ConvBn2d#
- class torch.ao.nn.intrinsic.qat.ConvBn2d(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=None, padding_mode='zeros', eps=1e-05, momentum=0.1, freeze_bn=False, qconfig=None)[source]#
A ConvBn2d module is a module fused from Conv2d and BatchNorm2d, attached with FakeQuantize modules for weight, used in quantization aware training.
We combined the interface of
torch.nn.Conv2dandtorch.nn.BatchNorm2d.Similar to
torch.nn.Conv2d, with FakeQuantize modules initialized to default.- Variables
freeze_bn –
weight_fake_quant – fake quant module for weight