本文介绍SINet和ExtremeC3Net两个轻量级人像分割模型，二者参数分别为0.087M、0.038M，Flop为0.064G、0.128G。可通过PaddleHub快速调用，也能基于PaddleInference推理部署，并给出了Paddle2.0上的实现代码。
☞☞☞AI 智能聊天, 问答助手, AI 智能搜索, 免费无限量使用 DeepSeek R1 模型☜☜☜
引入
- 随着算力和算法的不断提升，能够训练的模型也越来越大了，当然精度也越来越高了
- 不过过于巨大的模型也带来了部署上的不便
- 今天就介绍两个轻量级的人像分割模型：SINet 和 ExtremeC3Net
项目说明
- 项目模型转换至开源项目ext_portrait_segmentation
- 感谢上述项目提供的开源代码和模型
模型规格
- 具体的模型规格如下表：
model Param Flop SINet 0.087 M 0.064 G ExtremeC3 0.038 M 0.128 G - 可以看出这两个模型算是相当轻量的了
效果展示
- ExtremeC3Net：
- SINet：
快速使用
- 按照惯例已经将两个模型封装为PaddleHub Module
- 可通过PaddleHub进行快速调用
In [1]
!pip install paddlehub==2.0.0b2
In [9]
# 导入PaddleHubimport paddlehub as hub# 加载模型# 模型可选：SINet_Portrait_Segmentation 和 ExtremeC3_Portrait_Segmentationmodel = hub.Module(directory='SINet_Portrait_Segmentation')# 人像分割outputs = model.Segmentation(images=None,
paths=['00001.jpg'],
batch_size=1,
output_dir='output',
visualization=True)# 结果显示%matplotlib inlineimport cv2import numpy as npimport matplotlib.pyplot as plt
img = np.concatenate([
cv2.imread('00001.jpg'),
cv2.cvtColor(outputs[0]['mask'], cv2.COLOR_GRAY2BGR),
outputs[0]['result']
], 1)
plt.axis('off')
plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
plt.show()
In [3]
# 导入PaddleHubimport paddlehub as hub# 加载模型# 模型可选：SINet_Portrait_Segmentation 和 ExtremeC3_Portrait_Segmentationmodel = hub.Module(directory='ExtremeC3_Portrait_Segmentation')# 人像分割outputs = model.Segmentation(images=None,
paths=['00001.jpg'],
batch_size=1,
output_dir='output',
visualization=True)# 结果显示%matplotlib inlineimport cv2import numpy as npimport matplotlib.pyplot as plt
img = np.concatenate([
cv2.imread('00001.jpg'),
cv2.cvtColor(outputs[0]['mask'], cv2.COLOR_GRAY2BGR),
outputs[0]['result']
], 1)
plt.axis('off')
plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
plt.show()
推理部署
- 除了使用PaddleHub一键调用之外，当然也可以使用推理模型进行推理部署
- 接下来简单介绍一下如何基于PaddleInference完成推理部署
- 更多详情可以参考我的另一个项目：PaddleQuickInference
In [4]
# 安装PaddleQuickInference!pip install ppqi -i https://pypi.python.org/simple
In [5]
from ppqi import InferenceModelfrom SINet.processor import preprocess, postprocess# 参数配置configs = { 'img_path': '00001.jpg', 'save_dir': 'save_img', 'model_name': 'SINet_Portrait_Segmentation', 'use_gpu': False, 'use_mkldnn': False}# 第一步：数据预处理input_data = preprocess(configs['img_path'])# 第二步：加载模型model = InferenceModel(
modelpath='SINet/'+configs['model_name'],
use_gpu=configs['use_gpu'],
use_mkldnn=configs['use_mkldnn']
)
model.eval()# 第三步：模型推理output = model(input_data)# 第四步：结果后处理postprocess(
output,
configs['save_dir'],
configs['img_path'],
configs['model_name']
)
In [6]
from ppqi import InferenceModelfrom ExtremeC3Net.processor import preprocess, postprocess# 参数配置configs = { 'img_path': '00001.jpg', 'save_dir': 'save_img', 'model_name': 'ExtremeC3_Portrait_Segmentation', 'use_gpu': False, 'use_mkldnn': False}# 第一步：数据预处理input_data = preprocess(configs['img_path'])# 第二步：加载模型model = InferenceModel(
modelpath='ExtremeC3Net/'+configs['model_name'],
use_gpu=configs['use_gpu'],
use_mkldnn=configs['use_mkldnn']
)
model.eval()# 第三步：模型推理output = model(input_data)# 第四步：结果后处理postprocess(
output,
configs['save_dir'],
configs['img_path'],
configs['model_name']
)
模型实现
- 接下来再介绍一下如何在Paddle2.0上实现这两个模型吧
- 代码上与原项目的Pytorch代码是相似的
- 针对框架之间的差异将其中的一些算子做了替换
- 具体详情请参考下方代码
In [7]
# model/sinet.py'''
ExtPortraitSeg
Copyright (c) 2019-present NAVER Corp.
MIT license
'''import paddleimport paddle.nn as nn
BN_moment = 0.1def channel_shuffle(x, groups):
batchsize, num_channels, height, width = x.shape
channels_per_group = num_channels // groups # reshape
x = x.reshape([batchsize, groups,
channels_per_group, height, width]) # transpose
x = paddle.transpose(x, [0, 2, 1, 3, 4])
# flatten
x = x.reshape([batchsize, groups*channels_per_group, height, width]) return xclass CBR(nn.Layer):
'''
This class defines the convolution layer with batch normalization and PReLU activation
'''
def __init__(self, nIn, nOut, kSize, stride=1):
'''
:param nIn: number of input channels
:param nOut: number of output channels
:param kSize: kernel size
:param stride: stride rate for down-sampling. Default is 1
'''
super().__init__()
padding = int((kSize - 1) / 2)
self.conv = nn.Conv2D(nIn, nOut, (kSize, kSize), stride=stride, padding=(padding, padding), bias_attr=False)
self.bn = nn.BatchNorm2D(nOut, epsilon=1e-03, momentum=BN_moment)
self.act = nn.PReLU(nOut) def forward(self, input):
'''
:param input: input feature map
:return: transformed feature map
'''
output = self.conv(input)
output = self.bn(output)
output = self.act(output) return outputclass separableCBR(nn.Layer):
'''
This class defines the convolution layer with batch normalization and PReLU activation
'''
def __init__(self, nIn, nOut, kSize, stride=1):
'''
:param nIn: number of input channels
:param nOut: number of output channels
:param kSize: kernel size
:param stride: stride rate for down-sampling. Default is 1
'''
super().__init__()
padding = int((kSize - 1) / 2)
self.conv = nn.Sequential(
nn.Conv2D(nIn, nIn, (kSize, kSize), stride=stride, padding=(padding, padding), groups=nIn, bias_attr=False),
nn.Conv2D(nIn, nOut, kernel_size=1, stride=1, bias_attr=False),
)
self.bn = nn.BatchNorm2D(nOut, epsilon=1e-03, momentum= BN_moment)
self.act = nn.PReLU(nOut) def forward(self, input):
'''
:param input: input feature map
:return: transformed feature map
'''
output = self.conv(input)
output = self.bn(output)
output = self.act(output) return outputclass SqueezeBlock(nn.Layer):
def __init__(self, exp_size, divide=4.0):
super(SqueezeBlock, self).__init__() if divide > 1:
self.dense = nn.Sequential(
nn.Linear(exp_size, int(exp_size / divide)),
nn.PReLU(int(exp_size / divide)),
nn.Linear(int(exp_size / divide), exp_size),
nn.PReLU(exp_size),
) else:
self.dense = nn.Sequential(
nn.Linear(exp_size, exp_size),
nn.PReLU(exp_size)
) def forward(self, x):
batch, channels, height, width = x.shape
out = paddle.nn.functional.avg_pool2d(x, kernel_size=[height, width]).reshape([batch, channels])
out = self.dense(out)
out = out.reshape([batch, channels, 1, 1]) return paddle.multiply(out, x)class SEseparableCBR(nn.Layer):
'''
This class defines the convolution layer with batch normalization and PReLU activation
'''
def __init__(self, nIn, nOut, kSize, stride=1, divide=2.0):
'''
:param nIn: number of input channels
:param nOut: number of output channels
:param kSize: kernel size
:param stride: stride rate for down-sampling. Default is 1
'''
super().__init__()
padding = int((kSize - 1) / 2)
self.conv = nn.Sequential(
nn.Conv2D(nIn, nIn, (kSize, kSize), stride=stride, padding=(padding, padding), groups=nIn, bias_attr=False),
SqueezeBlock(nIn, divide=divide),
nn.Conv2D(nIn, nOut, kernel_size=1, stride=1, bias_attr=False),
)
self.bn = nn.BatchNorm2D(nOut, epsilon=1e-03, momentum= BN_moment)
self.act = nn.PReLU(nOut) def forward(self, input):
'''
:param input: input feature map
:return: transformed feature map
'''
output = self.conv(input)
output = self.bn(output)
output = self.act(output) return outputclass BR(nn.Layer):
'''
This class groups the batch normalization and PReLU activation
'''
def __init__(self, nOut):
'''
:param nOut: output feature maps
'''
super().__init__()
self.bn = nn.BatchNorm2D(nOut, epsilon=1e-03, momentum= BN_moment)
self.act = nn.PReLU(nOut) def forward(self, input):
'''
:param input: input feature map
:return: normalized and thresholded feature map
'''
output = self.bn(input)
output = self.act(output) return outputclass CB(nn.Layer):
'''
This class groups the convolution and batch normalization
'''
def __init__(self, nIn, nOut, kSize, stride=1):
'''
:param nIn: number of input channels
:param nOut: number of output channels
:param kSize: kernel size
:param stride: optinal stide for down-sampling
'''
super().__init__()
padding = int((kSize - 1) / 2)
self.conv = nn.Conv2D(nIn, nOut, (kSize, kSize), stride=stride, padding=(padding, padding), bias_attr=False)
self.bn = nn.BatchNorm2D(nOut, epsilon=1e-03, momentum= BN_moment) def forward(self, input):
'''
:param input: input feature map
:return: transformed feature map
'''
output = self.conv(input)
output = self.bn(output) return outputclass C(nn.Layer):
'''
This class is for a convolutional layer.
'''
def __init__(self, nIn, nOut, kSize, stride=1,group=1):
'''
:param nIn: number of input channels
:param nOut: number of output channels
:param kSize: kernel size
:param stride: optional stride rate for down-sampling
'''
super().__init__()
padding = int((kSize - 1) / 2)
self.conv = nn.Conv2D(nIn, nOut, (kSize, kSize), stride=stride,
padding=(padding, padding), bias_attr=False, groups=group) def forward(self, input):
'''
:param input: input feature map
:return: transformed feature map
'''
output = self.conv(input) return outputclass S2block(nn.Layer):
'''
This class defines the dilated convolution.
'''
def __init__(self, nIn, nOut, config):
'''
:param nIn: number of input channels
:param nOut: number of output channels
:param kSize: kernel size
:param stride: optional stride rate for down-sampling
:param d: optional dilation rate
'''
super().__init__()
kSize = config[0]
avgsize = config[1]
self.resolution_down = False
if avgsize >1:
self.resolution_down = True
self.down_res = nn.AvgPool2D(avgsize, avgsize)
self.up_res = nn.Upsample(mode='bilinear', align_corners=True, align_mode=0, scale_factor=avgsize)
self.avgsize = avgsize
padding = int((kSize - 1) / 2 )
self.conv = nn.Sequential(
nn.Conv2D(nIn, nIn, kernel_size=(kSize, kSize), stride=1,
padding=(padding, padding), groups=nIn, bias_attr=False),
nn.BatchNorm2D(nIn, epsilon=1e-03, momentum=BN_moment))
self.act_conv1x1 = nn.Sequential(
nn.PReLU(nIn),
nn.Conv2D(nIn, nOut, kernel_size=1, stride=1, bias_attr=False),
)
self.bn = nn.BatchNorm2D(nOut, epsilon=1e-03, momentum=BN_moment) def forward(self, input):
'''
:param input: input feature map
:return: transformed feature map
'''
if self.resolution_down: input = self.down_res(input)
output = self.conv(input)
output = self.act_conv1x1(output) if self.resolution_down:
output = self.up_res(output) return self.bn(output)class S2module(nn.Layer):
'''
This class defines the ESP block, which is based on the following principle
Reduce ---> Split ---> Transform --> Merge
'''
def __init__(self, nIn, nOut, add=True, config= [[3,1],[5,1]]):
'''
:param nIn: number of input channels
:param nOut: number of output channels
:param add: if true, add a residual connection through identity operation. You can use projection too as
in ResNet paper, but we avoid to use it if the dimensions are not the same because we do not want to
increase the module complexity
'''
super().__init__()
group_n = len(config)
n = int(nOut / group_n)
n1 = nOut - group_n * n
self.c1 = C(nIn, n, 1, 1, group=group_n) for i in range(group_n):
var_name = 'd{}'.format(i + 1) if i == 0:
self.__dict__["_sub_layers"][var_name] = S2block(n, n + n1, config[i]) else:
self.__dict__["_sub_layers"][var_name] = S2block(n, n, config[i])
self.BR = BR(nOut)
self.add = add
self.group_n = group_n def forward(self, input):
'''
:param input: input feature map
:return: transformed feature map
'''
# reduce
output1 = self.c1(input)
output1= channel_shuffle(output1, self.group_n) for i in range(self.group_n):
var_name = 'd{}'.format(i + 1)
result_d = self.__dict__["_sub_layers"][var_name](output1) if i == 0:
combine = result_d else:
combine = paddle.concat([combine, result_d], 1) # if residual version
if self.add:
combine = paddle.add(input, combine)
output = self.BR(combine) return outputclass InputProjectionA(nn.Layer):
'''
This class projects the input image to the same spatial dimensions as the feature map.
For example, if the input image is 512 x512 x3 and spatial dimensions of feature map size are 56x56xF, then
this class will generate an output of 56x56x3
'''
def __init__(self, samplingTimes):
'''
:param samplingTimes: The rate at which you want to down-sample the image
'''
super().__init__()
self.pool = nn.LayerList() for i in range(0, samplingTimes):
self.pool.append(nn.AvgPool2D(2, stride=2)) def forward(self, input):
'''
:param input: Input RGB Image
:return: down-sampled image (pyramid-based approach)
'''
for pool in self.pool: input = pool(input) return inputclass SINet_Encoder(nn.Layer):
def __init__(self, config,classes=20, p=5, q=3, chnn=1.0):
'''
:param classes: number of classes in the dataset. Default is 20 for the cityscapes
:param p: depth multiplier
:param q: depth multiplier
'''
super().__init__()
dim1 = 16
dim2 = 48 + 4 * (chnn - 1)
dim3 = 96 + 4 * (chnn - 1)
self.level1 = CBR(3, 12, 3, 2)
self.level2_0 = SEseparableCBR(12,dim1, 3,2, divide=1)
self.level2 = nn.LayerList() for i in range(0, p): if i ==0:
self.level2.append(S2module(dim1, dim2, config=config[i], add=False)) else:
self.level2.append(S2module(dim2, dim2,config=config[i]))
self.BR2 = BR(dim2+dim1)
self.level3_0 =SEseparableCBR(dim2+dim1,dim2, 3,2, divide=2)
self.level3 = nn.LayerList() for i in range(0, q): if i==0:
self.level3.append(S2module(dim2, dim3, config=config[2 + i], add=False)) else:
self.level3.append(S2module(dim3, dim3,config=config[2+i]))
self.BR3 = BR(dim3+dim2)
self.classifier = C(dim3+dim2, classes, 1, 1) def forward(self, input):
'''
:param input: Receives the input RGB image
:return: the transformed feature map with spatial dimensions 1/8th of the input image
'''
output1 = self.level1(input) #8h 8w
output2_0 = self.level2_0(output1) # 4h 4w
for i, layer in enumerate(self.level2): if i == 0:
output2 = layer(output2_0) else:
output2 = layer(output2) # 2h 2w
output3_0 = self.level3_0(self.BR2(paddle.concat([output2_0, output2],1))) # h w
for i, layer in enumerate(self.level3): if i == 0:
output3 = layer(output3_0) else:
output3 = layer(output3)
output3_cat = self.BR3(paddle.concat([output3_0, output3], 1))
classifier = self.classifier(output3_cat) return classifierclass SINet(nn.Layer):
def __init__(self,config, classes=20, p=2, q=3, chnn=1.0):
'''
:param classes: number of classes in the dataset. Default is 20 for the cityscapes
:param p: depth multiplier
:param q: depth multiplier
'''
super().__init__()
dim2 = 48 + 4 * (chnn - 1)
self.encoder = SINet_Encoder(config, classes, p, q, chnn)
self.up = nn.Upsample(mode='bilinear', align_corners=True, align_mode=0, scale_factor=2)
self.bn_3 = nn.BatchNorm2D(classes, epsilon=1e-03)
self.level2_C = CBR(dim2, classes, 1, 1)
self.bn_2 = nn.BatchNorm2D(classes, epsilon=1e-03)
self.classifier = nn.Sequential(
nn.Upsample(mode='bilinear', align_corners=True, align_mode=0, scale_factor=2),
nn.Conv2D(classes, classes, 3, 1, 1, bias_attr=False)) def forward(self, input):
'''
:param input: RGB image
:return: transformed feature map
'''
output1 = self.encoder.level1(input) # 8h 8w
output2_0 = self.encoder.level2_0(output1) # 4h 4w
for i, layer in enumerate(self.encoder.level2): if i == 0:
output2 = layer(output2_0) else:
output2 = layer(output2) # 2h 2w
output3_0 = self.encoder.level3_0(self.encoder.BR2(paddle.concat([output2_0, output2], 1))) # h w
for i, layer in enumerate(self.encoder.level3): if i == 0:
output3 = layer(output3_0) else:
output3 = layer(output3)
output3_cat = self.encoder.BR3(paddle.concat([output3_0, output3], 1))
Enc_final = self.encoder.classifier(output3_cat) #1/8
Dnc_stage1 = self.bn_3(self.up(Enc_final)) # 1/4
stage1_confidence = paddle.max(nn.functional.softmax(Dnc_stage1, 1), axis=1)
b, c, h, w = Dnc_stage1.shape
stage1_gate = (1-stage1_confidence).unsqueeze(1).expand([b, c, h, w])
Dnc_stage2_0 = self.level2_C(output2) # 2h 2w
Dnc_stage2 = self.bn_2(self.up(paddle.add(paddle.multiply(Dnc_stage2_0, stage1_gate), (Dnc_stage1)))) # 4h 4w
classifier = self.classifier(Dnc_stage2) return classifier
In [8]
# model/extremeC3.py'''
ExtPortraitSeg
Copyright (c) 2019-present NAVER Corp.
MIT license
'''import paddleimport paddle.nn as nn
basic_0 = 24basic_1 = 48basic_2 = 56basic_3 = 24class CBR(nn.Layer):
'''
This class defines the convolution layer with batch normalization and PReLU activation
'''
def __init__(self, nIn, nOut, kSize, stride=1):
'''
:param nIn: number of input channels
:param nOut: number of output channels
:param kSize: kernel size
:param stride: stride rate for down-sampling. Default is 1
'''
super().__init__()
padding = int((kSize - 1) / 2) # self.conv = nn.Conv2D(nIn, nOut, kSize, stride=stride, padding=padding, bias_attr=False)
self.conv = nn.Conv2D(nIn, nOut, (kSize, kSize), stride=stride, padding=(padding, padding), bias_attr=False) # self.conv1 = nn.Conv2D(nOut, nOut, (1, kSize), stride=1, padding=(0, padding), bias_attr=False)
self.bn = nn.BatchNorm2D(nOut, epsilon=1e-03)
self.act = nn.PReLU(nOut) # self.act = nn.ReLU()
def forward(self, input):
'''
:param input: input feature map
:return: transformed feature map
'''
output = self.conv(input) # output = self.conv1(output)
output = self.bn(output)
output = self.act(output) return outputclass BR(nn.Layer):
'''
This class groups the batch normalization and PReLU activation
'''
def __init__(self, nOut):
'''
:param nOut: output feature maps
'''
super().__init__()
self.bn = nn.BatchNorm2D(nOut, epsilon=1e-03)
self.act = nn.PReLU(nOut) # self.act = nn.ReLU()
def forward(self, input):
'''
:param input: input feature map
:return: normalized and thresholded feature map
'''
output = self.bn(input)
output = self.act(output) return outputclass CB(nn.Layer):
'''
This class groups the convolution and batch normalization
'''
def __init__(self, nIn, nOut, kSize, stride=1):
'''
:param nIn: number of input channels
:param nOut: number of output channels
:param kSize: kernel size
:param stride: optinal stide for down-sampling
'''
super().__init__()
padding = int((kSize - 1) / 2)
self.conv = nn.Conv2D(nIn, nOut, (kSize, kSize), stride=stride, padding=(padding, padding), bias_attr=False)
self.bn = nn.BatchNorm2D(nOut, epsilon=1e-03) def forward(self, input):
'''
:param input: input feature map
:return: transformed feature map
'''
output = self.conv(input)
output = self.bn(output) return outputclass C(nn.Layer):
'''
This class is for a convolutional layer.
'''
def __init__(self, nIn, nOut, kSize, stride=1):
'''
:param nIn: number of input channels
:param nOut: number of output channels
:param kSize: kernel size
:param stride: optional stride rate for down-sampling
'''
super().__init__()
padding = int((kSize - 1) / 2)
self.conv = nn.Conv2D(nIn, nOut, (kSize, kSize), stride=stride, padding=(padding, padding), bias_attr=False) def forward(self, input):
'''
:param input: input feature map
:return: transformed feature map
'''
output = self.conv(input) return outputclass C3block(nn.Layer):
'''
This class defines the dilated convolution.
'''
def __init__(self, nIn, nOut, kSize, stride=1, d=1):
'''
:param nIn: number of input channels
:param nOut: number of output channels
:param kSize: kernel size
:param stride: optional stride rate for down-sampling
:param d: optional dilation rate
'''
super().__init__()
padding = int((kSize - 1) / 2) * d if d == 1:
self.conv =nn.Sequential(
nn.Conv2D(nIn, nIn, (kSize, kSize), stride=stride, padding=(padding, padding), groups=nIn, bias_attr=False,
dilation=d),
nn.Conv2D(nIn, nOut, kernel_size=1, stride=1, bias_attr=False)
) else:
combine_kernel = 2 * d - 1
self.conv = nn.Sequential(
nn.Conv2D(nIn, nIn, kernel_size=(combine_kernel, 1), stride=stride, padding=(padding - 1, 0),
groups=nIn, bias_attr=False),
nn.BatchNorm2D(nIn),
nn.PReLU(nIn),
nn.Conv2D(nIn, nIn, kernel_size=(1, combine_kernel), stride=stride, padding=(0, padding - 1),
groups=nIn, bias_attr=False),
nn.BatchNorm2D(nIn),
nn.Conv2D(nIn, nIn, (kSize, kSize), stride=stride, padding=(padding, padding), groups=nIn, bias_attr=False,
dilation=d),
nn.Conv2D(nIn, nOut, kernel_size=1, stride=1, bias_attr=False)) def forward(self, input):
'''
:param input: input feature map
:return: transformed feature map
'''
output = self.conv(input) return outputclass Down_advancedC3(nn.Layer):
def __init__(self, nIn, nOut, ratio=[2,4,8]):
super().__init__()
n = int(nOut // 3)
n1 = nOut - 3 * n
self.c1 = C(nIn, n, 3, 2)
self.d1 = C3block(n, n+n1, 3, 1, ratio[0])
self.d2 = C3block(n, n, 3, 1, ratio[1])
self.d3 = C3block(n, n, 3, 1, ratio[2])
self.bn = nn.BatchNorm2D(nOut, epsilon=1e-3)
self.act = nn.PReLU(nOut) def forward(self, input):
output1 = self.c1(input)
d1 = self.d1(output1)
d2 = self.d2(output1)
d3 = self.d3(output1)
combine = paddle.concat([d1, d2, d3], 1)
output = self.bn(combine)
output = self.act(output) return outputclass AdvancedC3(nn.Layer):
'''
This class defines the ESP block, which is based on the following principle
Reduce ---> Split ---> Transform --> Merge
'''
def __init__(self, nIn, nOut, add=True, ratio=[2,4,8]):
'''
:param nIn: number of input channels
:param nOut: number of output channels
:param add: if true, add a residual connection through identity operation. You can use projection too as
in ResNet paper, but we avoid to use it if the dimensions are not the same because we do not want to
increase the module complexity
'''
super().__init__()
n = int(nOut // 3)
n1 = nOut - 3 * n
self.c1 = C(nIn, n, 1, 1)
self.d1 = C3block(n, n + n1, 3, 1, ratio[0])
self.d2 = C3block(n, n, 3, 1, ratio[1])
self.d3 = C3block(n, n, 3, 1, ratio[2]) # self.d4 = Double_CDilated(n, n, 3, 1, 12)
# self.conv =C(nOut, nOut, 1,1)
self.bn = BR(nOut)
self.add = add def forward(self, input):
'''
:param input: input feature map
:return: transformed feature map
'''
# reduce
output1 = self.c1(input)
d1 = self.d1(output1)
d2 = self.d2(output1)
d3 = self.d3(output1)
combine = paddle.concat([d1, d2, d3], 1) if self.add:
combine = paddle.add(input, combine)
output = self.bn(combine) return outputclass InputProjectionA(nn.Layer):
'''
This class projects the input image to the same spatial dimensions as the feature map.
For example, if the input image is 512 x512 x3 and spatial dimensions of feature map size are 56x56xF, then
this class will generate an output of 56x56x3
'''
def __init__(self, samplingTimes):
'''
:param samplingTimes: The rate at which you want to down-sample the image
'''
super().__init__()
self.pool = nn.LayerList() for i in range(0, samplingTimes): # pyramid-based approach for down-sampling
self.pool.append(nn.AvgPool2D(2, stride=2, padding=0)) def forward(self, input):
'''
:param input: Input RGB Image
:return: down-sampled image (pyramid-based approach)
'''
for pool in self.pool: input = pool(input) return inputclass ExtremeC3NetCoarse(nn.Layer):
'''
This class defines the ESPNet-C network in the paper
'''
def __init__(self, classes=20, p=5, q=3):
'''
:param classes: number of classes in the dataset. Default is 20 for the cityscapes
:param p: depth multiplier
:param q: depth multiplier
'''
super().__init__()
self.level1 = CBR(3, basic_0, 3, 2)
self.sample1 = InputProjectionA(1)
self.sample2 = InputProjectionA(2)
self.b1 = BR(basic_0 + 3)
self.level2_0 = Down_advancedC3(basic_0 + 3, basic_1, ratio=[1, 2, 3]) # , ratio=[1,2,3]
self.level2 = nn.LayerList() for i in range(0, p):
self.level2.append(
AdvancedC3(basic_1, basic_1, ratio=[1, 3, 4])) # , ratio=[1,3,4]
self.b2 = BR(basic_1 * 2 + 3)
self.level3_0 = AdvancedC3(basic_1 * 2 + 3, basic_2, add=False,
ratio=[1, 3, 5]) # , ratio=[1,3,5]
self.level3 = nn.LayerList() for i in range(0, q):
self.level3.append(AdvancedC3(basic_2, basic_2))
self.b3 = BR(basic_2 * 2)
self.Coarseclassifier = C(basic_2*2, classes, 1, 1) def forward(self, input):
'''
:param input: Receives the input RGB image
:return: the transformed feature map with spatial dimensions 1/8th of the input image
'''
output0 = self.level1(input)
inp1 = self.sample1(input)
inp2 = self.sample2(input)
output0_cat = self.b1(paddle.concat([output0, inp1], 1))
output1_0 = self.level2_0(output0_cat) # down-sampled
for i, layer in enumerate(self.level2): if i == 0:
output1 = layer(output1_0) else:
output1 = layer(output1)
output1_cat = self.b2(paddle.concat([output1, output1_0, inp2], 1))
output2_0 = self.level3_0(output1_cat) # down-sampled
for i, layer in enumerate(self.level3): if i == 0:
output2 = layer(output2_0) else:
output2 = layer(output2)
output2_cat = self.b3(paddle.concat([output2_0, output2], 1))
classifier = self.Coarseclassifier(output2_cat) return classifierclass ExtremeC3Net(nn.Layer):
'''
This class defines the ESPNet-C network in the paper
'''
def __init__(self, classes=20, p=5, q=3):
'''
:param classes: number of classes in the dataset. Default is 20 for the cityscapes
:param p: depth multiplier
:param q: depth multiplier
'''
super().__init__()
self.encoder = ExtremeC3NetCoarse(classes, p, q) # # load the encoder modules
del self.encoder.Coarseclassifier
self.upsample = nn.Sequential(
nn.Conv2D(kernel_size=(1, 1), in_channels=basic_2*2, out_channels=basic_3,bias_attr=False),
nn.BatchNorm2D(basic_3),
nn.Upsample(mode='bilinear', align_corners=True, align_mode=0, scale_factor=2)
)
self.Fine = nn.Sequential( # nn.Conv2D(kernel_size=3, stride=2, padding=1, in_channels=3, out_channels=basic_3,bias_attr=False),
C(3, basic_3, 3, 2),
AdvancedC3(basic_3, basic_3, add=True), # nn.BatchNorm2D(basic_3, epsilon=1e-03),
)
self.classifier = nn.Sequential(
BR(basic_3),
nn.Upsample(mode='bilinear', align_corners=True, align_mode=0, scale_factor=2),
nn.Conv2D(kernel_size=(1, 1), in_channels=basic_3, out_channels=classes, bias_attr=False),
) def forward(self, input):
'''
:param input: Receives the input RGB image
:return: the transformed feature map with spatial dimensions 1/8th of the input image
'''
output0 = self.encoder.level1(input)
inp1 = self.encoder.sample1(input)
inp2 = self.encoder.sample2(input)
output0_cat = self.encoder.b1(paddle.concat([output0, inp1], 1))
output1_0 = self.encoder.level2_0(output0_cat) # down-sampled
for i, layer in enumerate(self.encoder.level2): if i == 0:
output1 = layer(output1_0) else:
output1 = layer(output1)
output1_cat = self.encoder.b2(paddle.concat([output1, output1_0, inp2], 1))
output2_0 = self.encoder.level3_0(output1_cat) # down-sampled
for i, layer in enumerate(self.encoder.level3): if i == 0:
output2 = layer(output2_0) else:
output2 = layer(output2)
output2_cat = self.encoder.b3(paddle.concat([output2_0, output2], 1))
Coarse = self.upsample(output2_cat)
Fine = self.Fine(input)
classifier = self.classifier(paddle.add(Coarse, Fine))
return classifier
