def __init__(self, cfg = None, A=2): super(Net, self).__init__() if cfg is None: # 模型结构 cfg = [192, 160, 96, 192, 192, 192, 192, 192] self.tnn_bin = nn.Sequential( nn.Conv2d(3, cfg[0], kernel_size=5, stride=1, padding=2), bn.BatchNorm2d_bin(cfg[0], affine_flag=2), Tnn_Bin_Conv2d(cfg[0], cfg[1], kernel_size=1, stride=1, padding=0, A=A), Tnn_Bin_Conv2d(cfg[1], cfg[2], kernel_size=1, stride=1, padding=0, activation_mp=1, A=A), nn.MaxPool2d(kernel_size=3, stride=2, padding=1), Tnn_Bin_Conv2d(cfg[2], cfg[3], kernel_size=5, stride=1, padding=2, activation_nor=0, A=A), Tnn_Bin_Conv2d(cfg[3], cfg[4], kernel_size=1, stride=1, padding=0, A=A), Tnn_Bin_Conv2d(cfg[4], cfg[5], kernel_size=1, stride=1, padding=0, activation_mp=1, A=A), nn.AvgPool2d(kernel_size=3, stride=2, padding=1), Tnn_Bin_Conv2d(cfg[5], cfg[6], kernel_size=3, stride=1, padding=1, activation_nor=0, A=A), Tnn_Bin_Conv2d(cfg[6], cfg[7], kernel_size=1, stride=1, padding=0, last=0, last_relu=1, A=A), nn.Conv2d(cfg[7], 10, kernel_size=1, stride=1, padding=0), bn.BatchNorm2d_bin(10, affine_flag=2), nn.ReLU(inplace=True), nn.AvgPool2d(kernel_size=8, stride=1, padding=0), )
def __init__(self, cfg = None, A=2): super(Net, self).__init__() if cfg is None: # 模型结构 cfg = [256, 256, 256, 512, 512, 512, 1024, 1024] self.tnn_bin = nn.Sequential( nn.Conv2d(3, cfg[0], kernel_size=5, stride=1, padding=2), bn.BatchNorm2d_bin(cfg[0], affine_flag=2), Tnn_Bin_Conv2d(cfg[0], cfg[1], kernel_size=1, stride=1, padding=0, groups=2, channel_shuffle=0, A=A), Tnn_Bin_Conv2d(cfg[1], cfg[2], kernel_size=1, stride=1, padding=0, groups=2, channel_shuffle=1, shuffle_groups=2, activation_mp=1, A=A), nn.MaxPool2d(kernel_size=2, stride=2, padding=0), Tnn_Bin_Conv2d(cfg[2], cfg[3], kernel_size=3, stride=1, padding=1, groups=16, channel_shuffle=1, shuffle_groups=2, activation_nor=0, A=A), Tnn_Bin_Conv2d(cfg[3], cfg[4], kernel_size=1, stride=1, padding=0, groups=4, channel_shuffle=1, shuffle_groups=16, A=A), Tnn_Bin_Conv2d(cfg[4], cfg[5], kernel_size=1, stride=1, padding=0, groups=4, channel_shuffle=1, shuffle_groups=4, activation_mp=1, A=A), nn.MaxPool2d(kernel_size=2, stride=2, padding=0), Tnn_Bin_Conv2d(cfg[5], cfg[6], kernel_size=3, stride=1, padding=1, groups=32, channel_shuffle=1, shuffle_groups=4, activation_nor=0, A=A), Tnn_Bin_Conv2d(cfg[6], cfg[7], kernel_size=1, stride=1, padding=0, groups=8, channel_shuffle=1, shuffle_groups=32, last=0, last_relu=1, A=A), nn.Conv2d(cfg[7], 10, kernel_size=1, stride=1, padding=0), bn.BatchNorm2d_bin(10, affine_flag=2), nn.ReLU(inplace=True), nn.AvgPool2d(kernel_size=8, stride=1, padding=0), )
def __init__(self, input_channels, output_channels, kernel_size=-1, stride=-1, padding=-1, dropout=0, last=0, activation_mp=0, activation_nor=1, last_relu=0, A=2): super(Tnn_Bin_Conv2d, self).__init__() self.A = A self.dropout_ratio = dropout self.last = last self.activation_mp = activation_mp self.activation_nor = activation_nor self.last_relu = last_relu if dropout != 0: self.dropout = nn.Dropout(dropout) self.conv = nn.Conv2d(input_channels, output_channels, kernel_size=kernel_size, stride=stride, padding=padding) #self.bn = nn.BatchNorm2d(output_channels) self.bn = bn.BatchNorm2d_bin( output_channels, momentum=0.1, affine_flag=2) #自定义BN_γ=1、β-train;WbAb_momentum=0.8 self.relu = nn.ReLU(inplace=True)
def __init__(self, input_channels, output_channels, kernel_size=-1, stride=-1, padding=-1, dropout=0, groups=1, last_relu=0, A=2, W=2): super(Tnn_Bin_Conv2d, self).__init__() self.A = A self.W = W self.dropout_ratio = dropout self.last_relu = last_relu if self.dropout_ratio != 0: self.dropout = nn.Dropout(dropout) # ********************* 量化(三/二值)卷积 ********************* self.tnn_bin_conv = Conv2d_Q(input_channels, output_channels, kernel_size=kernel_size, stride=stride, padding=padding, groups=groups, A=A, W=W) #self.bn = nn.BatchNorm2d(output_channels) self.bn = bn.BatchNorm2d_bin(output_channels, affine_flag=2) #自定义BN_γ=1、β-train self.relu = nn.ReLU(inplace=True)
def __init__(self, input_channels, output_channels, kernel_size=-1, stride=-1, padding=-1, dropout=0, groups=1, channel_shuffle=0, shuffle_groups=1, last=0, activation_mp=0, activation_nor=1, last_relu=0, A=2): super(Tnn_Bin_Conv2d, self).__init__() self.A = A self.dropout_ratio = dropout self.channel_shuffle_flag = channel_shuffle self.shuffle_groups = shuffle_groups self.last = last self.activation_mp = activation_mp self.activation_nor = activation_nor self.last_relu = last_relu if dropout!=0: self.dropout = nn.Dropout(dropout) self.conv = nn.Conv2d(input_channels, output_channels, kernel_size=kernel_size, stride=stride, padding=padding, groups=groups) #self.bn = nn.BatchNorm2d(output_channels) self.bn = bn.BatchNorm2d_bin(output_channels, affine_flag=1)#自定义BN_γ=1、β-train self.relu = nn.ReLU(inplace=True)