Skip to content

tensorrt 7.2.2.3版本 #12

@xiangyun-hz

Description

@xiangyun-hz

对于大通道,比如1024个channels的网络,比如最简单的 单层网络

class conv_1024_1024_33_d1(torch.nn.Module):
def init(self):
self.in_channel = 1024
self.out_channel = 1024
self.kernel_w = 3
self.kernel_h = 3
super(conv_1024_1024_33_d1, self).init()
self.conv = torch.nn.Sequential(
OrderedDict([("conv1", torch.nn.Conv2d(self.in_channel, self.out_channel, self.kernel_w, 1, 1))])
)

def forward(self, x):
    return self.conv(x)

用这个pytorch转onnx,然后调用trt的execute接口跑出来的耗时结合这层的计算量计算出来的算力(模型计算量/耗时)大于平台标称算力,这是由于trt加速的原因还是什么?有大神知道的么?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions