用法：pc上20ms
aspp = ASPP(320, [3, 6, 9]) input = torch.randn(2, 320, 10, 10) # torch.onnx.export(pelee_net, input, "pelee_net.onnx", verbose=True) for i in range(10): start=time.time() # x, *shortcuts = net(input) # print(time.time()-start,x.shape) start = time.time() x=aspp(input) print(2,time.time() - start, x.shape) from torch.nn import functional as Fclass ASPPPooling(nn.Sequential): def __init__(self, in_channels, out_channels): super(ASPPPooling, self).__init__( nn.AdaptiveAvgPool2d(1), nn.Conv2d(in_channels, out_channels, 1, bias=False), nn.BatchNorm2d(out_channels), nn.ReLU()) def forward(self, x): size = x.shape[-2:] for mod in self: x = mod(x) return F.interpola