Training with 800x800 on mobile0.25 but get worse result

[twmht]

@biubug6

Have you ever trained on more larger resolution like 800x800 with mobile0.25? I got much more worse result than default 640x640.

Any idea?

[biubug6]

800*800 would get better result in small face, but it may influences the result of big face.

[twmht]

@biubug6

As compared to the original mxnet implementation, is the difference that you add ssh head?

[zhen8838]

这个SSH里面定义的conv5X5和self.conv7X7还是在使用卷积核为3....

def conv_bn(inp, oup, stride = 1, leaky = 0):
    return nn.Sequential(
        nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
        nn.BatchNorm2d(oup),
        nn.LeakyReLU(negative_slope=leaky, inplace=True)
    )
.
.
.
        self.conv5X5_1 = conv_bn(in_channel, out_channel//4, stride=1, leaky = leaky)
        self.conv5X5_2 = conv_bn_no_relu(out_channel//4, out_channel//4, stride=1)

        self.conv7X7_2 = conv_bn(out_channel//4, out_channel//4, stride=1, leaky = leaky)
        self.conv7x7_3 = conv_bn_no_relu(out_channel//4, out_channel//4, stride=1)

[zhangyuan1994511]

@zhen8838 @biubug6 这块用7x7；5x5；3x3好像没啥影响吧？

[zhangyuan1994511]

感觉只是个命名问题

[zhen8838]

我不知道作者本来想用什么卷积核.

[zhangyuan1994511]

这个就需要问作者了。。。 @biubug6

[biubug6]

It just means to have a 5 5 or 77 receptive field. Large kernels are not recommended at the back of the network because of expensive computing.

[zhangyuan1994511]

@OK! I get it, Thank you for your reply!