def _make_layer(self, block, planes, blocks, stride=1, dilation=1):
"""
Make resnet layer
:param block: Bottleneck
:param planes:
:param blocks:
:param stride:
:param dilation:
:return:
"""
BatchNorm = self._BatchNorm
downsample = None
# need conv in shortcut
if stride != 1 or self.in_planes != planes * block.expansion:
downsample = nn.Sequential(
conv1x1(self.in_planes, planes * block.expansion, stride),
BatchNorm(planes * block.expansion),
)
layers = list()
layers.append(
block(self.in_planes, planes, stride, downsample, dilation))
self.in_planes = planes * block.expansion
for _ in range(1, blocks):
layers.append(block(self.in_planes, planes, dilation=dilation))
return nn.Sequential(*layers)
def __init__(self):
super(ASPP, self).__init__()
in_planes = 2048
dilations = [1, 6, 12, 18]
# all aspp module output feature maps with channel 256
self.aspp1 = _ASPPModule(in_planes, planes=256, kernel_size=1, padding=0, dilation=dilations[0])
self.aspp2 = _ASPPModule(in_planes, planes=256, kernel_size=3, padding=dilations[1], dilation=dilations[1])
self.aspp3 = _ASPPModule(in_planes, planes=256, kernel_size=3, padding=dilations[2], dilation=dilations[2])
self.aspp4 = _ASPPModule(in_planes, planes=256, kernel_size=3, padding=dilations[3], dilation=dilations[3])
# perform global average pooling on the last feature map of the backbone
# batchsize must be greater than 1, otherwise exception will be thrown in calculating BatchNorm
self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
nn.Conv2d(in_planes, 256, 1, stride=1, bias=False),
nn.BatchNorm2d(256),
nn.ReLU())
self.p1 = nn.AdaptiveAvgPool2d(1)
self.p2 = nn.Conv2d(in_planes, 256, 1, stride=1, bias=False)
self.p3 = nn.BatchNorm2d(256)
self.conv1 = nn.Conv2d(1280, 256, 1, bias=False)
self.bn1 = nn.BatchNorm2d(256)
self.relu = nn.ReLU()
self.dropout = nn.Dropout(0.5)
self._init_weight()
def make_layers(self, block, planes, num_blocks, stride=1):
downsample = None
layers = []
if stride != 1 or self.inplanes != block.expansion * planes:
downsample = nn.Sequential(
nn.Conv2d(self.inplanes,
block.expansion * planes,
kernel_size=1,
stride=stride,
bias=False),
nn.BatchNorm2d(block.expansion * planes))
layers.append(
block(self.inplanes, planes, stride=stride, downsample=downsample))
self.inplanes = planes * block.expansion
for i in range(1, num_blocks):
layers.append(block(self.inplanes, planes))
return nn.Sequential(*layers)
def __init__(self, in_chn=1, cb_num=8):
super().__init__()
print('The Convolution Channel: {}'.format(Conv_CHN))
print('The Convolution Block: {}'.format(cb_num))
# self.se1 = SELayer(Convolution_CHN)
conv1 = conv_block(in_chn, Conv_CHN)
conv_more = [conv_block(Conv_CHN, Conv_CHN) for i in range(cb_num - 1)]
self.conv_blocks = nn.Sequential(conv1, *conv_more)
def __init__(self, in_planes, reduction=16):
super(SELayer, self).__init__()
# 返回1X1大小的特征图，通道数不变
self.avg_pool = nn.AdaptiveAvgPool1d(1)
self.fc = nn.Sequential(
nn.Linear(in_planes, in_planes // reduction, bias=False),
nn.ReLU(), # inplace = True, 计算值直接覆盖之前的值,最好默认为False,否则报错
nn.Linear(in_planes // reduction, in_planes, bias=False),
nn.Sigmoid())
def __init__(self, gate_channels, reduction_ratio=16):
super(ChannelGate, self).__init__()
self.gate_channels = gate_channels
self.gap1 = nn.AdaptiveAvgPool1d(output_size=1)
self.gap2 = nn.AdaptiveAvgPool1d(output_size=1)
self.mlp = nn.Sequential(
Flatten(),
nn.Linear(gate_channels, gate_channels // reduction_ratio),
nn.ReLU(),
nn.Linear(gate_channels // reduction_ratio, gate_channels))
def conv_block(in_channels, out_channels):
return nn.Sequential(
nn.Conv1d(in_channels,
out_channels,
kernel_size=K_SIZE,
padding=PADDING),
nn.BatchNorm1d(out_channels),
nn.ReLU(),
nn.MaxPool1d(kernel_size=2),
)
def _make_multi_grid_layer(self,
block,
planes,
blocks,
stride=1,
dilation=1):
"""
Multi-grid unit
:param block: Bottleneck
:param planes:
:param blocks:
:param stride:
:param dilation:
:return:
"""
downsample = None
if stride != 1 or self.in_planes != planes * block.expansion:
downsample = nn.Sequential(
nn.Conv2d(self.in_planes,
planes * block.expansion,
kernel_size=1,
stride=stride,
bias=False),
nn.BatchNorm2d(planes * block.expansion),
)
layers = list()
layers.append(
block(self.in_planes,
planes,
stride,
dilation=blocks[0] * dilation,
downsample=downsample))
self.in_planes = planes * block.expansion
for i in range(1, len(blocks)):
layers.append(
block(self.in_planes,
planes,
stride=1,
dilation=blocks[i] * dilation))
return nn.Sequential(*layers)
def __init__(self, cb_num=8):
super().__init__()
self.h_dim = 64
self.z_dim = 64
self.channel = 1
print('The Convolution Channel: {}'.format(self.h_dim))
print('The Convolution Block: {}'.format(cb_num))
conv1 = conv_block(self.channel, self.z_dim)
conv_more = [
conv_block(self.h_dim, self.z_dim) for i in range(cb_num - 1)
]
self.conv_blocks = nn.Sequential(conv1, *conv_more)
def __init__(self, in_channels, key_channels, out_channels, scale=1, dropout=0.1, bn_type=None):
super(SpatialOCR_Module, self).__init__()
self.object_context_block = ObjectAttentionBlock(in_channels, key_channels, scale, bn_type)
_in_channels = 2 * in_channels
self.conv_bn_dropout = nn.Sequential(
nn.Conv2d(_in_channels, out_channels, kernel_size=1, padding=0, bias=False),
nn.BatchNorm2d(out_channels),
nn.ReLU(inplace=True),
nn.Dropout2d(dropout)
)
def __init__(self, in_dim, n_class):
super(LeNet, self).__init__()
self.conv = nn.container.Sequential(
nn.conv.Conv2d(in_dim, 6, 5, stride=1,
padding=0), #1*28*18-->6*24*24
nn.activation.ReLU(True),
nn.pooling.MaxPool2d(2, 2), #6*24*24-->6*12*12
nn.conv.Conv2d(6, 16, 5, stride=1, padding=0), #6*12*12-->16*8*8
nn.activation.ReLU(True),
nn.pooling.MaxPool2d(2, 2)) #16*8*8-->16*4*4
self.fc = nn.Sequential(nn.Linear(16 * 4 * 4, 120), nn.Linear(120, 84),
nn.Linear(84, n_class))
def __init__(self, DIM):
super().__init__()
NUM_BLOCK = 8
FEATURE_CHN = 64
x_dim = int(FEATURE_CHN *
(DIM // 2**NUM_BLOCK)) # 2048:8192; 1024:4096
feature = 256 # 100(original), 64(CW2SQ), 32, 16. choose: 256, 16
# print('The NUM of ConvBlocK: {}'.format(NUM_BLOCK))
print('The FC features: {}\n'.format(feature))
self.create_feat = nn.Linear(x_dim,
feature) # weight shape: (feature, x_dim)
self.discriminator = nn.Sequential(nn.BatchNorm1d(feature), nn.ReLU(),
nn.Linear(feature, 2))
def __init__(self, in_channels, key_channels, scale=1, bn_type=None):
super(ObjectAttentionBlock, self).__init__()
self.scale = scale
self.in_channels = in_channels
self.key_channels = key_channels
self.pool = nn.MaxPool2d(kernel_size=(scale, scale))
self.f_pixel = nn.Sequential(
nn.Conv2d(in_channels=self.in_channels, out_channels=self.key_channels, kernel_size=1, stride=1,
padding=0, bias=False),
nn.BatchNorm2d(self.key_channels),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=self.key_channels, out_channels=self.key_channels, kernel_size=1, stride=1,
padding=0, bias=False),
nn.BatchNorm2d(self.key_channels),
nn.ReLU(inplace=True)
)
self.f_object = nn.Sequential(
nn.Conv2d(in_channels=self.in_channels, out_channels=self.key_channels, kernel_size=1, stride=1, padding=0,
bias=False),
nn.BatchNorm2d(self.key_channels),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=self.key_channels, out_channels=self.key_channels, kernel_size=1, stride=1, padding=0,
bias=False),
nn.BatchNorm2d(self.key_channels),
nn.ReLU(inplace=True)
)
self.f_down = nn.Sequential(
nn.Conv2d(in_channels=self.in_channels, out_channels=self.key_channels, kernel_size=1, stride=1, padding=0,
bias=False),
nn.BatchNorm2d(self.key_channels),
nn.ReLU(inplace=True)
)
self.f_up = nn.Sequential(
nn.Conv2d(in_channels=self.key_channels, out_channels=self.in_channels, kernel_size=1, stride=1, padding=0,
bias=False),
nn.BatchNorm2d(self.key_channels),
nn.ReLU(inplace=True)
)
def __init__(self, in_channels, out_channels, stride=1, M=2, r=16, L=32):
"""
:param in_channels: 输入通道维度
:param out_channels: 输出通道维度 原论文中 输入输出通道维度相同
:param stride: 步长，默认为1
:param M: 分支数
:param r: 特征Z的长度，计算其维度d 时所需的比率（论文中 特征S->Z 是降维，故需要规定 降维的下界）
:param L: 论文中规定特征Z的下界，默认为32
"""
super(SK_Conv1d, self).__init__()
d = max(in_channels // r, L) # 计算向量Z 的长度d
self.M = M
self.out_channels = out_channels
self.conv = nn.ModuleList() # 根据分支数量 添加 不同核的卷积操作
for i in range(M):
# 为提高效率，原论文中 扩张卷积5x5为 （3X3，dilation=2）来代替, 且论文中建议组卷积G=32,
# 每组计算只有out_channel/groups = 2 个channel参与.
self.conv.append(
nn.Sequential(
nn.Conv1d(in_channels,
out_channels,
3,
stride,
padding=1 + i,
dilation=1 + i,
groups=32,
bias=False), nn.BatchNorm1d(out_channels),
nn.ReLU(inplace=True)))
self.global_pool = nn.AdaptiveAvgPool1d(
1) # 自适应pool到指定维度, 这里指定为1，实现 GAP
self.fc1 = nn.Sequential(nn.Conv1d(out_channels, d, 1, bias=False),
nn.BatchNorm1d(d),
nn.ReLU(inplace=True)) # 降维
self.fc2 = nn.Conv1d(d, out_channels * M, 1, 1, bias=False) # 升维
# self.fcs = nn.ModuleList(self.fc1, self.fc2)
self.softmax = nn.Softmax(
dim=1) # 指定dim=1 使得两个全连接层对应位置进行softmax,保证 对应位置a+b+..=1
def __init__(self, n_states, n_actions, n_hidden, lr):
super(ActorCritic, self).__init__()
self.input = nn.Linear(n_states, n_hidden)
self.hidden_1 = nn.Linear(n_hidden, n_hidden)
self.hidden_2 = nn.Linear(n_hidden, n_hidden)
self.out_actor_sigma = nn.Linear(n_hidden, n_actions)
self.out_actor_mu = nn.Linear(n_hidden, n_actions)
self.out_critic = nn.Sequential(nn.Linear(n_hidden, n_hidden),
nn.ReLU(), nn.Linear(n_hidden, 1))
#self.bn_1 = nn.BatchNorm1d(n_hidden)
#self.bn_2 = nn.BatchNorm1d(n_hidden)
self.optimizer = optim.SGD(self.parameters(), lr=lr)
def __init__(self, num_classes):
super(Decoder, self).__init__()
low_level_in_planes = 256
self.conv1 = nn.Conv2d(low_level_in_planes, 48, 1, bias=False)
self.bn1 = nn.BatchNorm2d(48)
self.relu = nn.ReLU()
self.last_conv = nn.Sequential(
nn.Conv2d(304, 256, kernel_size=3, stride=1, padding=1,
bias=False), nn.BatchNorm2d(256), nn.ReLU(),
nn.Dropout(0.5),
nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1,
bias=False), nn.BatchNorm2d(256), nn.ReLU(),
nn.Dropout(0.1),
nn.Conv2d(256, num_classes, kernel_size=1, stride=1))
self._init_weight()
def __init__(self,
n_inputs,
n_outputs,
kernel_size,
stride,
dilation,
padding,
dropout=0.2):
super(TemporalBlock, self).__init__()
# 定义残差模块的第一层扩张卷积
# 经过conv，输出的size为(Batch, input_channel, seq_len + padding)，并归一化模型的参数
self.conv1 = weight_norm(
modules.Conv1d(n_inputs,
n_outputs,
kernel_size,
stride=stride,
padding=padding,
dilation=dilation))
# 裁剪掉多出来的padding部分，维持输出时间步为seq_len
self.chomp1 = Chomp1d(padding)
self.relu1 = modules.ReLU()
self.dropout1 = modules.Dropout(dropout)
#定义残差模块的第二层扩张卷积
self.conv2 = weight_norm(
modules.Conv1d(n_outputs,
n_outputs,
kernel_size,
stride=stride,
padding=padding,
dilation=dilation))
self.chomp2 = Chomp1d(padding)
self.relu2 = modules.ReLU()
self.dropout2 = modules.Dropout(dropout)
# 将卷积模块进行串联构成序列
self.net = modules.Sequential(self.conv1, self.chomp1, self.relu1,
self.dropout1, self.conv2, self.chomp2,
self.relu2, self.dropout2)
# 如果输入通道和输出通道不相同，那么通过1x1卷积进行降维，保持通道相同
self.downsample = modules.Conv1d(n_inputs, n_outputs,
1) if n_inputs != n_outputs else None
self.relu = modules.ReLU()
self.init_weights()
def __init__(self, input_num, hidden_num, output_num):
super(Net, self).__init__()
self.seq = nm.Sequential(
# BinarizeLinear(input_num, hidden_num),
# nm.ReLU(),
# BinarizeLinear(hidden_num, output_num),
# #______________________________________
# nn.BatchNorm1d(input_num),
# nn.Linear(input_num, hidden_num),
# nn.ReLU(),
# nn.BatchNorm1d(hidden_num),
# nn.Linear(hidden_num, output_num),
# #______________________________________
nm.BatchNorm1d(input_num),
BinarizeLinear(input_num, hidden_num),
nm.BatchNorm1d(hidden_num),
BinarizeLinear(hidden_num, output_num),
# BinarizeLinear(hidden_num, output_num),
)
def __init__(self, num_inputs, num_channels, kernel_size=2, dropout=0.2):
super(TemporalConvNet, self).__init__()
layers = []
num_levels = len(num_channels)
for i in range(num_levels):
# 扩张系数随着网络层级的增加而呈现指数级增加，以此来增大感受野
dilation_size = 2**i
# 计算输入通道数
in_channels = num_inputs if i == 0 else num_channels[i - 1]
# 确定每一层的输出通道数
out_channels = num_channels[i]
# 从num_channels中抽取每个残差模块的输入通道数和输出通道数
layers += [
TemporalBlock(in_channels,
out_channels,
kernel_size,
stride=1,
dilation=dilation_size,
padding=(kernel_size - 1) * dilation_size,
dropout=dropout)
]
# 将所有残差模块堆叠起来组成一个深度卷积网络
self.network = modules.Sequential(*layers)
def __init__(self, num_class=10):
super(VGG16, self).__init__()
self.feature = modules.Sequential(
# #1,
modules.Conv2d(3, 64, kernel_size=3, padding=1),
modules.BatchNorm2d(64),
modules.ReLU(True),
#2
modules.Conv2d(64, 64, kernel_size=3, padding=1),
modules.BatchNorm2d(64),
modules.ReLU(True),
modules.MaxPool2d(kernel_size=2, stride=2),
#3
modules.Conv2d(64, 128, kernel_size=3, padding=1),
modules.BatchNorm2d(128),
modules.ReLU(True),
# modules.MaxPool2d(kernel_size=2,stride=2),
#4
modules.Conv2d(128, 128, kernel_size=3, padding=1),
modules.BatchNorm2d(128),
modules.ReLU(True),
modules.MaxPool2d(kernel_size=2, stride=2),
#5
modules.Conv2d(128, 256, kernel_size=3, padding=1),
modules.BatchNorm2d(256),
modules.ReLU(True),
#6
modules.Conv2d(256, 256, kernel_size=3, padding=1),
modules.BatchNorm2d(256),
modules.ReLU(True),
#7
modules.Conv2d(256, 256, kernel_size=3, padding=1),
modules.BatchNorm2d(256),
modules.ReLU(True),
modules.MaxPool2d(kernel_size=2, stride=2),
#8
modules.Conv2d(256, 512, kernel_size=3, padding=1),
modules.BatchNorm2d(512),
modules.ReLU(True),
#9
modules.Conv2d(512, 512, kernel_size=3, padding=1),
modules.BatchNorm2d(512),
modules.ReLU(True),
#10
modules.Conv2d(512, 512, kernel_size=3, padding=1),
modules.BatchNorm2d(512),
modules.ReLU(True),
modules.MaxPool2d(kernel_size=2, stride=2),
#11
modules.Conv2d(512, 512, kernel_size=3, padding=1),
modules.BatchNorm2d(512),
modules.ReLU(True),
#12
modules.Conv2d(512, 512, kernel_size=3, padding=1),
modules.BatchNorm2d(512),
modules.ReLU(True),
#13
modules.Conv2d(512, 512, kernel_size=3, padding=1),
modules.BatchNorm2d(512),
modules.ReLU(True),
modules.MaxPool2d(kernel_size=2, stride=2),
modules.AvgPool2d(kernel_size=1, stride=1),
)
# 全连接层
self.classifier = modules.Sequential(
# #14
modules.Linear(512, 4096),
modules.ReLU(True),
modules.Dropout(),
#15
modules.Linear(4096, 4096),
modules.ReLU(True),
modules.Dropout(),
#16
modules.Linear(4096, num_class),
)
def __init__(self):
super(CNNNet, self).__init__()
self.cnn_layer = nn.Sequential(nn.Conv2d(in_channels=1))
self.fc_layer = nn.Sequential()
def __init__(self, input_num, hidden_num, output_num):
super(Net, self).__init__()
self.seq = nm.Sequential(nm.Linear(input_num, hidden_num), nm.ReLU(),
nm.Linear(hidden_num, output_num))
def __init__(self):
super(ModelClass, self).__init__()
device = torch.device('cpu')
net = tvM.resnet50()
summary(net.to(device), (3, 224, 224))
self.backbone = tvM.resnet50(pretrained=True)
# 固定输入尺寸224后才可用的全连接层，无法多尺度训练
# self.backbone.fc = nnM.Sequential(nnM.Linear(2048, 1024), nnM.Linear(1024, 9))
#Expected 3-dimensional tensor, but got 2-dimensional tensor for argument #1 'self' (while checking arguments for adaptive_avg_pool1d)
# self.backbone.fc = nnM.Sequential(nnM.AdaptiveAvgPool1d(1024), nnM.AdaptiveAvgPool1d(9))
#Expected 3-dimensional tensor, but got 2-dimensional tensor for argument #1 'self' (while checking arguments for adaptive_avg_pool1d)
# self.backbone.fc = nnM.Sequential(nnM.AdaptiveAvgPool1d(9))
#Expected 3-dimensional tensor, but got 4-dimensional tensor for argument #1 'self' (while checking arguments for adaptive_avg_pool1d)
# self.backbone.avgpool = nnM.Sequential( nnM.AdaptiveAvgPool1d(9))
#Expected 3-dimensional tensor, but got 2-dimensional tensor for argument #1 'self' (while checking arguments for adaptive_avg_pool1d) ???
# self.backbone.avgpool = nnM.AvgPool3d(kernel_size=7, stride=1, padding=0)
# self.backbone.fc = nnM.AdaptiveAvgPool1d(9)
# 打印出来resnet后两层观察（224*224输入）：
# Bottleneck-172 [-1, 2048, 7, 7] 0
# AvgPool2d-173 [-1, 2048, 1, 1] 0
# Linear-174 [-1, 1000] 2,049,000
# 在AvgPool2d之后，由于池化kernal等于7，已经相当于是全局pool了，输出一定是只有一个像素的2d(channel和batch两个维度)，所以上面一直报错“got 2-dimensional”
#Expected 3-dimensional tensor, but got 4-dimensional tensor for argument #1 'self' (while checking arguments for adaptive_avg_pool1d) ???
# self.backbone.avgpool = nnM.AdaptiveAvgPool2d((5,5))
# self.backbone.fc = nnM.AdaptiveAvgPool1d(9)
#size mismatch, m1: [35 x 51200], m2: [25 x 9] at /Users/soumith/minicondabuild3/conda-bld/pytorch_1524590658547/work/aten/src/TH/generic/THTensorMath.c:2033
# self.backbone.avgpool = nnM.AdaptiveAvgPool2d((5,5))
# self.backbone.fc = nnM.Linear(25,9)
#avgpool output is: (35, 2048, 5, 1)
#Expected 3-dimensional tensor, but got 4-dimensional tensor for argument #1 'self' (while checking arguments for adaptive_avg_pool1d)
# self.backbone.avgpool = nnM.AdaptiveAvgPool2d((5,1))
# self.backbone.fc = nnM.AdaptiveAvgPool1d(9)
#avgpool output is: (35,) 说明0尺度还是可以设的
#Expected 3-dimensional tensor, but got 1-dimensional tensor for argument #1 'self' (while checking arguments for adaptive_avg_pool1d)
# self.backbone.avgpool = nnM.AdaptiveAvgPool3d((0,5,5))
# self.backbone.fc = nnM.AdaptiveAvgPool1d(9)
#avgpool output is: (35, 5, 5)
#fc output is: (35, 5, 9) 为啥AdaptiveAvgPool1d输出是3d？？？见结论2
# self.backbone.avgpool = nnM.AdaptiveAvgPool3d((5,5,0))
# self.backbone.fc = nnM.AdaptiveAvgPool1d(9)
#avgpool output is: (35, 9, 1, 1)
#求loss时 RuntimeError: invalid argument 3: only batches of spatial targets supported (3D tensors) but got targets of dimension: 1 at /Users/soumith/minicondabuild3/conda-bld/pytorch_1524590658547/work/aten/src/THNN/generic/SpatialClassNLLCriterion.c:60
# self.backbone.avgpool = nnM.AdaptiveAvgPool3d((9,1,1))
#成功一、删掉 fc层，直接一个GAP-3D搞定。
# self.backbone.avgpool = nnM.AdaptiveAvgPool3d((9,0,0))
#成功二、留下fc层，用GAP-2D代替原来的pool2d。
self.backbone.avgpool = nnM.AdaptiveAvgPool2d((1, 1))
self.backbone.fc = nnM.Sequential(nnM.Linear(2048, 1024),
nnM.Linear(1024, 9))
