
return y return y
......
ppcls/modeling/architectures/se_resnet_vd.py
浏览文件 @ ea746480
...@@ -17,9 +17,12 @@ from __future__ import print_function ...@@ -17,9 +17,12 @@ from __future__ import print_function
import numpy as np import numpy as np
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
...@@ -29,7 +32,7 @@ __all__ = [ ...@@ -29,7 +32,7 @@ __all__ = [
] ]
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__( def __init__(
self, self,
num_channels, num_channels,
...@@ -43,21 +46,17 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -43,21 +46,17 @@ class ConvBNLayer(fluid.dygraph.Layer):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self.is_vd_mode = is_vd_mode self.is_vd_mode = is_vd_mode
self._pool2d_avg = Pool2D( self._pool2d_avg = AvgPool2d(
pool_size=2, kernel_size=2, stride=2, padding=0, ceil_mode=True)
pool_stride=2,
pool_padding=0, self._conv = Conv2d(
pool_type='avg', in_channels=num_channels,
ceil_mode=True) out_channels=num_filters,
self._conv = Conv2D( kernel_size=filter_size,
num_channels=num_channels,
num_filters=num_filters,
filter_size=filter_size,
stride=stride, stride=stride,
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
if name == "conv1": if name == "conv1":
bn_name = "bn_" + name bn_name = "bn_" + name
...@@ -79,7 +78,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -79,7 +78,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return y return y
class BottleneckBlock(fluid.dygraph.Layer): class BottleneckBlock(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -136,11 +135,11 @@ class BottleneckBlock(fluid.dygraph.Layer): ...@@ -136,11 +135,11 @@ class BottleneckBlock(fluid.dygraph.Layer):
short = inputs short = inputs
else: else:
short = self.short(inputs) short = self.short(inputs)
y = fluid.layers.elementwise_add(x=short, y=scale, act='relu') y = paddle.elementwise_add(x=short, y=scale, act='relu')
return y return y
class BasicBlock(fluid.dygraph.Layer): class BasicBlock(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -191,15 +190,15 @@ class BasicBlock(fluid.dygraph.Layer): ...@@ -191,15 +190,15 @@ class BasicBlock(fluid.dygraph.Layer):
short = inputs short = inputs
else: else:
short = self.short(inputs) short = self.short(inputs)
y = fluid.layers.elementwise_add(x=short, y=scale, act='relu') y = paddle.elementwise_add(x=short, y=scale, act='relu')
return y return y
class SELayer(fluid.dygraph.Layer): class SELayer(nn.Layer):
def __init__(self, num_channels, num_filters, reduction_ratio, name=None): def __init__(self, num_channels, num_filters, reduction_ratio, name=None):
super(SELayer, self).__init__() super(SELayer, self).__init__()
self.pool2d_gap = Pool2D(pool_type='avg', global_pooling=True) self.pool2d_gap = AdaptiveAvgPool2d(1)
self._num_channels = num_channels self._num_channels = num_channels
...@@ -208,34 +207,32 @@ class SELayer(fluid.dygraph.Layer): ...@@ -208,34 +207,32 @@ class SELayer(fluid.dygraph.Layer):
self.squeeze = Linear( self.squeeze = Linear(
num_channels, num_channels,
med_ch, med_ch,
act="relu", weight_attr=ParamAttr(
param_attr=ParamAttr( initializer=Uniform(-stdv, stdv), name=name + "_sqz_weights"),
initializer=fluid.initializer.Uniform(-stdv, stdv),
name=name + "_sqz_weights"),
bias_attr=ParamAttr(name=name + '_sqz_offset')) bias_attr=ParamAttr(name=name + '_sqz_offset'))
stdv = 1.0 / math.sqrt(med_ch * 1.0) stdv = 1.0 / math.sqrt(med_ch * 1.0)
self.excitation = Linear( self.excitation = Linear(
med_ch, med_ch,
num_filters, num_filters,
act="sigmoid", weight_attr=ParamAttr(
param_attr=ParamAttr( initializer=Uniform(-stdv, stdv), name=name + "_exc_weights"),
initializer=fluid.initializer.Uniform(-stdv, stdv),
name=name + "_exc_weights"),
bias_attr=ParamAttr(name=name + '_exc_offset')) bias_attr=ParamAttr(name=name + '_exc_offset'))
def forward(self, input): def forward(self, input):
pool = self.pool2d_gap(input) pool = self.pool2d_gap(input)
pool = fluid.layers.reshape(pool, shape=[-1, self._num_channels]) pool = paddle.reshape(pool, shape=[-1, self._num_channels])
squeeze = self.squeeze(pool) squeeze = self.squeeze(pool)
squeeze = F.relu(squeeze)
excitation = self.excitation(squeeze) excitation = self.excitation(squeeze)
excitation = fluid.layers.reshape( excitation = F.sigmoid(excitation)
excitation = paddle.reshape(
excitation, shape=[-1, self._num_channels, 1, 1]) excitation, shape=[-1, self._num_channels, 1, 1])
out = input * excitation out = input * excitation
return out return out
class SE_ResNet_vd(fluid.dygraph.Layer): class SE_ResNet_vd(nn.Layer):
def __init__(self, layers=50, class_dim=1000): def __init__(self, layers=50, class_dim=1000):
super(SE_ResNet_vd, self).__init__() super(SE_ResNet_vd, self).__init__()
...@@ -280,8 +277,7 @@ class SE_ResNet_vd(fluid.dygraph.Layer): ...@@ -280,8 +277,7 @@ class SE_ResNet_vd(fluid.dygraph.Layer):
stride=1, stride=1,
act='relu', act='relu',
name="conv1_3") name="conv1_3")
self.pool2d_max = Pool2D( self.pool2d_max = MaxPool2d(kernel_size=3, stride=2, padding=1)
pool_size=3, pool_stride=2, pool_padding=1, pool_type='max')
self.block_list = [] self.block_list = []
if layers >= 50: if layers >= 50:
...@@ -325,8 +321,7 @@ class SE_ResNet_vd(fluid.dygraph.Layer): ...@@ -325,8 +321,7 @@ class SE_ResNet_vd(fluid.dygraph.Layer):
self.block_list.append(basic_block) self.block_list.append(basic_block)
shortcut = True shortcut = True
self.pool2d_avg = Pool2D( self.pool2d_avg = AdaptiveAvgPool2d(1)
pool_size=7, pool_type='avg', global_pooling=True)
self.pool2d_avg_channels = num_channels[-1] * 2 self.pool2d_avg_channels = num_channels[-1] * 2
...@@ -335,9 +330,8 @@ class SE_ResNet_vd(fluid.dygraph.Layer): ...@@ -335,9 +330,8 @@ class SE_ResNet_vd(fluid.dygraph.Layer):
self.out = Linear( self.out = Linear(
self.pool2d_avg_channels, self.pool2d_avg_channels,
class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv), initializer=Uniform(-stdv, stdv), name="fc6_weights"),
name="fc6_weights"),
bias_attr=ParamAttr(name="fc6_offset")) bias_attr=ParamAttr(name="fc6_offset"))
def forward(self, inputs): def forward(self, inputs):
...@@ -348,7 +342,7 @@ class SE_ResNet_vd(fluid.dygraph.Layer): ...@@ -348,7 +342,7 @@ class SE_ResNet_vd(fluid.dygraph.Layer):
for block in self.block_list: for block in self.block_list:
y = block(y) y = block(y)
y = self.pool2d_avg(y) y = self.pool2d_avg(y)
y = fluid.layers.reshape(y, shape=[-1, self.pool2d_avg_channels]) y = paddle.reshape(y, shape=[-1, self.pool2d_avg_channels])
y = self.out(y) y = self.out(y)
return y return y
......
ppcls/modeling/architectures/shufflenet_v2.py
浏览文件 @ ea746480
...@@ -18,15 +18,17 @@ from __future__ import print_function ...@@ -18,15 +18,17 @@ from __future__ import print_function
import numpy as np import numpy as np
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout import paddle.nn.functional as F
from paddle.fluid.initializer import MSRA from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import MSRA
import math import math
__all__ = [ __all__ = [
"ShuffleNetV2_x0_25", "ShuffleNetV2_x0_33", "ShuffleNetV2_x0_5", "ShuffleNetV2_x0_25", "ShuffleNetV2_x0_33", "ShuffleNetV2_x0_5",
"ShuffleNetV2_x1_0", "ShuffleNetV2_x1_5", "ShuffleNetV2_x2_0", "ShuffleNetV2", "ShuffleNetV2_x1_5", "ShuffleNetV2_x2_0",
"ShuffleNetV2_swish" "ShuffleNetV2_swish"
] ]
...@@ -37,17 +39,16 @@ def channel_shuffle(x, groups): ...@@ -37,17 +39,16 @@ def channel_shuffle(x, groups):
channels_per_group = num_channels // groups channels_per_group = num_channels // groups
# reshape # reshape
x = fluid.layers.reshape( x = paddle.reshape(
x=x, shape=[batchsize, groups, channels_per_group, height, width]) x=x, shape=[batchsize, groups, channels_per_group, height, width])
x = fluid.layers.transpose(x=x, perm=[0, 2, 1, 3, 4]) x = paddle.transpose(x=x, perm=[0, 2, 1, 3, 4])
# flatten # flatten
x = fluid.layers.reshape( x = paddle.reshape(x=x, shape=[batchsize, num_channels, height, width])
x=x, shape=[batchsize, num_channels, height, width])
return x return x
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
filter_size, filter_size,
...@@ -58,24 +59,21 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -58,24 +59,21 @@ class ConvBNLayer(fluid.dygraph.Layer):
num_groups=1, num_groups=1,
if_act=True, if_act=True,
act='relu', act='relu',
name=None, name=None):
use_cudnn=True):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self._if_act = if_act self._if_act = if_act
assert act in ['relu', 'swish'], \ assert act in ['relu', 'swish'], \
"supported act are {} but your act is {}".format( "supported act are {} but your act is {}".format(
['relu', 'swish'], act) ['relu', 'swish'], act)
self._act = act self._act = act
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=padding, padding=padding,
groups=num_groups, groups=num_groups,
act=None, weight_attr=ParamAttr(
use_cudnn=use_cudnn,
param_attr=ParamAttr(
initializer=MSRA(), name=name + "_weights"), initializer=MSRA(), name=name + "_weights"),
bias_attr=False) bias_attr=False)
...@@ -90,12 +88,11 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -90,12 +88,11 @@ class ConvBNLayer(fluid.dygraph.Layer):
y = self._conv(inputs) y = self._conv(inputs)
y = self._batch_norm(y) y = self._batch_norm(y)
if self._if_act: if self._if_act:
y = fluid.layers.relu( y = F.relu(y) if self._act == 'relu' else F.swish(y)
y) if self._act == 'relu' else fluid.layers.swish(y)
return y return y
class InvertedResidualUnit(fluid.dygraph.Layer): class InvertedResidualUnit(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -130,7 +127,6 @@ class InvertedResidualUnit(fluid.dygraph.Layer): ...@@ -130,7 +127,6 @@ class InvertedResidualUnit(fluid.dygraph.Layer):
num_groups=oup_inc, num_groups=oup_inc,
if_act=False, if_act=False,
act=act, act=act,
use_cudnn=False,
name='stage_' + name + '_conv2') name='stage_' + name + '_conv2')
self._conv_linear = ConvBNLayer( self._conv_linear = ConvBNLayer(
num_channels=oup_inc, num_channels=oup_inc,
...@@ -153,7 +149,6 @@ class InvertedResidualUnit(fluid.dygraph.Layer): ...@@ -153,7 +149,6 @@ class InvertedResidualUnit(fluid.dygraph.Layer):
num_groups=inp, num_groups=inp,
if_act=False, if_act=False,
act=act, act=act,
use_cudnn=False,
name='stage_' + name + '_conv4') name='stage_' + name + '_conv4')
self._conv_linear_1 = ConvBNLayer( self._conv_linear_1 = ConvBNLayer(
num_channels=inp, num_channels=inp,
...@@ -185,7 +180,6 @@ class InvertedResidualUnit(fluid.dygraph.Layer): ...@@ -185,7 +180,6 @@ class InvertedResidualUnit(fluid.dygraph.Layer):
num_groups=oup_inc, num_groups=oup_inc,
if_act=False, if_act=False,
act=act, act=act,
use_cudnn=False,
name='stage_' + name + '_conv2') name='stage_' + name + '_conv2')
self._conv_linear_2 = ConvBNLayer( self._conv_linear_2 = ConvBNLayer(
num_channels=oup_inc, num_channels=oup_inc,
...@@ -200,14 +194,14 @@ class InvertedResidualUnit(fluid.dygraph.Layer): ...@@ -200,14 +194,14 @@ class InvertedResidualUnit(fluid.dygraph.Layer):
def forward(self, inputs): def forward(self, inputs):
if self.benchmodel == 1: if self.benchmodel == 1:
x1, x2 = fluid.layers.split( x1, x2 = paddle.split(
inputs, inputs,
num_or_sections=[inputs.shape[1] // 2, inputs.shape[1] // 2], num_or_sections=[inputs.shape[1] // 2, inputs.shape[1] // 2],
dim=1) axis=1)
x2 = self._conv_pw(x2) x2 = self._conv_pw(x2)
x2 = self._conv_dw(x2) x2 = self._conv_dw(x2)
x2 = self._conv_linear(x2) x2 = self._conv_linear(x2)
out = fluid.layers.concat([x1, x2], axis=1) out = paddle.concat([x1, x2], axis=1)
else: else:
x1 = self._conv_dw_1(inputs) x1 = self._conv_dw_1(inputs)
x1 = self._conv_linear_1(x1) x1 = self._conv_linear_1(x1)
...@@ -215,12 +209,12 @@ class InvertedResidualUnit(fluid.dygraph.Layer): ...@@ -215,12 +209,12 @@ class InvertedResidualUnit(fluid.dygraph.Layer):
x2 = self._conv_pw_2(inputs) x2 = self._conv_pw_2(inputs)
x2 = self._conv_dw_2(x2) x2 = self._conv_dw_2(x2)
x2 = self._conv_linear_2(x2) x2 = self._conv_linear_2(x2)
out = fluid.layers.concat([x1, x2], axis=1) out = paddle.concat([x1, x2], axis=1)
return channel_shuffle(out, 2) return channel_shuffle(out, 2)
class ShuffleNet(fluid.dygraph.Layer): class ShuffleNet(nn.Layer):
def __init__(self, class_dim=1000, scale=1.0, act='relu'): def __init__(self, class_dim=1000, scale=1.0, act='relu'):
super(ShuffleNet, self).__init__() super(ShuffleNet, self).__init__()
self.scale = scale self.scale = scale
...@@ -252,8 +246,7 @@ class ShuffleNet(fluid.dygraph.Layer): ...@@ -252,8 +246,7 @@ class ShuffleNet(fluid.dygraph.Layer):
if_act=True, if_act=True,
act=act, act=act,
name='stage1_conv') name='stage1_conv')
self._max_pool = Pool2D( self._max_pool = MaxPool2d(kernel_size=3, stride=2, padding=1)
pool_type='max', pool_size=3, pool_stride=2, pool_padding=1)
# 2. bottleneck sequences # 2. bottleneck sequences
self._block_list = [] self._block_list = []
...@@ -298,13 +291,13 @@ class ShuffleNet(fluid.dygraph.Layer): ...@@ -298,13 +291,13 @@ class ShuffleNet(fluid.dygraph.Layer):
name='conv5') name='conv5')
# 4. pool # 4. pool
self._pool2d_avg = Pool2D(pool_type='avg', global_pooling=True) self._pool2d_avg = AdaptiveAvgPool2d(1)
self._out_c = stage_out_channels[-1] self._out_c = stage_out_channels[-1]
# 5. fc # 5. fc
self._fc = Linear( self._fc = Linear(
stage_out_channels[-1], stage_out_channels[-1],
class_dim, class_dim,
param_attr=ParamAttr(name='fc6_weights'), weight_attr=ParamAttr(name='fc6_weights'),
bias_attr=ParamAttr(name='fc6_offset')) bias_attr=ParamAttr(name='fc6_offset'))
def forward(self, inputs): def forward(self, inputs):
...@@ -314,7 +307,7 @@ class ShuffleNet(fluid.dygraph.Layer): ...@@ -314,7 +307,7 @@ class ShuffleNet(fluid.dygraph.Layer):
y = inv(y) y = inv(y)
y = self._last_conv(y) y = self._last_conv(y)
y = self._pool2d_avg(y) y = self._pool2d_avg(y)
y = fluid.layers.reshape(y, shape=[-1, self._out_c]) y = paddle.reshape(y, shape=[-1, self._out_c])
y = self._fc(y) y = self._fc(y)
return y return y
......
ppcls/modeling/architectures/squeezenet.py
浏览文件 @ ea746480
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
__all__ = ["SqueezeNet1_0", "SqueezeNet1_1"] __all__ = ["SqueezeNet1_0", "SqueezeNet1_1"]
class MakeFireConv(fluid.dygraph.Layer):
def __init__(self, class MakeFireConv(nn.Layer):
input_channels, def __init__(self,
output_channels, input_channels,
filter_size, output_channels,
padding=0, filter_size,
name=None): padding=0,
name=None):
super(MakeFireConv, self).__init__() super(MakeFireConv, self).__init__()
self._conv = Conv2D(input_channels, self._conv = Conv2d(
output_channels, input_channels,
filter_size, output_channels,
padding=padding, filter_size,
act="relu", padding=padding,
param_attr=ParamAttr(name=name + "_weights"), weight_attr=ParamAttr(name=name + "_weights"),
bias_attr=ParamAttr(name=name + "_offset")) bias_attr=ParamAttr(name=name + "_offset"))
def forward(self, x):
x = self._conv(x)
x = F.relu(x)
return x
def forward(self, inputs):
return self._conv(inputs)
class MakeFire(fluid.dygraph.Layer): class MakeFire(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
squeeze_channels, squeeze_channels,
expand1x1_channels, expand1x1_channels,
expand3x3_channels, expand3x3_channels,
name=None): name=None):
super(MakeFire, self).__init__() super(MakeFire, self).__init__()
self._conv = MakeFireConv(input_channels, self._conv = MakeFireConv(
squeeze_channels, input_channels, squeeze_channels, 1, name=name + "_squeeze1x1")
1, self._conv_path1 = MakeFireConv(
name=name + "_squeeze1x1") squeeze_channels, expand1x1_channels, 1, name=name + "_expand1x1")
self._conv_path1 = MakeFireConv(squeeze_channels, self._conv_path2 = MakeFireConv(
expand1x1_channels, squeeze_channels,
1, expand3x3_channels,
name=name + "_expand1x1") 3,
self._conv_path2 = MakeFireConv(squeeze_channels, padding=1,
expand3x3_channels, name=name + "_expand3x3")
3,
padding=1,
name=name + "_expand3x3")
def forward(self, inputs): def forward(self, inputs):
x = self._conv(inputs) x = self._conv(inputs)
x1 = self._conv_path1(x) x1 = self._conv_path1(x)
x2 = self._conv_path2(x) x2 = self._conv_path2(x)
return fluid.layers.concat([x1, x2], axis=1) return paddle.concat([x1, x2], axis=1)
class SqueezeNet(fluid.dygraph.Layer):
class SqueezeNet(nn.Layer):
def __init__(self, version, class_dim=1000): def __init__(self, version, class_dim=1000):
super(SqueezeNet, self).__init__() super(SqueezeNet, self).__init__()
self.version = version self.version = version
if self.version == "1.0": if self.version == "1.0":
self._conv = Conv2D(3, self._conv = Conv2d(
96, 3,
7, 96,
stride=2, 7,
act="relu", stride=2,
param_attr=ParamAttr(name="conv1_weights"), weight_attr=ParamAttr(name="conv1_weights"),
bias_attr=ParamAttr(name="conv1_offset")) bias_attr=ParamAttr(name="conv1_offset"))
self._pool = Pool2D(pool_size=3, self._pool = MaxPool2d(kernel_size=3, stride=2, padding=0)
pool_stride=2,
pool_type="max")
self._conv1 = MakeFire(96, 16, 64, 64, name="fire2") self._conv1 = MakeFire(96, 16, 64, 64, name="fire2")
self._conv2 = MakeFire(128, 16, 64, 64, name="fire3") self._conv2 = MakeFire(128, 16, 64, 64, name="fire3")
self._conv3 = MakeFire(128, 32, 128, 128, name="fire4") self._conv3 = MakeFire(128, 32, 128, 128, name="fire4")
...@@ -79,17 +81,15 @@ class SqueezeNet(fluid.dygraph.Layer): ...@@ -79,17 +81,15 @@ class SqueezeNet(fluid.dygraph.Layer):
self._conv8 = MakeFire(512, 64, 256, 256, name="fire9") self._conv8 = MakeFire(512, 64, 256, 256, name="fire9")
else: else:
self._conv = Conv2D(3, self._conv = Conv2d(
64, 3,
3, 64,
stride=2, 3,
padding=1, stride=2,
act="relu", padding=1,
param_attr=ParamAttr(name="conv1_weights"), weight_attr=ParamAttr(name="conv1_weights"),
bias_attr=ParamAttr(name="conv1_offset")) bias_attr=ParamAttr(name="conv1_offset"))
self._pool = Pool2D(pool_size=3, self._pool = MaxPool2d(kernel_size=3, stride=2, padding=0)
pool_stride=2,
pool_type="max")
self._conv1 = MakeFire(64, 16, 64, 64, name="fire2") self._conv1 = MakeFire(64, 16, 64, 64, name="fire2")
self._conv2 = MakeFire(128, 16, 64, 64, name="fire3") self._conv2 = MakeFire(128, 16, 64, 64, name="fire3")
...@@ -101,20 +101,20 @@ class SqueezeNet(fluid.dygraph.Layer): ...@@ -101,20 +101,20 @@ class SqueezeNet(fluid.dygraph.Layer):
self._conv7 = MakeFire(384, 64, 256, 256, name="fire8") self._conv7 = MakeFire(384, 64, 256, 256, name="fire8")
self._conv8 = MakeFire(512, 64, 256, 256, name="fire9") self._conv8 = MakeFire(512, 64, 256, 256, name="fire9")
self._drop = Dropout(p=0.5) self._drop = Dropout(p=0.5, mode="downscale_in_infer")
self._conv9 = Conv2D(512, self._conv9 = Conv2d(
class_dim, 512,
1, class_dim,
act="relu", 1,
param_attr=ParamAttr(name="conv10_weights"), weight_attr=ParamAttr(name="conv10_weights"),
bias_attr=ParamAttr(name="conv10_offset")) bias_attr=ParamAttr(name="conv10_offset"))
self._avg_pool = Pool2D(pool_type="avg", self._avg_pool = AdaptiveAvgPool2d(1)
global_pooling=True)
def forward(self, inputs): def forward(self, inputs):
x = self._conv(inputs) x = self._conv(inputs)
x = F.relu(x)
x = self._pool(x) x = self._pool(x)
if self.version=="1.0": if self.version == "1.0":
x = self._conv1(x) x = self._conv1(x)
x = self._conv2(x) x = self._conv2(x)
x = self._conv3(x) x = self._conv3(x)
...@@ -138,14 +138,17 @@ class SqueezeNet(fluid.dygraph.Layer): ...@@ -138,14 +138,17 @@ class SqueezeNet(fluid.dygraph.Layer):
x = self._conv8(x) x = self._conv8(x)
x = self._drop(x) x = self._drop(x)
x = self._conv9(x) x = self._conv9(x)
x = F.relu(x)
x = self._avg_pool(x) x = self._avg_pool(x)
x = fluid.layers.squeeze(x, axes=[2,3]) x = paddle.squeeze(x, axis=[2, 3])
return x return x
def SqueezeNet1_0(**args): def SqueezeNet1_0(**args):
model = SqueezeNet(version="1.0", **args) model = SqueezeNet(version="1.0", **args)
return model return model
def SqueezeNet1_1(**args): def SqueezeNet1_1(**args):
model = SqueezeNet(version="1.1", **args) model = SqueezeNet(version="1.1", **args)
return model return model
\ No newline at end of file
ppcls/modeling/architectures/vgg.py
浏览文件 @ ea746480
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
__all__ = ["VGG11", "VGG13", "VGG16", "VGG19"] __all__ = ["VGG11", "VGG13", "VGG16", "VGG19"]
class ConvBlock(fluid.dygraph.Layer):
def __init__(self, class ConvBlock(nn.Layer):
input_channels, def __init__(self, input_channels, output_channels, groups, name=None):
output_channels,
groups,
name=None):
super(ConvBlock, self).__init__() super(ConvBlock, self).__init__()
self.groups = groups self.groups = groups
self._conv_1 = Conv2D(num_channels=input_channels, self._conv_1 = Conv2d(
num_filters=output_channels, in_channels=input_channels,
filter_size=3, out_channels=output_channels,
stride=1, kernel_size=3,
padding=1, stride=1,
act="relu", padding=1,
param_attr=ParamAttr(name=name + "1_weights"), weight_attr=ParamAttr(name=name + "1_weights"),
bias_attr=False) bias_attr=False)
if groups == 2 or groups == 3 or groups == 4: if groups == 2 or groups == 3 or groups == 4:
self._conv_2 = Conv2D(num_channels=output_channels, self._conv_2 = Conv2d(
num_filters=output_channels, in_channels=output_channels,
filter_size=3, out_channels=output_channels,
stride=1, kernel_size=3,
padding=1, stride=1,
act="relu", padding=1,
param_attr=ParamAttr(name=name + "2_weights"), weight_attr=ParamAttr(name=name + "2_weights"),
bias_attr=False) bias_attr=False)
if groups == 3 or groups == 4: if groups == 3 or groups == 4:
self._conv_3 = Conv2D(num_channels=output_channels, self._conv_3 = Conv2d(
num_filters=output_channels, in_channels=output_channels,
filter_size=3, out_channels=output_channels,
stride=1, kernel_size=3,
padding=1, stride=1,
act="relu", padding=1,
param_attr=ParamAttr(name=name + "3_weights"), weight_attr=ParamAttr(name=name + "3_weights"),
bias_attr=False) bias_attr=False)
if groups == 4: if groups == 4:
self._conv_4 = Conv2D(num_channels=output_channels, self._conv_4 = Conv2d(
num_filters=output_channels, in_channels=output_channels,
filter_size=3, out_channels=output_channels,
stride=1, kernel_size=3,
padding=1, stride=1,
act="relu", padding=1,
param_attr=ParamAttr(name=name + "4_weights"), weight_attr=ParamAttr(name=name + "4_weights"),
bias_attr=False) bias_attr=False)
self._pool = Pool2D(pool_size=2,
pool_type="max", self._pool = MaxPool2d(kernel_size=2, stride=2, padding=0)
pool_stride=2)
def forward(self, inputs): def forward(self, inputs):
x = self._conv_1(inputs) x = self._conv_1(inputs)
x = F.relu(x)
if self.groups == 2 or self.groups == 3 or self.groups == 4: if self.groups == 2 or self.groups == 3 or self.groups == 4:
x = self._conv_2(x) x = self._conv_2(x)
if self.groups == 3 or self.groups == 4 : x = F.relu(x)
if self.groups == 3 or self.groups == 4:
x = self._conv_3(x) x = self._conv_3(x)
x = F.relu(x)
if self.groups == 4: if self.groups == 4:
x = self._conv_4(x) x = self._conv_4(x)
x = F.relu(x)
x = self._pool(x) x = self._pool(x)
return x return x
class VGGNet(fluid.dygraph.Layer):
class VGGNet(nn.Layer):
def __init__(self, layers=11, class_dim=1000): def __init__(self, layers=11, class_dim=1000):
super(VGGNet, self).__init__() super(VGGNet, self).__init__()
self.layers = layers self.layers = layers
self.vgg_configure = {11: [1, 1, 2, 2, 2], self.vgg_configure = {
13: [2, 2, 2, 2, 2], 11: [1, 1, 2, 2, 2],
16: [2, 2, 3, 3, 3], 13: [2, 2, 2, 2, 2],
19: [2, 2, 4, 4, 4]} 16: [2, 2, 3, 3, 3],
19: [2, 2, 4, 4, 4]
}
assert self.layers in self.vgg_configure.keys(), \ assert self.layers in self.vgg_configure.keys(), \
"supported layers are {} but input layer is {}".format(vgg_configure.keys(), layers) "supported layers are {} but input layer is {}".format(
vgg_configure.keys(), layers)
self.groups = self.vgg_configure[self.layers] self.groups = self.vgg_configure[self.layers]
self._conv_block_1 = ConvBlock(3, 64, self.groups[0], name="conv1_") self._conv_block_1 = ConvBlock(3, 64, self.groups[0], name="conv1_")
...@@ -83,21 +89,22 @@ class VGGNet(fluid.dygraph.Layer): ...@@ -83,21 +89,22 @@ class VGGNet(fluid.dygraph.Layer):
self._conv_block_4 = ConvBlock(256, 512, self.groups[3], name="conv4_") self._conv_block_4 = ConvBlock(256, 512, self.groups[3], name="conv4_")
self._conv_block_5 = ConvBlock(512, 512, self.groups[4], name="conv5_") self._conv_block_5 = ConvBlock(512, 512, self.groups[4], name="conv5_")
self._drop = fluid.dygraph.Dropout(p=0.5) self._drop = Dropout(p=0.5, mode="downscale_in_infer")
self._fc1 = Linear(input_dim=7*7*512, self._fc1 = Linear(
output_dim=4096, 7 * 7 * 512,
act="relu", 4096,
param_attr=ParamAttr(name="fc6_weights"), weight_attr=ParamAttr(name="fc6_weights"),
bias_attr=ParamAttr(name="fc6_offset")) bias_attr=ParamAttr(name="fc6_offset"))
self._fc2 = Linear(input_dim=4096, self._fc2 = Linear(
output_dim=4096, 4096,
act="relu", 4096,
param_attr=ParamAttr(name="fc7_weights"), weight_attr=ParamAttr(name="fc7_weights"),
bias_attr=ParamAttr(name="fc7_offset")) bias_attr=ParamAttr(name="fc7_offset"))
self._out = Linear(input_dim=4096, self._out = Linear(
output_dim=class_dim, 4096,
param_attr=ParamAttr(name="fc8_weights"), class_dim,
bias_attr=ParamAttr(name="fc8_offset")) weight_attr=ParamAttr(name="fc8_weights"),
bias_attr=ParamAttr(name="fc8_offset"))
def forward(self, inputs): def forward(self, inputs):
x = self._conv_block_1(inputs) x = self._conv_block_1(inputs)
...@@ -106,26 +113,32 @@ class VGGNet(fluid.dygraph.Layer): ...@@ -106,26 +113,32 @@ class VGGNet(fluid.dygraph.Layer):
x = self._conv_block_4(x) x = self._conv_block_4(x)
x = self._conv_block_5(x) x = self._conv_block_5(x)
x = fluid.layers.reshape(x, [0,-1]) x = paddle.reshape(x, [0, -1])
x = self._fc1(x) x = self._fc1(x)
x = F.relu(x)
x = self._drop(x) x = self._drop(x)
x = self._fc2(x) x = self._fc2(x)
x = F.relu(x)
x = self._drop(x) x = self._drop(x)
x = self._out(x) x = self._out(x)
return x return x
def VGG11(**args): def VGG11(**args):
model = VGGNet(layers=11, **args) model = VGGNet(layers=11, **args)
return model return model
def VGG13(**args): def VGG13(**args):
model = VGGNet(layers=13, **args) model = VGGNet(layers=13, **args)
return model return model
def VGG16(**args): def VGG16(**args):
model = VGGNet(layers=16, **args) model = VGGNet(layers=16, **args)
return model return model
def VGG19(**args): def VGG19(**args):
model = VGGNet(layers=19, **args) model = VGGNet(layers=19, **args)
return model return model
\ No newline at end of file
ppcls/modeling/architectures/xception.py
浏览文件 @ ea746480
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
from paddle.nn.initializer import Uniform
import math import math
__all__ = ['Xception41', 'Xception65', 'Xception71'] __all__ = ['Xception41', 'Xception65', 'Xception71']
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__(self, def __init__(self,
num_channels, num_channels,
num_filters, num_filters,
...@@ -18,15 +21,14 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -18,15 +21,14 @@ class ConvBNLayer(fluid.dygraph.Layer):
name=None): name=None):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self._conv = Conv2D( self._conv = Conv2d(
num_channels=num_channels, in_channels=num_channels,
num_filters=num_filters, out_channels=num_filters,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=groups, groups=groups,
act=None, weight_attr=ParamAttr(name=name + "_weights"),
param_attr=ParamAttr(name=name + "_weights"),
bias_attr=False) bias_attr=False)
bn_name = "bn_" + name bn_name = "bn_" + name
self._batch_norm = BatchNorm( self._batch_norm = BatchNorm(
...@@ -43,7 +45,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -43,7 +45,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return y return y
class SeparableConv(fluid.dygraph.Layer): class SeparableConv(nn.Layer):
def __init__(self, input_channels, output_channels, stride=1, name=None): def __init__(self, input_channels, output_channels, stride=1, name=None):
super(SeparableConv, self).__init__() super(SeparableConv, self).__init__()
...@@ -63,7 +65,7 @@ class SeparableConv(fluid.dygraph.Layer): ...@@ -63,7 +65,7 @@ class SeparableConv(fluid.dygraph.Layer):
return x return x
class EntryFlowBottleneckBlock(fluid.dygraph.Layer): class EntryFlowBottleneckBlock(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
output_channels, output_channels,
...@@ -73,14 +75,13 @@ class EntryFlowBottleneckBlock(fluid.dygraph.Layer): ...@@ -73,14 +75,13 @@ class EntryFlowBottleneckBlock(fluid.dygraph.Layer):
super(EntryFlowBottleneckBlock, self).__init__() super(EntryFlowBottleneckBlock, self).__init__()
self.relu_first = relu_first self.relu_first = relu_first
self._short = Conv2D( self._short = Conv2d(
num_channels=input_channels, in_channels=input_channels,
num_filters=output_channels, out_channels=output_channels,
filter_size=1, kernel_size=1,
stride=stride, stride=stride,
padding=0, padding=0,
act=None, weight_attr=ParamAttr(name + "_branch1_weights"),
param_attr=ParamAttr(name + "_branch1_weights"),
bias_attr=False) bias_attr=False)
self._conv1 = SeparableConv( self._conv1 = SeparableConv(
input_channels, input_channels,
...@@ -92,22 +93,21 @@ class EntryFlowBottleneckBlock(fluid.dygraph.Layer): ...@@ -92,22 +93,21 @@ class EntryFlowBottleneckBlock(fluid.dygraph.Layer):
output_channels, output_channels,
stride=1, stride=1,
name=name + "_branch2b_weights") name=name + "_branch2b_weights")
self._pool = Pool2D( self._pool = MaxPool2d(kernel_size=3, stride=stride, padding=1)
pool_size=3, pool_stride=stride, pool_padding=1, pool_type="max")
def forward(self, inputs): def forward(self, inputs):
conv0 = inputs conv0 = inputs
short = self._short(inputs) short = self._short(inputs)
if self.relu_first: if self.relu_first:
conv0 = fluid.layers.relu(conv0) conv0 = F.relu(conv0)
conv1 = self._conv1(conv0) conv1 = self._conv1(conv0)
conv2 = fluid.layers.relu(conv1) conv2 = F.relu(conv1)
conv2 = self._conv2(conv2) conv2 = self._conv2(conv2)
pool = self._pool(conv2) pool = self._pool(conv2)
return fluid.layers.elementwise_add(x=short, y=pool) return paddle.elementwise_add(x=short, y=pool)
class EntryFlow(fluid.dygraph.Layer): class EntryFlow(nn.Layer):
def __init__(self, block_num=3): def __init__(self, block_num=3):
super(EntryFlow, self).__init__() super(EntryFlow, self).__init__()
...@@ -154,7 +154,7 @@ class EntryFlow(fluid.dygraph.Layer): ...@@ -154,7 +154,7 @@ class EntryFlow(fluid.dygraph.Layer):
return x return x
class MiddleFlowBottleneckBlock(fluid.dygraph.Layer): class MiddleFlowBottleneckBlock(nn.Layer):
def __init__(self, input_channels, output_channels, name): def __init__(self, input_channels, output_channels, name):
super(MiddleFlowBottleneckBlock, self).__init__() super(MiddleFlowBottleneckBlock, self).__init__()
...@@ -175,16 +175,16 @@ class MiddleFlowBottleneckBlock(fluid.dygraph.Layer): ...@@ -175,16 +175,16 @@ class MiddleFlowBottleneckBlock(fluid.dygraph.Layer):
name=name + "_branch2c_weights") name=name + "_branch2c_weights")
def forward(self, inputs): def forward(self, inputs):
conv0 = fluid.layers.relu(inputs) conv0 = F.relu(inputs)
conv0 = self._conv_0(conv0) conv0 = self._conv_0(conv0)
conv1 = fluid.layers.relu(conv0) conv1 = F.relu(conv0)
conv1 = self._conv_1(conv1) conv1 = self._conv_1(conv1)
conv2 = fluid.layers.relu(conv1) conv2 = F.relu(conv1)
conv2 = self._conv_2(conv2) conv2 = self._conv_2(conv2)
return fluid.layers.elementwise_add(x=inputs, y=conv2) return paddle.elementwise_add(x=inputs, y=conv2)
class MiddleFlow(fluid.dygraph.Layer): class MiddleFlow(nn.Layer):
def __init__(self, block_num=8): def __init__(self, block_num=8):
super(MiddleFlow, self).__init__() super(MiddleFlow, self).__init__()
...@@ -244,19 +244,18 @@ class MiddleFlow(fluid.dygraph.Layer): ...@@ -244,19 +244,18 @@ class MiddleFlow(fluid.dygraph.Layer):
return x return x
class ExitFlowBottleneckBlock(fluid.dygraph.Layer): class ExitFlowBottleneckBlock(nn.Layer):
def __init__(self, input_channels, output_channels1, output_channels2, def __init__(self, input_channels, output_channels1, output_channels2,
name): name):
super(ExitFlowBottleneckBlock, self).__init__() super(ExitFlowBottleneckBlock, self).__init__()
self._short = Conv2D( self._short = Conv2d(
num_channels=input_channels, in_channels=input_channels,
num_filters=output_channels2, out_channels=output_channels2,
filter_size=1, kernel_size=1,
stride=2, stride=2,
padding=0, padding=0,
act=None, weight_attr=ParamAttr(name + "_branch1_weights"),
param_attr=ParamAttr(name + "_branch1_weights"),
bias_attr=False) bias_attr=False)
self._conv_1 = SeparableConv( self._conv_1 = SeparableConv(
input_channels, input_channels,
...@@ -268,20 +267,19 @@ class ExitFlowBottleneckBlock(fluid.dygraph.Layer): ...@@ -268,20 +267,19 @@ class ExitFlowBottleneckBlock(fluid.dygraph.Layer):
output_channels2, output_channels2,
stride=1, stride=1,
name=name + "_branch2b_weights") name=name + "_branch2b_weights")
self._pool = Pool2D( self._pool = MaxPool2d(kernel_size=3, stride=2, padding=1)
pool_size=3, pool_stride=2, pool_padding=1, pool_type="max")
def forward(self, inputs): def forward(self, inputs):
short = self._short(inputs) short = self._short(inputs)
conv0 = fluid.layers.relu(inputs) conv0 = F.relu(inputs)
conv1 = self._conv_1(conv0) conv1 = self._conv_1(conv0)
conv2 = fluid.layers.relu(conv1) conv2 = F.relu(conv1)
conv2 = self._conv_2(conv2) conv2 = self._conv_2(conv2)
pool = self._pool(conv2) pool = self._pool(conv2)
return fluid.layers.elementwise_add(x=short, y=pool) return paddle.elementwise_add(x=short, y=pool)
class ExitFlow(fluid.dygraph.Layer): class ExitFlow(nn.Layer):
def __init__(self, class_dim): def __init__(self, class_dim):
super(ExitFlow, self).__init__() super(ExitFlow, self).__init__()
...@@ -291,29 +289,28 @@ class ExitFlow(fluid.dygraph.Layer): ...@@ -291,29 +289,28 @@ class ExitFlow(fluid.dygraph.Layer):
728, 728, 1024, name=name + "_1") 728, 728, 1024, name=name + "_1")
self._conv_1 = SeparableConv(1024, 1536, stride=1, name=name + "_2") self._conv_1 = SeparableConv(1024, 1536, stride=1, name=name + "_2")
self._conv_2 = SeparableConv(1536, 2048, stride=1, name=name + "_3") self._conv_2 = SeparableConv(1536, 2048, stride=1, name=name + "_3")
self._pool = Pool2D(pool_type="avg", global_pooling=True) self._pool = AdaptiveAvgPool2d(1)
stdv = 1.0 / math.sqrt(2048 * 1.0) stdv = 1.0 / math.sqrt(2048 * 1.0)
self._out = Linear( self._out = Linear(
2048, 2048,
class_dim, class_dim,
param_attr=ParamAttr( weight_attr=ParamAttr(
name="fc_weights", name="fc_weights", initializer=Uniform(-stdv, stdv)),
initializer=fluid.initializer.Uniform(-stdv, stdv)),
bias_attr=ParamAttr(name="fc_offset")) bias_attr=ParamAttr(name="fc_offset"))
def forward(self, inputs): def forward(self, inputs):
conv0 = self._conv_0(inputs) conv0 = self._conv_0(inputs)
conv1 = self._conv_1(conv0) conv1 = self._conv_1(conv0)
conv1 = fluid.layers.relu(conv1) conv1 = F.relu(conv1)
conv2 = self._conv_2(conv1) conv2 = self._conv_2(conv1)
conv2 = fluid.layers.relu(conv2) conv2 = F.relu(conv2)
pool = self._pool(conv2) pool = self._pool(conv2)
pool = fluid.layers.reshape(pool, [0, -1]) pool = paddle.reshape(pool, [0, -1])
out = self._out(pool) out = self._out(pool)
return out return out
class Xception(fluid.dygraph.Layer): class Xception(nn.Layer):
def __init__(self, def __init__(self,
entry_flow_block_num=3, entry_flow_block_num=3,
middle_flow_block_num=8, middle_flow_block_num=8,
...@@ -344,4 +341,4 @@ def Xception65(**args): ...@@ -344,4 +341,4 @@ def Xception65(**args):
def Xception71(**args): def Xception71(**args):
model = Xception(entry_flow_block_num=5, middle_flow_block_num=16, **args) model = Xception(entry_flow_block_num=5, middle_flow_block_num=16, **args)
return model return model
\ No newline at end of file
ppcls/modeling/architectures/xception_deeplab.py
浏览文件 @ ea746480
import paddle import paddle
import paddle.fluid as fluid from paddle import ParamAttr
from paddle.fluid.param_attr import ParamAttr import paddle.nn as nn
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear, Dropout import paddle.nn.functional as F
from paddle.nn import Conv2d, BatchNorm, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2d, MaxPool2d, AvgPool2d
__all__ = ["Xception41_deeplab", "Xception65_deeplab", "Xception71_deeplab"] __all__ = ["Xception41_deeplab", "Xception65_deeplab", "Xception71_deeplab"]
...@@ -56,7 +58,7 @@ def gen_bottleneck_params(backbone='xception_65'): ...@@ -56,7 +58,7 @@ def gen_bottleneck_params(backbone='xception_65'):
return bottleneck_params return bottleneck_params
class ConvBNLayer(fluid.dygraph.Layer): class ConvBNLayer(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
output_channels, output_channels,
...@@ -67,13 +69,13 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -67,13 +69,13 @@ class ConvBNLayer(fluid.dygraph.Layer):
name=None): name=None):
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
self._conv = Conv2D( self._conv = Conv2d(
num_channels=input_channels, in_channels=input_channels,
num_filters=output_channels, out_channels=output_channels,
filter_size=filter_size, kernel_size=filter_size,
stride=stride, stride=stride,
padding=padding, padding=padding,
param_attr=ParamAttr(name=name + "/weights"), weight_attr=ParamAttr(name=name + "/weights"),
bias_attr=False) bias_attr=False)
self._bn = BatchNorm( self._bn = BatchNorm(
num_channels=output_channels, num_channels=output_channels,
...@@ -89,7 +91,7 @@ class ConvBNLayer(fluid.dygraph.Layer): ...@@ -89,7 +91,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
return self._bn(self._conv(inputs)) return self._bn(self._conv(inputs))
class Seperate_Conv(fluid.dygraph.Layer): class Seperate_Conv(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
output_channels, output_channels,
...@@ -100,15 +102,15 @@ class Seperate_Conv(fluid.dygraph.Layer): ...@@ -100,15 +102,15 @@ class Seperate_Conv(fluid.dygraph.Layer):
name=None): name=None):
super(Seperate_Conv, self).__init__() super(Seperate_Conv, self).__init__()
self._conv1 = Conv2D( self._conv1 = Conv2d(
num_channels=input_channels, in_channels=input_channels,
num_filters=input_channels, out_channels=input_channels,
filter_size=filter, kernel_size=filter,
stride=stride, stride=stride,
groups=input_channels, groups=input_channels,
padding=(filter) // 2 * dilation, padding=(filter) // 2 * dilation,
dilation=dilation, dilation=dilation,
param_attr=ParamAttr(name=name + "/depthwise/weights"), weight_attr=ParamAttr(name=name + "/depthwise/weights"),
bias_attr=False) bias_attr=False)
self._bn1 = BatchNorm( self._bn1 = BatchNorm(
input_channels, input_channels,
...@@ -119,14 +121,14 @@ class Seperate_Conv(fluid.dygraph.Layer): ...@@ -119,14 +121,14 @@ class Seperate_Conv(fluid.dygraph.Layer):
bias_attr=ParamAttr(name=name + "/depthwise/BatchNorm/beta"), bias_attr=ParamAttr(name=name + "/depthwise/BatchNorm/beta"),
moving_mean_name=name + "/depthwise/BatchNorm/moving_mean", moving_mean_name=name + "/depthwise/BatchNorm/moving_mean",
moving_variance_name=name + "/depthwise/BatchNorm/moving_variance") moving_variance_name=name + "/depthwise/BatchNorm/moving_variance")
self._conv2 = Conv2D( self._conv2 = Conv2d(
input_channels, input_channels,
output_channels, output_channels,
1, 1,
stride=1, stride=1,
groups=1, groups=1,
padding=0, padding=0,
param_attr=ParamAttr(name=name + "/pointwise/weights"), weight_attr=ParamAttr(name=name + "/pointwise/weights"),
bias_attr=False) bias_attr=False)
self._bn2 = BatchNorm( self._bn2 = BatchNorm(
output_channels, output_channels,
...@@ -146,7 +148,7 @@ class Seperate_Conv(fluid.dygraph.Layer): ...@@ -146,7 +148,7 @@ class Seperate_Conv(fluid.dygraph.Layer):
return x return x
class Xception_Block(fluid.dygraph.Layer): class Xception_Block(nn.Layer):
def __init__(self, def __init__(self,
input_channels, input_channels,
output_channels, output_channels,
...@@ -226,11 +228,11 @@ class Xception_Block(fluid.dygraph.Layer): ...@@ -226,11 +228,11 @@ class Xception_Block(fluid.dygraph.Layer):
def forward(self, inputs): def forward(self, inputs):
if not self.activation_fn_in_separable_conv: if not self.activation_fn_in_separable_conv:
x = fluid.layers.relu(inputs) x = F.relu(inputs)
x = self._conv1(x) x = self._conv1(x)
x = fluid.layers.relu(x) x = F.relu(x)
x = self._conv2(x) x = self._conv2(x)
x = fluid.layers.relu(x) x = F.relu(x)
x = self._conv3(x) x = self._conv3(x)
else: else:
x = self._conv1(inputs) x = self._conv1(inputs)
...@@ -242,10 +244,10 @@ class Xception_Block(fluid.dygraph.Layer): ...@@ -242,10 +244,10 @@ class Xception_Block(fluid.dygraph.Layer):
skip = self._short(inputs) skip = self._short(inputs)
else: else:
skip = inputs skip = inputs
return fluid.layers.elementwise_add(x, skip) return paddle.elementwise_add(x, skip)
class XceptionDeeplab(fluid.dygraph.Layer): class XceptionDeeplab(nn.Layer):
def __init__(self, backbone, class_dim=1000): def __init__(self, backbone, class_dim=1000):
super(XceptionDeeplab, self).__init__() super(XceptionDeeplab, self).__init__()
...@@ -344,12 +346,12 @@ class XceptionDeeplab(fluid.dygraph.Layer): ...@@ -344,12 +346,12 @@ class XceptionDeeplab(fluid.dygraph.Layer):
self.stride = s self.stride = s
self._drop = Dropout(p=0.5) self._drop = Dropout(p=0.5, mode="downscale_in_infer")
self._pool = Pool2D(pool_type="avg", global_pooling=True) self._pool = AdaptiveAvgPool2d(1)
self._fc = Linear( self._fc = Linear(
self.chns[1][-1], self.chns[1][-1],
class_dim, class_dim,
param_attr=ParamAttr(name="fc_weights"), weight_attr=ParamAttr(name="fc_weights"),
bias_attr=ParamAttr(name="fc_bias")) bias_attr=ParamAttr(name="fc_bias"))
def forward(self, inputs): def forward(self, inputs):
...@@ -363,7 +365,7 @@ class XceptionDeeplab(fluid.dygraph.Layer): ...@@ -363,7 +365,7 @@ class XceptionDeeplab(fluid.dygraph.Layer):
x = self._exit_flow_2(x) x = self._exit_flow_2(x)
x = self._drop(x) x = self._drop(x)
x = self._pool(x) x = self._pool(x)
x = fluid.layers.squeeze(x, axes=[2, 3]) x = paddle.squeeze(x, axis=[2, 3])
x = self._fc(x) x = self._fc(x)
return x return x
...@@ -380,4 +382,4 @@ def Xception65_deeplab(**args): ...@@ -380,4 +382,4 @@ def Xception65_deeplab(**args):
def Xception71_deeplab(**args): def Xception71_deeplab(**args):
model = XceptionDeeplab("xception_71", **args) model = XceptionDeeplab("xception_71", **args)
return model return model
\ No newline at end of file
ppcls/modeling/loss.py
浏览文件 @ ea746480
...@@ -12,7 +12,8 @@ ...@@ -12,7 +12,8 @@
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
import paddle.fluid as fluid import paddle
import paddle.nn.functional as F
__all__ = ['CELoss', 'MixCELoss', 'GoogLeNetLoss', 'JSDivLoss'] __all__ = ['CELoss', 'MixCELoss', 'GoogLeNetLoss', 'JSDivLoss']
...@@ -34,35 +35,37 @@ class Loss(object): ...@@ -34,35 +35,37 @@ class Loss(object):
def _labelsmoothing(self, target): def _labelsmoothing(self, target):
if target.shape[-1] != self._class_dim: if target.shape[-1] != self._class_dim:
one_hot_target = fluid.one_hot(input=target, depth=self._class_dim) one_hot_target = F.one_hot(target, self._class_dim)
else: else:
one_hot_target = target one_hot_target = target
soft_target = fluid.layers.label_smooth( soft_target = F.label_smooth(
label=one_hot_target, epsilon=self._epsilon, dtype="float32") one_hot_target, epsilon=self._epsilon, dtype="float32")
soft_target = fluid.layers.reshape( soft_target = paddle.reshape(soft_target, shape=[-1, self._class_dim])
soft_target, shape=[-1, self._class_dim])
return soft_target return soft_target
def _crossentropy(self, input, target): def _crossentropy(self, input, target):
if self._label_smoothing: if self._label_smoothing:
target = self._labelsmoothing(target) target = self._labelsmoothing(target)
softmax_out = fluid.layers.softmax(input, use_cudnn=False) input = -F.log_softmax(input, axis=-1)
cost = fluid.layers.cross_entropy( log_probs = -F.log_softmax(input, axis=-1)
input=softmax_out, label=target, soft_label=self._label_smoothing) cost = paddle.reduce_sum(target * log_probs, dim=-1)
avg_cost = fluid.layers.mean(cost) else:
# softmax_out = F.softmax(input)
cost = F.cross_entropy(input=input, label=target)
avg_cost = paddle.mean(cost)
return avg_cost return avg_cost
def _kldiv(self, input, target): def _kldiv(self, input, target):
cost = target * fluid.layers.log(target / input) * self._class_dim cost = target * F.log(target / input) * self._class_dim
cost = fluid.layers.sum(cost) cost = paddle.sum(cost)
return cost return cost
def _jsdiv(self, input, target): def _jsdiv(self, input, target):
input = fluid.layers.softmax(input, use_cudnn=False) input = F.softmax(input)
target = fluid.layers.softmax(target, use_cudnn=False) target = F.softmax(target)
cost = self._kldiv(input, target) + self._kldiv(target, input) cost = self._kldiv(input, target) + self._kldiv(target, input)
cost = cost / 2 cost = cost / 2
avg_cost = fluid.layers.mean(cost) avg_cost = paddle.mean(cost)
return avg_cost return avg_cost
def __call__(self, input, target): def __call__(self, input, target):
...@@ -94,7 +97,7 @@ class MixCELoss(Loss): ...@@ -94,7 +97,7 @@ class MixCELoss(Loss):
cost0 = self._crossentropy(input, target0) cost0 = self._crossentropy(input, target0)
cost1 = self._crossentropy(input, target1) cost1 = self._crossentropy(input, target1)
cost = lam * cost0 + (1.0 - lam) * cost1 cost = lam * cost0 + (1.0 - lam) * cost1
avg_cost = fluid.layers.mean(cost) avg_cost = paddle.mean(cost)
return avg_cost return avg_cost
...@@ -111,7 +114,7 @@ class GoogLeNetLoss(Loss): ...@@ -111,7 +114,7 @@ class GoogLeNetLoss(Loss):
cost1 = self._crossentropy(input1, target) cost1 = self._crossentropy(input1, target)
cost2 = self._crossentropy(input2, target) cost2 = self._crossentropy(input2, target)
cost = cost0 + 0.3 * cost1 + 0.3 * cost2 cost = cost0 + 0.3 * cost1 + 0.3 * cost2
avg_cost = fluid.layers.mean(cost) avg_cost = paddle.mean(cost)
return avg_cost return avg_cost
......
ppcls/optimizer/learning_rate.py
浏览文件 @ ea746480
...@@ -19,36 +19,15 @@ from __future__ import print_function ...@@ -19,36 +19,15 @@ from __future__ import print_function
import sys import sys
import math import math
import paddle.fluid as fluid from paddle.optimizer.lr_scheduler import LinearLrWarmup
import paddle.fluid.layers.ops as ops from paddle.optimizer.lr_scheduler import PiecewiseLR
from paddle.fluid.layers.learning_rate_scheduler import _decay_step_counter from paddle.optimizer.lr_scheduler import CosineAnnealingLR
from paddle.optimizer.lr_scheduler import ExponentialLR
__all__ = ['LearningRateBuilder'] __all__ = ['LearningRateBuilder']
class Linear(object): class Cosine(CosineAnnealingLR):
"""
Linear learning rate decay
Args:
lr(float): initial learning rate
steps(int): total decay steps
end_lr(float): end learning rate, default: 0.0.
"""
def __init__(self, lr, steps, end_lr=0.0, **kwargs):
super(Linear, self).__init__()
self.lr = lr
self.steps = steps
self.end_lr = end_lr
def __call__(self):
learning_rate = fluid.layers.polynomial_decay(
self.lr, self.steps, self.end_lr, power=1)
return learning_rate
class Cosine(object):
""" """
Cosine learning rate decay Cosine learning rate decay
lr = 0.05 * (math.cos(epoch * (math.pi / epochs)) + 1) lr = 0.05 * (math.cos(epoch * (math.pi / epochs)) + 1)
...@@ -60,20 +39,14 @@ class Cosine(object): ...@@ -60,20 +39,14 @@ class Cosine(object):
""" """
def __init__(self, lr, step_each_epoch, epochs, **kwargs): def __init__(self, lr, step_each_epoch, epochs, **kwargs):
super(Cosine, self).__init__() super(Cosine, self).__init__(
self.lr = lr learning_rate=lr,
self.step_each_epoch = step_each_epoch T_max=step_each_epoch * epochs, )
self.epochs = epochs
def __call__(self): self.update_specified = False
learning_rate = fluid.layers.cosine_decay(
learning_rate=self.lr,
step_each_epoch=self.step_each_epoch,
epochs=self.epochs)
return learning_rate
class Piecewise(object): class Piecewise(PiecewiseLR):
""" """
Piecewise learning rate decay Piecewise learning rate decay
...@@ -85,16 +58,15 @@ class Piecewise(object): ...@@ -85,16 +58,15 @@ class Piecewise(object):
""" """
def __init__(self, lr, step_each_epoch, decay_epochs, gamma=0.1, **kwargs): def __init__(self, lr, step_each_epoch, decay_epochs, gamma=0.1, **kwargs):
super(Piecewise, self).__init__() boundaries = [step_each_epoch * e for e in decay_epochs]
self.bd = [step_each_epoch * e for e in decay_epochs] lr_values = [lr * (gamma**i) for i in range(len(boundaries) + 1)]
self.lr = [lr * (gamma**i) for i in range(len(self.bd) + 1)] super(Piecewise, self).__init__(
boundaries=boundaries, values=lr_values)
def __call__(self): self.update_specified = False
learning_rate = fluid.layers.piecewise_decay(self.bd, self.lr)
return learning_rate
class CosineWarmup(object): class CosineWarmup(LinearLrWarmup):
""" """
Cosine learning rate decay with warmup Cosine learning rate decay with warmup
[0, warmup_epoch): linear warmup [0, warmup_epoch): linear warmup
...@@ -108,28 +80,23 @@ class CosineWarmup(object): ...@@ -108,28 +80,23 @@ class CosineWarmup(object):
""" """
def __init__(self, lr, step_each_epoch, epochs, warmup_epoch=5, **kwargs): def __init__(self, lr, step_each_epoch, epochs, warmup_epoch=5, **kwargs):
super(CosineWarmup, self).__init__() assert epochs > warmup_epoch, "total epoch({}) should be larger than warmup_epoch({}) in CosineWarmup.".format(
self.lr = lr epochs, warmup_epoch)
self.step_each_epoch = step_each_epoch warmup_step = warmup_epoch * step_each_epoch
self.epochs = epochs start_lr = 0.0
self.warmup_epoch = warmup_epoch end_lr = lr
lr_sch = Cosine(lr, step_each_epoch, epochs - warmup_epoch)
def __call__(self):
learning_rate = fluid.layers.cosine_decay(
learning_rate=self.lr,
step_each_epoch=self.step_each_epoch,
epochs=self.epochs)
learning_rate = fluid.layers.linear_lr_warmup( super(CosineWarmup, self).__init__(
learning_rate, learning_rate=lr_sch,
warmup_steps=self.warmup_epoch * self.step_each_epoch, warmup_steps=warmup_step,
start_lr=0.0, start_lr=start_lr,
end_lr=self.lr) end_lr=end_lr)
return learning_rate self.update_specified = False
class ExponentialWarmup(object): class ExponentialWarmup(LinearLrWarmup):
""" """
Exponential learning rate decay with warmup Exponential learning rate decay with warmup
[0, warmup_epoch): linear warmup [0, warmup_epoch): linear warmup
...@@ -150,27 +117,22 @@ class ExponentialWarmup(object): ...@@ -150,27 +117,22 @@ class ExponentialWarmup(object):
decay_rate=0.97, decay_rate=0.97,
warmup_epoch=5, warmup_epoch=5,
**kwargs): **kwargs):
super(ExponentialWarmup, self).__init__() warmup_step = warmup_epoch * step_each_epoch
self.lr = lr start_lr = 0.0
end_lr = lr
lr_sch = ExponentialLR(lr, decay_rate)
super(ExponentialWarmup, self).__init__(
learning_rate=lr_sch,
warmup_steps=warmup_step,
start_lr=start_lr,
end_lr=end_lr)
# NOTE: hac method to update exponential lr scheduler
self.update_specified = True
self.update_start_step = warmup_step
self.update_step_interval = int(decay_epochs * step_each_epoch)
self.step_each_epoch = step_each_epoch self.step_each_epoch = step_each_epoch
self.decay_epochs = decay_epochs
self.decay_rate = decay_rate
self.warmup_epoch = warmup_epoch
def __call__(self):
learning_rate = fluid.layers.exponential_decay(
learning_rate=self.lr,
decay_steps=self.decay_epochs * self.step_each_epoch,
decay_rate=self.decay_rate,
staircase=False)
learning_rate = fluid.layers.linear_lr_warmup(
learning_rate,
warmup_steps=self.warmup_epoch * self.step_each_epoch,
start_lr=0.0,
end_lr=self.lr)
return learning_rate
class LearningRateBuilder(): class LearningRateBuilder():
...@@ -193,5 +155,5 @@ class LearningRateBuilder(): ...@@ -193,5 +155,5 @@ class LearningRateBuilder():
def __call__(self): def __call__(self):
mod = sys.modules[__name__] mod = sys.modules[__name__]
lr = getattr(mod, self.function)(**self.params)() lr = getattr(mod, self.function)(**self.params)
return lr return lr
ppcls/optimizer/optimizer.py
浏览文件 @ ea746480
...@@ -18,7 +18,7 @@ from __future__ import print_function ...@@ -18,7 +18,7 @@ from __future__ import print_function
import sys import sys
import paddle.fluid as fluid import paddle
__all__ = ['OptimizerBuilder'] __all__ = ['OptimizerBuilder']
...@@ -33,11 +33,10 @@ class L1Decay(object): ...@@ -33,11 +33,10 @@ class L1Decay(object):
def __init__(self, factor=0.0): def __init__(self, factor=0.0):
super(L1Decay, self).__init__() super(L1Decay, self).__init__()
self.regularization_coeff = factor self.factor = factor
def __call__(self): def __call__(self):
reg = fluid.regularizer.L1Decay( reg = paddle.regularizer.L1Decay(self.factor)
regularization_coeff=self.regularization_coeff)
return reg return reg
...@@ -51,11 +50,10 @@ class L2Decay(object): ...@@ -51,11 +50,10 @@ class L2Decay(object):
def __init__(self, factor=0.0): def __init__(self, factor=0.0):
super(L2Decay, self).__init__() super(L2Decay, self).__init__()
self.regularization_coeff = factor self.factor = factor
def __call__(self): def __call__(self):
reg = fluid.regularizer.L2Decay( reg = paddle.regularizer.L2Decay(self.factor)
regularization_coeff=self.regularization_coeff)
return reg return reg
...@@ -83,11 +81,11 @@ class Momentum(object): ...@@ -83,11 +81,11 @@ class Momentum(object):
self.regularization = regularization self.regularization = regularization
def __call__(self): def __call__(self):
opt = fluid.optimizer.Momentum( opt = paddle.optimizer.Momentum(
learning_rate=self.learning_rate, learning_rate=self.learning_rate,
momentum=self.momentum, momentum=self.momentum,
parameter_list=self.parameter_list, parameters=self.parameter_list,
regularization=self.regularization) weight_decay=self.regularization)
return opt return opt
...@@ -121,13 +119,13 @@ class RMSProp(object): ...@@ -121,13 +119,13 @@ class RMSProp(object):
self.regularization = regularization self.regularization = regularization
def __call__(self): def __call__(self):
opt = fluid.optimizer.RMSProp( opt = paddle.optimizer.RMSProp(
learning_rate=self.learning_rate, learning_rate=self.learning_rate,
momentum=self.momentum, momentum=self.momentum,
rho=self.rho, rho=self.rho,
epsilon=self.epsilon, epsilon=self.epsilon,
parameter_list=self.parameter_list, parameters=self.parameter_list,
regularization=self.regularization) weight_decay=self.regularization)
return opt return opt
......
ppcls/utils/check.py
浏览文件 @ ea746480
...@@ -19,7 +19,9 @@ from __future__ import print_function ...@@ -19,7 +19,9 @@ from __future__ import print_function
import os import os
import sys import sys
import paddle.fluid as fluid import paddle
# TODO: need to be fixed in the future.
from paddle.fluid import is_compiled_with_cuda
from ppcls.modeling import get_architectures from ppcls.modeling import get_architectures
from ppcls.modeling import similar_architectures from ppcls.modeling import similar_architectures
...@@ -33,10 +35,9 @@ def check_version(): ...@@ -33,10 +35,9 @@ def check_version():
""" """
err = "PaddlePaddle version 1.8.0 or higher is required, " \ err = "PaddlePaddle version 1.8.0 or higher is required, " \
"or a suitable develop version is satisfied as well. \n" \ "or a suitable develop version is satisfied as well. \n" \
"Please make sure the version is good with your code." \ "Please make sure the version is good with your code."
try: try:
fluid.require_version('1.8.0') paddle.utils.require_version('0.0.0')
except Exception: except Exception:
logger.error(err) logger.error(err)
sys.exit(1) sys.exit(1)
...@@ -50,7 +51,7 @@ def check_gpu(): ...@@ -50,7 +51,7 @@ def check_gpu():
"install paddlepaddle-gpu to run model on GPU." "install paddlepaddle-gpu to run model on GPU."
try: try:
assert fluid.is_compiled_with_cuda() assert is_compiled_with_cuda()
except AssertionError: except AssertionError:
logger.error(err) logger.error(err)
sys.exit(1) sys.exit(1)
......
ppcls/utils/logger.py
浏览文件 @ ea746480
...@@ -23,7 +23,7 @@ logging.basicConfig( ...@@ -23,7 +23,7 @@ logging.basicConfig(
def time_zone(sec, fmt): def time_zone(sec, fmt):
real_time = datetime.datetime.now() + datetime.timedelta(hours=8) real_time = datetime.datetime.now()
return real_time.timetuple() return real_time.timetuple()
......
ppcls/utils/save_load.py
浏览文件 @ ea746480
...@@ -22,7 +22,8 @@ import re ...@@ -22,7 +22,8 @@ import re
import shutil import shutil
import tempfile import tempfile
import paddle.fluid as fluid import paddle
from paddle.io import load_program_state
from ppcls.utils import logger from ppcls.utils import logger
...@@ -50,7 +51,7 @@ def load_dygraph_pretrain(model, path=None, load_static_weights=False): ...@@ -50,7 +51,7 @@ def load_dygraph_pretrain(model, path=None, load_static_weights=False):
raise ValueError("Model pretrain path {} does not " raise ValueError("Model pretrain path {} does not "
"exists.".format(path)) "exists.".format(path))
if load_static_weights: if load_static_weights:
pre_state_dict = fluid.load_program_state(path) pre_state_dict = load_program_state(path)
param_state_dict = {} param_state_dict = {}
model_dict = model.state_dict() model_dict = model.state_dict()
for key in model_dict.keys(): for key in model_dict.keys():
...@@ -64,7 +65,7 @@ def load_dygraph_pretrain(model, path=None, load_static_weights=False): ...@@ -64,7 +65,7 @@ def load_dygraph_pretrain(model, path=None, load_static_weights=False):
model.set_dict(param_state_dict) model.set_dict(param_state_dict)
return return
param_state_dict, optim_state_dict = fluid.load_dygraph(path) param_state_dict, optim_state_dict = paddle.load(path)
model.set_dict(param_state_dict) model.set_dict(param_state_dict)
return return
...@@ -105,7 +106,7 @@ def init_model(config, net, optimizer=None): ...@@ -105,7 +106,7 @@ def init_model(config, net, optimizer=None):
"Given dir {}.pdparams not exist.".format(checkpoints) "Given dir {}.pdparams not exist.".format(checkpoints)
assert os.path.exists(checkpoints + ".pdopt"), \ assert os.path.exists(checkpoints + ".pdopt"), \
"Given dir {}.pdopt not exist.".format(checkpoints) "Given dir {}.pdopt not exist.".format(checkpoints)
para_dict, opti_dict = fluid.dygraph.load_dygraph(checkpoints) para_dict, opti_dict = paddle(checkpoints)
net.set_dict(para_dict) net.set_dict(para_dict)
optimizer.set_dict(opti_dict) optimizer.set_dict(opti_dict)
logger.info( logger.info(
...@@ -141,8 +142,8 @@ def save_model(net, optimizer, model_path, epoch_id, prefix='ppcls'): ...@@ -141,8 +142,8 @@ def save_model(net, optimizer, model_path, epoch_id, prefix='ppcls'):
_mkdir_if_not_exist(model_path) _mkdir_if_not_exist(model_path)
model_prefix = os.path.join(model_path, prefix) model_prefix = os.path.join(model_path, prefix)
fluid.dygraph.save_dygraph(net.state_dict(), model_prefix) paddle.save(net.state_dict(), model_prefix)
fluid.dygraph.save_dygraph(optimizer.state_dict(), model_prefix) paddle.save(optimizer.state_dict(), model_prefix)
logger.info( logger.info(
logger.coloring("Already save model in {}".format(model_path), logger.coloring("Already save model in {}".format(model_path),
"HEADER")) "HEADER"))
tools/eval.py
浏览文件 @ ea746480
...@@ -12,24 +12,22 @@ ...@@ -12,24 +12,22 @@
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
from __future__ import absolute_import
import program
from ppcls.utils import logger
from ppcls.utils.save_load import init_model
from ppcls.utils.config import get_config
from ppcls.data import Reader
import paddle.fluid as fluid
import paddle import paddle
import argparse from paddle.distributed import ParallelEnv
from __future__ import division
from __future__ import print_function
import argparse
import os import os
import sys import sys
__dir__ = os.path.dirname(os.path.abspath(__file__)) __dir__ = os.path.dirname(os.path.abspath(__file__))
sys.path.append(__dir__) sys.path.append(__dir__)
sys.path.append(os.path.abspath(os.path.join(__dir__, '..'))) sys.path.append(os.path.abspath(os.path.join(__dir__, '..')))
from ppcls.utils import logger
from ppcls.utils.save_load import init_model
from ppcls.utils.config import get_config
from ppcls.data import Reader
import program
def parse_args(): def parse_args():
parser = argparse.ArgumentParser("PaddleClas eval script") parser = argparse.ArgumentParser("PaddleClas eval script")
...@@ -54,20 +52,23 @@ def main(args): ...@@ -54,20 +52,23 @@ def main(args):
# assign place # assign place
use_gpu = config.get("use_gpu", True) use_gpu = config.get("use_gpu", True)
if use_gpu: if use_gpu:
gpu_id = fluid.dygraph.ParallelEnv().dev_id gpu_id = ParallelEnv().dev_id
place = fluid.CUDAPlace(gpu_id) place = paddle.CUDAPlace(gpu_id)
else: else:
place = fluid.CPUPlace() place = paddle.CPUPlace()
with fluid.dygraph.guard(place):
strategy = fluid.dygraph.parallel.prepare_context() paddle.disable_static(place)
net = program.create_model(config.ARCHITECTURE, config.classes_num)
net = fluid.dygraph.parallel.DataParallel(net, strategy) strategy = paddle.distributed.init_parallel_env()
init_model(config, net, optimizer=None) net = program.create_model(config.ARCHITECTURE, config.classes_num)
valid_dataloader = program.create_dataloader() net = paddle.DataParallel(net, strategy)
valid_reader = Reader(config, 'valid')() init_model(config, net, optimizer=None)
valid_dataloader.set_sample_list_generator(valid_reader, place) valid_dataloader = program.create_dataloader()
net.eval() valid_reader = Reader(config, 'valid')()
top1_acc = program.run(valid_dataloader, config, net, None, 0, 'valid') valid_dataloader.set_sample_list_generator(valid_reader, place)
net.eval()
top1_acc = program.run(valid_dataloader, config, net, None, None, 0,
'valid')
if __name__ == '__main__': if __name__ == '__main__':
......
tools/infer/infer.py
浏览文件 @ ea746480
...@@ -12,9 +12,6 @@ ...@@ -12,9 +12,6 @@
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
from ppcls.utils.save_load import load_dygraph_pretrain
from ppcls.modeling import architectures
import paddle.fluid as fluid
import numpy as np import numpy as np
import argparse import argparse
import utils import utils
...@@ -24,6 +21,12 @@ __dir__ = os.path.dirname(os.path.abspath(__file__)) ...@@ -24,6 +21,12 @@ __dir__ = os.path.dirname(os.path.abspath(__file__))
sys.path.append(__dir__) sys.path.append(__dir__)
sys.path.append(os.path.abspath(os.path.join(__dir__, '../..'))) sys.path.append(os.path.abspath(os.path.join(__dir__, '../..')))
from ppcls.utils.save_load import load_dygraph_pretrain
from ppcls.modeling import architectures
import paddle
from paddle.distributed import ParallelEnv
import paddle.nn.functional as F
def parse_args(): def parse_args():
def str2bool(v): def str2bool(v):
...@@ -92,32 +95,35 @@ def main(): ...@@ -92,32 +95,35 @@ def main():
operators = create_operators() operators = create_operators()
# assign the place # assign the place
if args.use_gpu: if args.use_gpu:
gpu_id = fluid.dygraph.parallel.Env().dev_id gpu_id = ParallelEnv().dev_id
place = fluid.CUDAPlace(gpu_id) place = paddle.CUDAPlace(gpu_id)
else: else:
place = fluid.CPUPlace() place = paddle.CPUPlace()
with fluid.dygraph.guard(place): paddle.disable_static(place)
net = architectures.__dict__[args.model]()
load_dygraph_pretrain(net, args.pretrained_model, net = architectures.__dict__[args.model]()
args.load_static_weights) load_dygraph_pretrain(net, args.pretrained_model, args.load_static_weights)
image_list = get_image_list(args.image_file) image_list = get_image_list(args.image_file)
for idx, filename in enumerate(image_list): for idx, filename in enumerate(image_list):
data = preprocess(filename, operators) data = preprocess(filename, operators)
data = np.expand_dims(data, axis=0) data = np.expand_dims(data, axis=0)
data = fluid.dygraph.to_variable(data) data = paddle.to_tensor(data)
net.eval() net.eval()
outputs = net(data) outputs = net(data)
outputs = fluid.layers.softmax(outputs) if args.model == "GoogLeNet":
outputs = outputs.numpy() outputs = outputs[0]
else:
probs = postprocess(outputs) outputs = F.softmax(outputs)
rank = 1 outputs = outputs.numpy()
print("Current image file: {}".format(filename))
for idx, prob in probs: probs = postprocess(outputs)
print("\ttop{:d}, class id: {:d}, probability: {:.4f}".format( rank = 1
rank, idx, prob)) print("Current image file: {}".format(filename))
rank += 1 for idx, prob in probs:
print("\ttop{:d}, class id: {:d}, probability: {:.4f}".format(
rank, idx, prob))
rank += 1
return return
......
tools/program.py
浏览文件 @ ea746480
...@@ -18,11 +18,12 @@ from __future__ import print_function ...@@ -18,11 +18,12 @@ from __future__ import print_function
import os import os
import time import time
from collections import OrderedDict from collections import OrderedDict
import paddle import paddle
import paddle.fluid as fluid from paddle import to_tensor
import paddle.nn as nn
import paddle.nn.functional as F
from ppcls.optimizer import LearningRateBuilder from ppcls.optimizer import LearningRateBuilder
from ppcls.optimizer import OptimizerBuilder from ppcls.optimizer import OptimizerBuilder
...@@ -34,8 +35,6 @@ from ppcls.modeling.loss import GoogLeNetLoss ...@@ -34,8 +35,6 @@ from ppcls.modeling.loss import GoogLeNetLoss
from ppcls.utils.misc import AverageMeter from ppcls.utils.misc import AverageMeter
from ppcls.utils import logger from ppcls.utils import logger
from paddle.fluid.dygraph.base import to_variable
def create_dataloader(): def create_dataloader():
""" """
...@@ -45,11 +44,11 @@ def create_dataloader(): ...@@ -45,11 +44,11 @@ def create_dataloader():
feeds(dict): dict of model input variables feeds(dict): dict of model input variables
Returns: Returns:
dataloader(fluid dataloader): dataloader(paddle dataloader):
""" """
trainer_num = int(os.environ.get('PADDLE_TRAINERS_NUM', 1)) trainer_num = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))
capacity = 64 if trainer_num == 1 else 8 capacity = 64 if trainer_num == 1 else 8
dataloader = fluid.io.DataLoader.from_generator( dataloader = paddle.io.DataLoader.from_generator(
capacity=capacity, use_double_buffer=True, iterable=True) capacity=capacity, use_double_buffer=True, iterable=True)
return dataloader return dataloader
...@@ -70,8 +69,6 @@ def create_model(architecture, classes_num): ...@@ -70,8 +69,6 @@ def create_model(architecture, classes_num):
""" """
name = architecture["name"] name = architecture["name"]
params = architecture.get("params", {}) params = architecture.get("params", {})
print(name)
print(params)
return architectures.__dict__[name](class_dim=classes_num, **params) return architectures.__dict__[name](class_dim=classes_num, **params)
...@@ -149,15 +146,15 @@ def create_metric(out, ...@@ -149,15 +146,15 @@ def create_metric(out,
# just need student label to get metrics # just need student label to get metrics
if use_distillation: if use_distillation:
out = out[1] out = out[1]
softmax_out = fluid.layers.softmax(out, use_cudnn=False) softmax_out = F.softmax(out)
fetchs = OrderedDict() fetchs = OrderedDict()
# set top1 to fetchs # set top1 to fetchs
top1 = fluid.layers.accuracy(softmax_out, label=label, k=1) top1 = paddle.metric.accuracy(softmax_out, label=label, k=1)
fetchs['top1'] = top1 fetchs['top1'] = top1
# set topk to fetchs # set topk to fetchs
k = min(topk, classes_num) k = min(topk, classes_num)
topk = fluid.layers.accuracy(softmax_out, label=label, k=k) topk = paddle.metric.accuracy(softmax_out, label=label, k=k)
topk_name = 'top{}'.format(k) topk_name = 'top{}'.format(k)
fetchs[topk_name] = topk fetchs[topk_name] = topk
...@@ -238,28 +235,34 @@ def create_optimizer(config, parameter_list=None): ...@@ -238,28 +235,34 @@ def create_optimizer(config, parameter_list=None):
# create optimizer instance # create optimizer instance
opt_config = config['OPTIMIZER'] opt_config = config['OPTIMIZER']
opt = OptimizerBuilder(**opt_config) opt = OptimizerBuilder(**opt_config)
return opt(lr, parameter_list) return opt(lr, parameter_list), lr
def create_feeds(batch, use_mix): def create_feeds(batch, use_mix):
image = batch[0] image = batch[0]
if use_mix: if use_mix:
y_a = to_variable(batch[1].numpy().astype("int64").reshape(-1, 1)) y_a = to_tensor(batch[1].numpy().astype("int64").reshape(-1, 1))
y_b = to_variable(batch[2].numpy().astype("int64").reshape(-1, 1)) y_b = to_tensor(batch[2].numpy().astype("int64").reshape(-1, 1))
lam = to_variable(batch[3].numpy().astype("float32").reshape(-1, 1)) lam = to_tensor(batch[3].numpy().astype("float32").reshape(-1, 1))
feeds = {"image": image, "y_a": y_a, "y_b": y_b, "lam": lam} feeds = {"image": image, "y_a": y_a, "y_b": y_b, "lam": lam}
else: else:
label = to_variable(batch[1].numpy().astype('int64').reshape(-1, 1)) label = to_tensor(batch[1].numpy().astype('int64').reshape(-1, 1))
feeds = {"image": image, "label": label} feeds = {"image": image, "label": label}
return feeds return feeds
def run(dataloader, config, net, optimizer=None, epoch=0, mode='train'): def run(dataloader,
config,
net,
optimizer=None,
lr_scheduler=None,
epoch=0,
mode='train'):
""" """
Feed data to the model and fetch the measures and loss Feed data to the model and fetch the measures and loss
Args: Args:
dataloader(fluid dataloader): dataloader(paddle dataloader):
exe(): exe():
program(): program():
fetchs(dict): dict of measures and the loss fetchs(dict): dict of measures and the loss
...@@ -303,6 +306,17 @@ def run(dataloader, config, net, optimizer=None, epoch=0, mode='train'): ...@@ -303,6 +306,17 @@ def run(dataloader, config, net, optimizer=None, epoch=0, mode='train'):
metric_list['lr'].update( metric_list['lr'].update(
optimizer._global_learning_rate().numpy()[0], batch_size) optimizer._global_learning_rate().numpy()[0], batch_size)
if lr_scheduler is not None:
if lr_scheduler.update_specified:
curr_global_counter = lr_scheduler.step_each_epoch * epoch + idx
update = max(
0, curr_global_counter - lr_scheduler.update_start_step
) % lr_scheduler.update_step_interval == 0
if update:
lr_scheduler.step()
else:
lr_scheduler.step()
for name, fetch in fetchs.items(): for name, fetch in fetchs.items():
metric_list[name].update(fetch.numpy()[0], batch_size) metric_list[name].update(fetch.numpy()[0], batch_size)
metric_list['batch_time'].update(time.time() - tic) metric_list['batch_time'].update(time.time() - tic)
......
tools/train.py
浏览文件 @ ea746480
...@@ -13,12 +13,6 @@ ...@@ -13,12 +13,6 @@
# limitations under the License. # limitations under the License.
from __future__ import absolute_import from __future__ import absolute_import
import program
from ppcls.utils import logger
from ppcls.utils.save_load import init_model, save_model
from ppcls.utils.config import get_config
from ppcls.data import Reader
import paddle.fluid as fluid
from __future__ import division from __future__ import division
from __future__ import print_function from __future__ import print_function
...@@ -30,6 +24,15 @@ sys.path.append(__dir__) ...@@ -30,6 +24,15 @@ sys.path.append(__dir__)
sys.path.append(os.path.abspath(os.path.join(__dir__, '..'))) sys.path.append(os.path.abspath(os.path.join(__dir__, '..')))
import paddle
from paddle.distributed import ParallelEnv
from ppcls.data import Reader
from ppcls.utils.config import get_config
from ppcls.utils.save_load import init_model, save_model
from ppcls.utils import logger
import program
def parse_args(): def parse_args():
parser = argparse.ArgumentParser("PaddleClas train script") parser = argparse.ArgumentParser("PaddleClas train script")
parser.add_argument( parser.add_argument(
...@@ -53,69 +56,66 @@ def main(args): ...@@ -53,69 +56,66 @@ def main(args):
# assign the place # assign the place
use_gpu = config.get("use_gpu", True) use_gpu = config.get("use_gpu", True)
if use_gpu: if use_gpu:
gpu_id = fluid.dygraph.ParallelEnv().dev_id gpu_id = ParallelEnv().dev_id
place = fluid.CUDAPlace(gpu_id) place = paddle.CUDAPlace(gpu_id)
else: else:
place = fluid.CPUPlace() place = paddle.CPUPlace()
use_data_parallel = int(os.getenv("PADDLE_TRAINERS_NUM", 1)) != 1 use_data_parallel = int(os.getenv("PADDLE_TRAINERS_NUM", 1)) != 1
config["use_data_parallel"] = use_data_parallel config["use_data_parallel"] = use_data_parallel
with fluid.dygraph.guard(place): paddle.disable_static(place)
net = program.create_model(config.ARCHITECTURE, config.classes_num)
net = program.create_model(config.ARCHITECTURE, config.classes_num)
optimizer = program.create_optimizer(
config, parameter_list=net.parameters()) optimizer, lr_scheduler = program.create_optimizer(
config, parameter_list=net.parameters())
if config["use_data_parallel"]:
strategy = fluid.dygraph.parallel.prepare_context() if config["use_data_parallel"]:
net = fluid.dygraph.parallel.DataParallel(net, strategy) strategy = paddle.distributed.init_parallel_env()
net = paddle.DataParallel(net, strategy)
# load model from checkpoint or pretrained model
init_model(config, net, optimizer) # load model from checkpoint or pretrained model
init_model(config, net, optimizer)
train_dataloader = program.create_dataloader()
train_reader = Reader(config, 'train')() train_dataloader = program.create_dataloader()
train_dataloader.set_sample_list_generator(train_reader, place) train_reader = Reader(config, 'train')()
train_dataloader.set_sample_list_generator(train_reader, place)
if config.validate:
valid_dataloader = program.create_dataloader() if config.validate:
valid_reader = Reader(config, 'valid')() valid_dataloader = program.create_dataloader()
valid_dataloader.set_sample_list_generator(valid_reader, place) valid_reader = Reader(config, 'valid')()
valid_dataloader.set_sample_list_generator(valid_reader, place)
best_top1_acc = 0.0 # best top1 acc record
for epoch_id in range(config.epochs): best_top1_acc = 0.0 # best top1 acc record
net.train() for epoch_id in range(config.epochs):
# 1. train with train dataset net.train()
program.run(train_dataloader, config, net, optimizer, epoch_id, # 1. train with train dataset
'train') program.run(train_dataloader, config, net, optimizer, lr_scheduler,
epoch_id, 'train')
if not config["use_data_parallel"] or fluid.dygraph.parallel.Env(
).local_rank == 0: if not config["use_data_parallel"] or ParallelEnv().local_rank == 0:
# 2. validate with validate dataset # 2. validate with validate dataset
if config.validate and epoch_id % config.valid_interval == 0: if config.validate and epoch_id % config.valid_interval == 0:
net.eval() net.eval()
top1_acc = program.run(valid_dataloader, config, net, None, top1_acc = program.run(valid_dataloader, config, net, None, None,
epoch_id, 'valid') epoch_id, 'valid')
if top1_acc > best_top1_acc: if top1_acc > best_top1_acc:
best_top1_acc = top1_acc best_top1_acc = top1_acc
message = "The best top1 acc {:.5f}, in epoch: {:d}".format( message = "The best top1 acc {:.5f}, in epoch: {:d}".format(
best_top1_acc, epoch_id) best_top1_acc, epoch_id)
logger.info("{:s}".format( logger.info("{:s}".format(logger.coloring(message, "RED")))
logger.coloring(message, "RED"))) if epoch_id % config.save_interval == 0:
if epoch_id % config.save_interval == 0:
model_path = os.path.join(config.model_save_dir,
model_path = os.path.join( config.ARCHITECTURE["name"])
config.model_save_dir, save_model(net, optimizer, model_path, "best_model")
config.ARCHITECTURE["name"])
save_model(net, optimizer, model_path, # 3. save the persistable model
"best_model") if epoch_id % config.save_interval == 0:
model_path = os.path.join(config.model_save_dir,
# 3. save the persistable model config.ARCHITECTURE["name"])
if epoch_id % config.save_interval == 0: save_model(net, optimizer, model_path, epoch_id)
model_path = os.path.join(config.model_save_dir,
config.ARCHITECTURE["name"])
save_model(net, optimizer, model_path, epoch_id)
if __name__ == '__main__': if __name__ == '__main__':
......
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