Closed pawopawo closed 3 years ago
支持的。 groups 可以是任意的。repvgg.py里有相关的代码
input_dim = self.in_channels // self.groups kernel_value = np.zeros((self.in_channels, input_dim, 3, 3), dtype=np.float32) for i in range(self.in_channels): kernel_value[i, i % input_dim, 1, 1] = 1
shawnding@outlook.com
发件人: pawopawo 发送时间: 2021-01-14 15:34 收件人: DingXiaoH/ACNet 抄送: Subscribed 主题: [DingXiaoH/ACNet] depth conv的实验 (#39) 请问使用depth conv的mobilenet支持这种权重融合吗 — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.
不好意思 刚才发错了 我以为是另一个repo的
shawnding@outlook.com
发件人: pawopawo 发送时间: 2021-01-14 15:34 收件人: DingXiaoH/ACNet 抄送: Subscribed 主题: [DingXiaoH/ACNet] depth conv的实验 (#39) 请问使用depth conv的mobilenet支持这种权重融合吗 — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.
ACNet也是支持depth conv的。ACBlock里的groups可以是任意的。
shawnding@outlook.com
发件人: pawopawo 发送时间: 2021-01-14 15:34 收件人: DingXiaoH/ACNet 抄送: Subscribed 主题: [DingXiaoH/ACNet] depth conv的实验 (#39) 请问使用depth conv的mobilenet支持这种权重融合吗 — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.
谢谢你的回复~ 如果是mobilenetv2 这种含较多1x1conv的网络,是不是1x1conv就不能做re-param了?repvgg或acnet还能有较大的收益吗?
对,1x1没法用1x3或3x1 reparam。这是因为前者的感受野小于后者。但是1x1上可以加identity。
shawnding@outlook.com
发件人: pawopawo 发送时间: 2021-01-15 13:58 收件人: DingXiaoH/ACNet 抄送: Ding Xiaohan; Comment 主题: Re: [DingXiaoH/ACNet] depth conv的实验 (#39) 谢谢你的回复~ 如果是mobilenetv2 这种含较多11conv的网络,是不是11conv就不能做re-param了,repvgg或acnet还能有较大的收益吗? — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.
请问使用depth conv的mobilenet支持这种权重融合吗