search

0aqz0 / SLR

isolated & continuous sign language recognition using CNN+LSTM/3D CNN/GCN/Encoder-Decoder
198 stars 41 forks source link

attention model #1

Closed HaminyG closed 4 years ago

HaminyG commented 4 years ago

Excellent job!! Cause I am a fresh bird in this area, could you please tell me what is l and g which are as inputs for linearattentionblock?

0aqz0 commented 4 years ago

I add the attention blocks and refer to this paper: Learn to Pay Attention(ICLR2018). And this figure shows the architecture in the paper.

image

To my understanding, l represents features extracted from different layers and g represents features extracted from the final layer.

You can check out this paper. In my job, I try to add similar attention blocks to 3D ResNet. But the result now is a bit worse than 3D ResNet without attention :) and I am still tuning the model.

HaminyG commented 4 years ago

Thanks so much. This figure makes more sense. 

发自我的iPhone

------------------ Original ------------------ From: Haodong Zhang <notifications@github.com> Date: Mon,Mar 23,2020 7:57 PM To: 0aqz0/SLR <SLR@noreply.github.com> Cc: HaminyG <992532036@qq.com>, Author <author@noreply.github.com> Subject: Re: [0aqz0/SLR] attention model (#1)

I add the attention blocks and refer to this paper: Learn to Pay Attention(ICLR2018). And this figure shows the architecture in the paper.

To my understanding, l represents features extracted from different layers and g represents features extracted from the final layer.

You can check out this paper. In my job, I try to add similar attention blocks to 3D ResNet. But the result now is a bit worse than 3D ResNet without attention :) and I am still tuning the model.

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

0aqz0 commented 4 years ago

You're welcome.