PyTorch implementation of LS-CNN architecture from the paper
- LS-CNN: Characterizing Local Patches at Multiple Scales for Face Recognition by Qiangchang Wang and Guodong Guo
Since similar discriminative face regions may occur at different scales, a new backbone network HSNet which extracts multi-scale features has been proposed in the paper. HSNet excracts multi-scale features from two harmonious perspectives.
-
Different kernels in a single layer, concatenation of features from different layers.
-
To identify local patches from which facial features can be extracted, new spatial attention is used in the paper.
-
Also in a CNN, channels in high-level layers represent high-level representations. So, a channel attention is also used in the paper to emphasize important channels and suppress less informative ones automatically.
-
This spatial and channel attention is called Dual Face Attention (DFA)
The goal of this package is to provide a nice and simple object-oriented implementation of the architecture. The individual submodules are cleanly separated into self-contained blocks, that come with documentation and typings, and that are therefore easy to import and reuse.
This project is based on Python 3.6+ and PyTorch 1.7.0+
Within the correct environment, install the package from the repository:
pip install git+https://github.com/Ksuryateja/LS-CNN Either load the predefined network from model.py
from model import LSCNN
net = LSCNN()Or use the modules for custom architecture:
from torch.nn import Sequential
from LSCNN import TransitionBlock, DenseBlock
class Module(Sequential):
def __init__(self, in_channels, out_channels):
super(Module, self).__init__()
self.dense = DenseBlock(in_channels, growth_rate=4, num_layers=2,
concat_input=True, dense_layer_params={'dropout': 0.2})
self.transition = TransitionBlock(self.dense.out_channels, out_channels)
net = Module(10, 4)Module((dense): DenseBlock(60, 1*30+60=90)(
(DenseBlock_0): Sequential(
(DFA_0): DFA(60, 16)(
(LANet): LALayer(
(spatial_atten): Sequential(
(0): Conv2d(60, 3, kernel_size=(1, 1), stride=(1, 1))
(1): ReLU(inplace=True)
(2): Conv2d(3, 1, kernel_size=(1, 1), stride=(1, 1))
(3): Sigmoid()
)
)
(SENet): SELayer(
(avg_pool): AdaptiveAvgPool2d(output_size=1)
(fc1): Linear(in_features=60, out_features=3, bias=False)
(relu): ReLU(inplace=True)
(fc2): Linear(in_features=3, out_features=60, bias=False)
(sigmoid): Sigmoid()
)
)
(DenseLayer_0): DenseLayer(60, 30)(
(branch1_1x1): BasicConv2d(
(conv): Conv2d(60, 120, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(120, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)
)
(branch2_1x1): BasicConv2d(
(conv): Conv2d(60, 120, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(120, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)
)
(branch3_1x1): BasicConv2d(
(conv): Conv2d(60, 30, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(30, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)
)
(branch1_3x3): BasicConv2d(
(conv): Conv2d(120, 30, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(30, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)
)
(branch1_3x3_2): BasicConv2d(
(conv): Conv2d(30, 30, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(30, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)
)
(branch2_3x3): BasicConv2d(
(conv): Conv2d(120, 30, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(30, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)
)
(output): BasicConv2d(
(conv): Conv2d(90, 30, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(30, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)
)
)
)
)
(transition): TransitionBlock()(
(transition_block): Sequential(
(DFA): DFA(90, 16)(
(LANet): LALayer(
(spatial_atten): Sequential(
(0): Conv2d(90, 5, kernel_size=(1, 1), stride=(1, 1))
(1): ReLU(inplace=True)
(2): Conv2d(5, 1, kernel_size=(1, 1), stride=(1, 1))
(3): Sigmoid()
)
)
(SENet): SELayer(
(avg_pool): AdaptiveAvgPool2d(output_size=1)
(fc1): Linear(in_features=90, out_features=5, bias=False)
(relu): ReLU(inplace=True)
(fc2): Linear(in_features=5, out_features=90, bias=False)
(sigmoid): Sigmoid()
)
)
(Transition layer): TransitionLayer(90, 45)(
(branch1_1x1): BasicConv2d(
(conv): Conv2d(90, 45, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(45, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)
)
(branch2_1x1): BasicConv2d(
(conv): Conv2d(90, 45, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(45, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)
)
(branch3_1x1): BasicConv2d(
(conv): Conv2d(90, 45, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(45, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)
)
(branch1_3x3): BasicConv2d(
(conv): Conv2d(45, 45, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(45, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)
)
(branch1_3x3_2): BasicConv2d(
(conv): Conv2d(45, 45, kernel_size=(3, 3), stride=(2, 2), bias=False)
(bn): BatchNorm2d(45, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)
)
(branch2_3x3): BasicConv2d(
(conv): Conv2d(45, 45, kernel_size=(3, 3), stride=(2, 2), bias=False)
(bn): BatchNorm2d(45, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)
)
(branch3_3x3): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
(output): BasicConv2d(
(conv): Conv2d(135, 45, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(45, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)
)
)
)
)
)