def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True, padding_mode='zeros', momentum=0.01, a_bits=8, w_bits=8, bn_fuse=0, q_type=1, q_level=0, first_layer=0): super(QuantConvBNReLU, self).__init__() self.bn_fuse = bn_fuse if self.bn_fuse == 1: self.quant_bn_fuse_conv = QuantBNFuseConv2d(in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=False, padding_mode=padding_mode, a_bits=a_bits, w_bits=w_bits, q_type=q_type, q_level=q_level, first_layer=first_layer) else: self.quant_conv = QuantConv2d(in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias, padding_mode=padding_mode, a_bits=a_bits, w_bits=w_bits, q_type=q_type, q_level=q_level, first_layer=first_layer) self.bn = nn.BatchNorm2d(out_channels, momentum=momentum) # 考虑量化带来的抖动影响,对momentum进行调整(0.1 ——> 0.01),削弱batch统计参数占比,一定程度抑制抖动。经实验量化训练效果更好,acc提升1%左右 self.relu = QuantReLU(inplace=True, a_bits=a_bits, q_type=q_type)
def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True, padding_mode='zeros', momentum=0.1, channel_shuffle=0, shuffle_groups=1, a_bits=8, w_bits=8, first_layer=0): super(QuantConvBNReLU, self).__init__() self.channel_shuffle_flag = channel_shuffle self.shuffle_groups = shuffle_groups self.quant_conv = QuantConv2d(in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias, padding_mode=padding_mode, a_bits=a_bits, w_bits=w_bits, first_layer=first_layer) self.bn = nn.BatchNorm2d(out_channels, momentum=momentum) self.relu = nn.ReLU(inplace=True)
def __init__(self, input_channels, output_channels, kernel_size=-1, stride=-1, padding=-1, groups=1, channel_shuffle=0, shuffle_groups=1, abits=8, wbits=8, first_layer=0): super(QuantConvBNReLU, self).__init__() self.channel_shuffle_flag = channel_shuffle self.shuffle_groups = shuffle_groups self.quant_conv = QuantConv2d(input_channels, output_channels, kernel_size=kernel_size, stride=stride, padding=padding, groups=groups, a_bits=abits, w_bits=wbits, first_layer=first_layer) self.bn = nn.BatchNorm2d(output_channels) self.relu = nn.ReLU(inplace=True)
def __init__(self, input_channels, output_channels, kernel_size=-1, stride=-1, padding=-1, groups=1, channel_shuffle=0, shuffle_groups=1, abits=8, wbits=8, bn_fuse=0, q_type=1, q_level=0, first_layer=0): super(QuantConvBNReLU, self).__init__() self.channel_shuffle_flag = channel_shuffle self.shuffle_groups = shuffle_groups self.bn_fuse = bn_fuse if self.bn_fuse == 1: self.quant_bn_fuse_conv = QuantBNFuseConv2d( input_channels, output_channels, kernel_size=kernel_size, stride=stride, padding=padding, groups=groups, a_bits=abits, w_bits=wbits, q_type=q_type, q_level=q_level, first_layer=first_layer) else: self.quant_conv = QuantConv2d(input_channels, output_channels, kernel_size=kernel_size, stride=stride, padding=padding, groups=groups, a_bits=abits, w_bits=wbits, q_type=q_type, q_level=q_level, first_layer=first_layer) self.bn = nn.BatchNorm2d( output_channels, momentum=0.01 ) # 考虑量化带来的抖动影响,对momentum进行调整(0.1 ——> 0.01),削弱batch统计参数占比,一定程度抑制抖动。经实验量化训练效果更好,acc提升1%左右 self.relu = QuantReLU(inplace=True, a_bits=abits, q_type=q_type)