网络结构

[shajie17]

你好，num_segments=4,numsegments=16画出来的网络结构是一样的，这是为啥呢?

[zhang-can]

你好，“res5b_pool”层会将时序长度pooling为1。

• num_segments=4时，“res5b_pool”的“kernel_d”为1；
• num_segments=16时，“res5b_pool”的“kernel_d”为4；

[shajie17]

感谢您的回复， 是在ECO.yaml中的这一层吧？

attrs: {kernel_d: 1, kernel_h: 7, kernel_w: 7, stride: 1, mode: ave}

expr: global_pool<=Pooling3d<=res5b_bnid: global_pool

------------------ Original ------------------ From: "Can Zhang"[email protected]; Date: Mon, Feb 18, 2019 12:58 PM To: "zhang-can/ECO-pytorch"[email protected]; Cc: "shajie17"[email protected]; "Author"[email protected]; Subject: Re: [zhang-can/ECO-pytorch] 网络结构 (#33)

你好，“res5b_pool”层会将时序长度pooling为1。

num_segments=4时，“res5b_pool”的“kernel_d”为1；

num_segments=16时，“res5b_pool”的“kernel_d”为4；

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[chen-ming2019]

@ zhang-can Hi，for the ECO-full model,how do 3D nets(3D resnet18) and 2D nets (inception-4a later) exploit different length of input videos ? Specifically,the code for the paper didn't give a weight combination with different input videos(increase or supress ) instead of giving in two branches simply.I don't know that what I said is right or wrong .so can you give me some advice? Thank you!